{"evidence_id": "Orca.ev_0001", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Introduction", "quote": "We argue that the essential next step toward general intelligence is to build a model that can continuously learn and self-evolve like a human, and ultimately transcend human cognitive boundaries. As it internalizes physical laws, causal r…", "summary": "We argue that the essential next step toward general intelligence is to build a model that can continuously learn and self-evolve like a human, and ultimately transcend human cogn…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0001", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Introduction"}}}
{"evidence_id": "Orca.ev_0002", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Introduction", "quote": "From this perspective, intelligence should not merely be Next-Token-Prediction model that can respond to instructions (DeepSeek-AI et al., 2026;Qwen Team, 2026b;Wang et al., 2026a;OpenAI, 2026b), Next-Frame-Prediction model that can genera…", "summary": "From this perspective, intelligence should not merely be Next-Token-Prediction model that can respond to instructions (DeepSeek-AI et al., 2026;Qwen Team, 2026b;Wang et al., 2026a…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0002", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Introduction"}}}
{"evidence_id": "Orca.ev_0003", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Introduction", "quote": "We present Orca, a world learner that takes an initial step toward the above goal by learning a world latent space. Figure 1 shows the Orca's overall framework. In this version, Encoder focuses on two fundamental signal types: visual and l…", "summary": "We present Orca, a world learner that takes an initial step toward the above goal by learning a world latent space. Figure 1 shows the Orca's overall framework. In this version, E…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0003", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Introduction"}}}
{"evidence_id": "Orca.ev_0004", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Introduction", "quote": "1) Unconscious learning aims to learn natural and dense state transitions from continuous video. This process does not rely on labeled tags, but instead uses the supervision provided by itself. The model learns natural evolution by predict…", "summary": "1) Unconscious learning aims to learn natural and dense state transitions from continuous video. This process does not rely on labeled tags, but instead uses the supervision provi…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0004", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Introduction"}}}
{"evidence_id": "Orca.ev_0005", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Introduction", "quote": "2) Conscious learning aims to learn meaningful and sparse state transitions under the constraints of instructions. The model uses textual constraints to learn meaningful state transitions at the event level.", "summary": "2) Conscious learning aims to learn meaningful and sparse state transitions under the constraints of instructions. The model uses textual constraints to learn meaningful state tra…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0005", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Introduction"}}}
{"evidence_id": "Orca.ev_0006", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Introduction", "quote": "Orca builds a world latent space through the two paradigms. The Decoder reads out text, images, and actions. Note that these readouts are not intended to chase task-specific SOTA performance, but to examine two core questions: 1) the propo…", "summary": "Orca builds a world latent space through the two paradigms. The Decoder reads out text, images, and actions. Note that these readouts are not intended to chase task-specific SOTA…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0006", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Introduction"}}}
{"evidence_id": "Orca.ev_0007", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Introduction", "quote": "• We propose Orca. Orca learns a world latent space from multimodal world signals. This latent space can serve as a general interface for multimodal downstream readouts. • We design two complementary learning paradigms. Unconscious learnin…", "summary": "• We propose Orca. Orca learns a world latent space from multimodal world signals. This latent space can serve as a general interface for multimodal downstream readouts. • We desi…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0007", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Introduction"}}}
{"evidence_id": "Orca.ev_0008", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Introduction", "quote": "Encoder. Orca focuses on the Encoder, which learns a unified world latent space for state abstraction and state transition. The overview of the Encoder is shown in Figure 2. It uses a native pre-trained VLM (Qwen Team, 2026a) aligned with…", "summary": "Encoder. Orca focuses on the Encoder, which learns a unified world latent space for state abstraction and state transition. The overview of the Encoder is shown in Figure 2. It us…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0008", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Introduction"}}}
{"evidence_id": "Orca.ev_0009", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Introduction", "quote": "The input of unconscious learning is a certain frame 𝑣 𝑡 of the video 𝑉, and the output is the prediction latent v𝑙 𝑡+1 of the next adjacent frame. After passing through the VLM, 𝑣 𝑡 will be used to obtain the predicted v𝑙 𝑡+1 through two…", "summary": "The input of unconscious learning is a certain frame 𝑣 𝑡 of the video 𝑉, and the output is the prediction latent v𝑙 𝑡+1 of the next adjacent frame. After passing through the VLM,…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0009", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Introduction"}}}
{"evidence_id": "Orca.ev_0010", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Introduction", "quote": "Decoder. The learned latent space is read out by the modality-specific Decoder to extract multi-modal information. Since the decoder is not the focus of this section, its details will be shown in Section 3.2.", "summary": "Decoder. The learned latent space is read out by the modality-specific Decoder to extract multi-modal information. Since the decoder is not the focus of this section, its details…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0010", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Introduction"}}}
{"evidence_id": "Orca.ev_0011", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Training", "quote": "Orca is trained in two stages. The pre-training stage learns the world latent through large-scale visual and language data. In the downstream post-training stage, the Orca's backbone is frozen. Only modality-specific readout modules are tr…", "summary": "Orca is trained in two stages. The pre-training stage learns the world latent through large-scale visual and language data. In the downstream post-training stage, the Orca's backb…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0011", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Training"}}}
{"evidence_id": "Orca.ev_0012", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Pre-Training Recipe", "quote": "Orca pre-training instantiates world-state modeling with three objectives: 1) observation-only state transition, 2) event-conditioned state transition, and 3) VQA response generation. The two state-transition objectives are implemented thr…", "summary": "Orca pre-training instantiates world-state modeling with three objectives: 1) observation-only state transition, 2) event-conditioned state transition, and 3) VQA response generat…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0012", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Pre-Training Recipe"}}}
{"evidence_id": "Orca.ev_0013", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Pre-Training Recipe", "quote": "Learning Objectives. We instantiate pre-training with the three objectives. 1) observation-only state transition and 2) event-conditioned state transition are implemented with learnable queries in the input of the VLM backbone. The input i…", "summary": "Learning Objectives. We instantiate pre-training with the three objectives. 1) observation-only state transition and 2) event-conditioned state transition are implemented with lea…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0013", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Pre-Training Recipe"}}}
{"evidence_id": "Orca.ev_0014", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Pre-Training Recipe", "quote": "2) Event-conditioned state transition. This objective forms conscious learning with language. Given 𝑣 𝑡 , <Query 1> 𝑞 1 , 𝑒 𝑡+Δ , and <Query 2> 𝑞 2 , the last-layer hidden state of 𝑞 2 is passed through the two layers of MLP to predict v𝑙…", "summary": "2) Event-conditioned state transition. This objective forms conscious learning with language. Given 𝑣 𝑡 , <Query 1> 𝑞 1 , 𝑒 𝑡+Δ , and <Query 2> 𝑞 2 , the last-layer hidden state o…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0014", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Pre-Training Recipe"}}}
{"evidence_id": "Orca.ev_0015", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Pre-Training Recipe", "quote": "The first two objectives are supervised in the latent space of the vision encoder. This design focuses on pre-training for state modeling rather than pixel-level reconstruction. The last objective uses the LM head. Given 𝑉 and 𝑙 𝑞 , the LM…", "summary": "The first two objectives are supervised in the latent space of the vision encoder. This design focuses on pre-training for state modeling rather than pixel-level reconstruction. T…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0015", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Pre-Training Recipe"}}}
{"evidence_id": "Orca.ev_0016", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Pre-Training Recipe", "quote": "Training Components. The full pre-training loss combines the three components as:", "summary": "Training Components. The full pre-training loss combines the three components as:", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0016", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Pre-Training Recipe"}}}
{"evidence_id": "Orca.ev_0017", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Pre-Training Recipe", "quote": "where, 𝜆 obs , 𝜆 evt , and 𝜆 vqa are weighting coefficients that balance the contributions of the two objectives.", "summary": "where, 𝜆 obs , 𝜆 evt , and 𝜆 vqa are weighting coefficients that balance the contributions of the two objectives.", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0017", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Pre-Training Recipe"}}}
{"evidence_id": "Orca.ev_0018", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Pre-Training Recipe", "quote": "Here, L obs corresponds to unconscious learning from naturally occurring visual transition. L evt and L vqa correspond to conscious learning through language-specified transitions and common sense. L evt uses the ground truth latent of a f…", "summary": "Here, L obs corresponds to unconscious learning from naturally occurring visual transition. L evt and L vqa correspond to conscious learning through language-specified transitions…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0018", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Pre-Training Recipe"}}}
{"evidence_id": "Orca.ev_0019", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Pre-Training Recipe", "quote": "𝑡+Δ under the constraint of 𝑒 𝑡+Δ . L vqa represents the standard VQA loss. Given a visual information 𝑉 and a question 𝑙 𝑞 , Orca learns to produce the target response 𝑙 𝑎 . The details of the sampling ratio, loss coefficients, and optimi…", "summary": "𝑡+Δ under the constraint of 𝑒 𝑡+Δ . L vqa represents the standard VQA loss. Given a visual information 𝑉 and a question 𝑙 𝑞 , Orca learns to produce the target response 𝑙 𝑎 . The…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0019", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Pre-Training Recipe"}}}
{"evidence_id": "Orca.ev_0020", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Pre-Training Data", "quote": "Data Organization. The three collections provide complementary supervision for learning world states and their transitions. The pre-training data is shown in Figure 3. A. Video Data is built from visual signals and covers four types of rea…", "summary": "Data Organization. The three collections provide complementary supervision for learning world states and their transitions. The pre-training data is shown in Figure 3. A. Video Da…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0020", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Pre-Training Data"}}}
{"evidence_id": "Orca.ev_0021", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Pre-Training Data", "quote": "B. Event data is derived from A. Video Data through multi-level event segmentation and language annotation. Coarse events describe the main steps of a temporal process, while fine-grained events capture the shorter state transitions within…", "summary": "B. Event data is derived from A. Video Data through multi-level event segmentation and language annotation. Coarse events describe the main steps of a temporal process, while fine…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0021", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Pre-Training Data"}}}
{"evidence_id": "Orca.ev_0022", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Pre-Training Data", "quote": "Across these collections, data construction is grounded in the real world. The video data is built from real-world videos, while event and VQA data are constructed on top of these observations to describe state transition, physical relatio…", "summary": "Across these collections, data construction is grounded in the real world. The video data is built from real-world videos, while event and VQA data are constructed on top of these…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0022", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Pre-Training Data"}}}
{"evidence_id": "Orca.ev_0023", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Pre-Training Data", "quote": "The existing data includes 125K hours of general video data, 160M of event annotations, and 11.5M of general VQA data. In this version, only one-tenth of the video data are used. The remaining data will be used in Orca's subsequent version…", "summary": "The existing data includes 125K hours of general video data, 160M of event annotations, and 11.5M of general VQA data. In this version, only one-tenth of the video data are used.…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0023", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Pre-Training Data"}}}
{"evidence_id": "Orca.ev_0024", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Downstream Post-Training", "quote": "After pre-training, Orca is connected to downstream readout interfaces for language, vision, and action.", "summary": "After pre-training, Orca is connected to downstream readout interfaces for language, vision, and action.", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0024", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Downstream Post-Training"}}}
{"evidence_id": "Orca.ev_0025", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Downstream Post-Training", "quote": "Note that our goal is to explore, in essence, whether the learned latent is effective for downstream tasks. So, Orca's backbone is always frozen, and only the corresponding readout modules are trainable. In other words, if Orca is intended…", "summary": "Note that our goal is to explore, in essence, whether the learned latent is effective for downstream tasks. So, Orca's backbone is always frozen, and only the corresponding readou…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0025", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Downstream Post-Training"}}}
{"evidence_id": "Orca.ev_0026", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "To Language: Text Generation", "quote": "As shown in Figure 4(a). Given a visual observation and an instruction, Orca produces the response through LM head, without attaching an additional decoder. It expresses Orca's latent in natural language.", "summary": "As shown in Figure 4(a). Given a visual observation and an instruction, Orca produces the response through LM head, without attaching an additional decoder. It expresses Orca's la…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0026", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "To Language: Text Generation"}}}
{"evidence_id": "Orca.ev_0027", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "To Vision: Image Prediction", "quote": "Vision Readout Recipe. As shown in Figure 4(b), the vision readout maps the latent to a pixel-level image. Since Orca focuses on the encoder, the decoder uses a pre-trained model to show the effectiveness of latent. The latent is passed th…", "summary": "Vision Readout Recipe. As shown in Figure 4(b), the vision readout maps the latent to a pixel-level image. Since Orca focuses on the encoder, the decoder uses a pre-trained model…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0027", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "To Vision: Image Prediction"}}}
{"evidence_id": "Orca.ev_0028", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "To Vision: Image Prediction", "quote": "Image Prediction Data. This readout training uses paired current and target frames sampled from A. Video Data in Figure 3. Given the image and an instruction, Orca first produces the latent of the target frame using frozen Orca, and then t…", "summary": "Image Prediction Data. This readout training uses paired current and target frames sampled from A. Video Data in Figure 3. Given the image and an instruction, Orca first produces…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0028", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "To Vision: Image Prediction"}}}
{"evidence_id": "Orca.ev_0029", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Infrastructures", "quote": "Orca's infra uses the self-developed FlagScale (FlagOS, 2026) and makes the following improvements:", "summary": "Orca's infra uses the self-developed FlagScale (FlagOS, 2026) and makes the following improvements:", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0029", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Infrastructures"}}}
{"evidence_id": "Orca.ev_0030", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Infrastructures", "quote": "1) FlagScale training framework. We use FlagScale and rebuild the Orca training with FSDP2, enabling more flexible parameter sharding, better memory control, and stable training.", "summary": "1) FlagScale training framework. We use FlagScale and rebuild the Orca training with FSDP2, enabling more flexible parameter sharding, better memory control, and stable training.", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0030", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Infrastructures"}}}
{"evidence_id": "Orca.ev_0031", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Infrastructures", "quote": "2) Memory-efficient loss and recompute. We adopt Chunked Cross-Entropy Loss to avoid materializing full logits during loss computation, and further apply activation recomputation to trade moderate computation overhead for substantial memor…", "summary": "2) Memory-efficient loss and recompute. We adopt Chunked Cross-Entropy Loss to avoid materializing full logits during loss computation, and further apply activation recomputation…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0031", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Infrastructures"}}}
{"evidence_id": "Orca.ev_0032", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Infrastructures", "quote": "3) Communication scheduling. We introduce forward/backward pre-fetching to overlap FSDP allgather communication with computation, and remove unnecessary FSDP sharding for visual blocks.", "summary": "3) Communication scheduling. We introduce forward/backward pre-fetching to overlap FSDP allgather communication with computation, and remove unnecessary FSDP sharding for visual b…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0032", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Infrastructures"}}}
{"evidence_id": "Orca.ev_0033", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Effectiveness and Scaling Behavior", "quote": "Before giving the downstream-specific results, we first explore whether Orca's core hypotheses hold. As presented in the introduction, Orca is designed to learn a world latent space through next state prediction, and this latent space is e…", "summary": "Before giving the downstream-specific results, we first explore whether Orca's core hypotheses hold. As presented in the introduction, Orca is designed to learn a world latent spa…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0033", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Effectiveness and Scaling Behavior"}}}
{"evidence_id": "Orca.ev_0034", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Effectiveness and Scaling Behavior", "quote": "• The total loss of Orca decreases as the pre-training data scale up, and the larger Orca achieves a lower objective loss than the smaller one. This trend suggests that Orca provides an effective learning paradigm for building world latent…", "summary": "• The total loss of Orca decreases as the pre-training data scale up, and the larger Orca achieves a lower objective loss than the smaller one. This trend suggests that Orca provi…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0034", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Effectiveness and Scaling Behavior"}}}
{"evidence_id": "Orca.ev_0035", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Downstream Readout Analysis", "quote": "Following the above discussion of Question 1.2 and Answer 1.2, we present the quantitative evaluation results of Orca across three downstream tasks: text generation, image prediction, and action generation. Note that we do not construct or…", "summary": "Following the above discussion of Question 1.2 and Answer 1.2, we present the quantitative evaluation results of Orca across three downstream tasks: text generation, image predict…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0035", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Downstream Readout Analysis"}}}
{"evidence_id": "Orca.ev_0036", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Comparison on Text Generation", "quote": "To evaluate how state transition modeling enhances abstract reasoning, we conducted text generation assessments on OOD understanding. The details of benchmarks and baselines of the text generation can be seen in Appendix E.1.", "summary": "To evaluate how state transition modeling enhances abstract reasoning, we conducted text generation assessments on OOD understanding. The details of benchmarks and baselines of th…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0036", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Comparison on Text Generation"}}}
{"evidence_id": "Orca.ev_0037", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Comparison on Text Generation", "quote": "Benchmarks. We evaluate Orca on a complementary suite of benchmarks that probe different aspects of world-state and state-transition understanding: MVBench (Li et al., 2024), TemporalBench (Cai et al., 2024), 3DSRBench (Ma et al., 2025), a…", "summary": "Benchmarks. We evaluate Orca on a complementary suite of benchmarks that probe different aspects of world-state and state-transition understanding: MVBench (Li et al., 2024), Temp…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0037", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Comparison on Text Generation"}}}
{"evidence_id": "Orca.ev_0038", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Comparison on Text Generation", "quote": "Baselines. We compare Orca with two categories of baselines:", "summary": "Baselines. We compare Orca with two categories of baselines:", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0038", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Comparison on Text Generation"}}}
{"evidence_id": "Orca.ev_0039", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Comparison on Text Generation", "quote": "• World models: V-JEPA 2.1 (Mur-Labadia et al., 2026), Emu3 (Wang et al., 2026a), and Emu3.5 (Cui et al., 2025). • Vision-language models: Qwen3.5 (Qwen Team, 2026a), Gemma 4 (Deepmind, 2026b), DeepSeek-VL2 (Wu et al., 2025b), MiniCPM-V-4.…", "summary": "• World models: V-JEPA 2.1 (Mur-Labadia et al., 2026), Emu3 (Wang et al., 2026a), and Emu3.5 (Cui et al., 2025). • Vision-language models: Qwen3.5 (Qwen Team, 2026a), Gemma 4 (Dee…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0039", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Comparison on Text Generation"}}}
{"evidence_id": "Orca.ev_0040", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Results and Analysis. Based on", "quote": "1) State Transition. It focuses on state transitions induced by actions or temporal evolution.", "summary": "1) State Transition. It focuses on state transitions induced by actions or temporal evolution.", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0040", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Results and Analysis. Based on"}}}
{"evidence_id": "Orca.ev_0041", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Results and Analysis. Based on", "quote": "2) Commonsense Reasoning. It evaluates the internalization of social and physical commonsense knowledge, as well as the ability to reason causally.", "summary": "2) Commonsense Reasoning. It evaluates the internalization of social and physical commonsense knowledge, as well as the ability to reason causally.", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0041", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Results and Analysis. Based on"}}}
{"evidence_id": "Orca.ev_0042", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Results and Analysis. Based on", "quote": "3) Spatial Relations. It measures the understanding of three-dimensional geometric relationships. 4) Dynamic Motion. It assesses quantitative reasoning over kinematic properties, including velocity, direction vectors, and higher-order moti…", "summary": "3) Spatial Relations. It measures the understanding of three-dimensional geometric relationships. 4) Dynamic Motion. It assesses quantitative reasoning over kinematic properties,…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0042", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Results and Analysis. Based on"}}}
{"evidence_id": "Orca.ev_0043", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Results and Analysis. Based on", "quote": "1) More accurate state transition. Orca predicts future states more accurately and demonstrates a deeper understanding of temporal dynamics.", "summary": "1) More accurate state transition. Orca predicts future states more accurately and demonstrates a deeper understanding of temporal dynamics.", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0043", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Results and Analysis. Based on"}}}
{"evidence_id": "Orca.ev_0044", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Results and Analysis. Based on", "quote": "2) Common-sense and counterfactual reasoning. Orca achieves more reliable common-sense reasoning and counterfactual reasoning through causal alignment of conscious learning.", "summary": "2) Common-sense and counterfactual reasoning. Orca achieves more reliable common-sense reasoning and counterfactual reasoning through causal alignment of conscious learning.", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0044", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Results and Analysis. Based on"}}}
{"evidence_id": "Orca.ev_0045", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Results and Analysis. Based on", "quote": "3) Strong spatial understanding. Orca can capture geometric continuity, reduce spatial inconsistencies, and improve robustness under complex perspectives.", "summary": "3) Strong spatial understanding. Orca can capture geometric continuity, reduce spatial inconsistencies, and improve robustness under complex perspectives.", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0045", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Results and Analysis. Based on"}}}
{"evidence_id": "Orca.ev_0046", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "4) Dynamic motion consistency.", "quote": "Orca can better capture temporal continuity and motion inertia.", "summary": "Orca can better capture temporal continuity and motion inertia.", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0046", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "4) Dynamic motion consistency."}}}
{"evidence_id": "Orca.ev_0047", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Comparison on Image Prediction", "quote": "To visualize the capability of the state transition, we performed a comparison on image prediction. The details of benchmarks and baselines of the image prediction can be seen in Appendix E.2.", "summary": "To visualize the capability of the state transition, we performed a comparison on image prediction. The details of benchmarks and baselines of the image prediction can be seen in…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0047", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Comparison on Image Prediction"}}}
{"evidence_id": "Orca.ev_0048", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Comparison on Image Prediction", "quote": "Benchmark. Our motivation is not to create a painter, but to explore whether the latent possesses the ability to predict future states. So, instead of generating or simulating scenarios, we build a real-world dataset, PRICE-V0.1 (i.e., Pre…", "summary": "Benchmark. Our motivation is not to create a painter, but to explore whether the latent possesses the ability to predict future states. So, instead of generating or simulating sce…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0048", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Comparison on Image Prediction"}}}
{"evidence_id": "Orca.ev_0049", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Metrics.", "quote": "In PRICE-V0.1, we use Gemini 3.1 Pro (Deepmind, 2026a), GPT 5.4 (OpenAI, 2026b), Doubao-Seed-2.0-Pro-260215 (Doubao Team, 2026), and open-source Gemma 4-31B (Deepmind, 2026b) for evaluation. The specific evaluation prompt is shown in the L…", "summary": "In PRICE-V0.1, we use Gemini 3.1 Pro (Deepmind, 2026a), GPT 5.4 (OpenAI, 2026b), Doubao-Seed-2.0-Pro-260215 (Doubao Team, 2026), and open-source Gemma 4-31B (Deepmind, 2026b) for…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0049", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Metrics."}}}
{"evidence_id": "Orca.ev_0050", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Metrics.", "quote": "Baselines. We selected recent image generation models with a similar size to Orca as baselines: including OmniGen2 (Wu et al., 2026) Results and Analysis. Based on Table 3 and Figure 7, we obtained two conclusions: 2) Orca better predicts…", "summary": "Baselines. We selected recent image generation models with a similar size to Orca as baselines: including OmniGen2 (Wu et al., 2026) Results and Analysis. Based on Table 3 and Fig…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0050", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Metrics."}}}
{"evidence_id": "Orca.ev_0051", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Metrics.", "quote": "These results suggest that the learned world latent provides useful state-transition information for visual readout, enabling more physically grounded image prediction for real-world interactions.", "summary": "These results suggest that the learned world latent provides useful state-transition information for visual readout, enabling more physically grounded image prediction for real-wo…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0051", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Metrics."}}}
{"evidence_id": "Orca.ev_0052", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Comparison on Action Generation", "quote": "To truly apply state transition modeling capabilities to the real world, we performed the embodied realrobot tasks. The details of benchmarks, metrics, and baselines are shown in Appendix E.3.", "summary": "To truly apply state transition modeling capabilities to the real world, we performed the embodied realrobot tasks. The details of benchmarks, metrics, and baselines are shown in…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0052", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Comparison on Action Generation"}}}
{"evidence_id": "Orca.ev_0053", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Comparison on Action Generation", "quote": "Benchmarks. We used a dual-arm wheeled robot to collect data on five tasks: Take Book, Stacked Bowls, Pull Out Tissue, Stamp, and Scoop Sugar. We performed two OOD settings: environment and object OOD.", "summary": "Benchmarks. We used a dual-arm wheeled robot to collect data on five tasks: Take Book, Stacked Bowls, Pull Out Tissue, Stamp, and Scoop Sugar. We performed two OOD settings: envir…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0053", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Comparison on Action Generation"}}}
{"evidence_id": "Orca.ev_0054", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Comparison on Action Generation", "quote": "Metrics. We report the rule-based scores, which measure key-stage task completion. The rule-based score is shown in Table E2. We further employ PRM-as-a-Judge series (Ji et al., 2026; PRM-as-a-Judge Team, 2026) to provide dense trajectory-…", "summary": "Metrics. We report the rule-based scores, which measure key-stage task completion. The rule-based score is shown in Table E2. We further employ PRM-as-a-Judge series (Ji et al., 2…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0054", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Comparison on Action Generation"}}}
{"evidence_id": "Orca.ev_0055", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Comparison on Action Generation", "quote": "Baselines. We compare Orca with V-JEPA 2.1 (Mur-Labadia et al., 2026), Qwen3.5 (Qwen Team, 2026a), and 𝜋 0.5 (Physical Intelligence et al., 2025). For V-JEPA 2.1 and Qwen3.5, we connect them to the same Action Expert as Orca, i.e., V-JEPA…", "summary": "Baselines. We compare Orca with V-JEPA 2.1 (Mur-Labadia et al., 2026), Qwen3.5 (Qwen Team, 2026a), and 𝜋 0.5 (Physical Intelligence et al., 2025). For V-JEPA 2.1 and Qwen3.5, we c…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0055", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Comparison on Action Generation"}}}
{"evidence_id": "Orca.ev_0056", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Results and Analysis. Based on Table 4, we obtained two conclusions:", "quote": "1) Orca's learning paradigm and learned world latent transfers effectively to action readout. Under the from-scratch Action Expert, Orca outperforms Qwen3.5 in all OOD settings, achieving a breakthrough from 0% success rate. It is also com…", "summary": "1) Orca's learning paradigm and learned world latent transfers effectively to action readout. Under the from-scratch Action Expert, Orca outperforms Qwen3.5 in all OOD settings, a…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0056", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Results and Analysis. Based on Table 4, we obtained two conclusions:"}}}
{"evidence_id": "Orca.ev_0057", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Results and Analysis. Based on Table 4, we obtained two conclusions:", "quote": "The metrics are trajectory-level diagnostics from PRM-as-a-Judge. Note that the backbones of all methods are frozen, and Action Experts for Orca, V-JEPA 2.1, and Qwen3.5 are trainable from scratch. that Orca's learning paradigm has a signi…", "summary": "The metrics are trajectory-level diagnostics from PRM-as-a-Judge. Note that the backbones of all methods are frozen, and Action Experts for Orca, V-JEPA 2.1, and Qwen3.5 are train…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0057", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Results and Analysis. Based on Table 4, we obtained two conclusions:"}}}
{"evidence_id": "Orca.ev_0058", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Results and Analysis. Based on Table 4, we obtained two conclusions:", "quote": "2) Orca consistently advances the task and recovers better from execution errors. The metrics show that Orca is more likely to make meaningful intermediate progress during execution, while suffering less from stagnation. Its higher FNS ind…", "summary": "2) Orca consistently advances the task and recovers better from execution errors. The metrics show that Orca is more likely to make meaningful intermediate progress during executi…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0058", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Results and Analysis. Based on Table 4, we obtained two conclusions:"}}}
{"evidence_id": "Orca.ev_0059", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Ablation", "quote": "In the current learning paradigm, there are three losses, i.e., 1) Observation-only state transition 𝜆 obs ; 2) Event-conditioned state transition 𝜆 evt ; 3) VQA response generation 𝜆 vqa . So we ablated the different losses.", "summary": "In the current learning paradigm, there are three losses, i.e., 1) Observation-only state transition 𝜆 obs ; 2) Event-conditioned state transition 𝜆 evt ; 3) VQA response generati…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0059", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Ablation"}}}
{"evidence_id": "Orca.ev_0060", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Ablation", "quote": "The ablation results are shown in Table 5. The results in Table 5 demonstrate that: 1) The three pre-training objectives provide the most balanced downstream readouts. When 𝜆 obs , 𝜆 evt , and 𝜆 vqa are jointly used, Orca achieves the most…", "summary": "The ablation results are shown in Table 5. The results in Table 5 demonstrate that: 1) The three pre-training objectives provide the most balanced downstream readouts. When 𝜆 obs…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0060", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Ablation"}}}
{"evidence_id": "Orca.ev_0061", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Ablation", "quote": "2) Observation-only transition is especially important for action readout. Adding 𝜆 obs clearly improves action generation. This suggests that dense natural dynamics from continuous videos provide useful information about temporal changes,…", "summary": "2) Observation-only transition is especially important for action readout. Adding 𝜆 obs clearly improves action generation. This suggests that dense natural dynamics from continuo…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0061", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Ablation"}}}
{"evidence_id": "Orca.ev_0062", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "3) Event-conditioned transition is the key supervision for vision readout.", "quote": "Image prediction requires the model to infer a target state under a semantic condition. 𝜆 evt aligns language-described events with visual state transition, enabling Orca to predict instruction-or event-guided target states rather than onl…", "summary": "Image prediction requires the model to infer a target state under a semantic condition. 𝜆 evt aligns language-described events with visual state transition, enabling Orca to predi…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0062", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "3) Event-conditioned transition is the key supervision for vision readout."}}}
{"evidence_id": "Orca.ev_0063", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "3) Event-conditioned transition is the key supervision for vision readout.", "quote": "4) VQA response generation preserves the language interface and strengthens semantic grounding. 𝜆 vqa enables Orca to maintain natural-language readout ability and provides semantic and commonsense constraints for the learned world latent…", "summary": "4) VQA response generation preserves the language interface and strengthens semantic grounding. 𝜆 vqa enables Orca to maintain natural-language readout ability and provides semant…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0063", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "3) Event-conditioned transition is the key supervision for vision readout."}}}
{"evidence_id": "Orca.ev_0064", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Conclusion", "quote": "We presented Orca, a world learner built around a world latent space. Rather than being purpose-built for isolated downstream tasks such as question answering, visual frame prediction, or action generation, Orca adopts a fundamentally diff…", "summary": "We presented Orca, a world learner built around a world latent space. Rather than being purpose-built for isolated downstream tasks such as question answering, visual frame predic…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0064", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Conclusion"}}}
{"evidence_id": "Orca.ev_0065", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "Conclusion", "quote": "Discussion & Limitation. Orca is still an early step toward general world foundation models. We discuss its current boundaries together with the research directions they suggest for the community.", "summary": "Discussion & Limitation. Orca is still an early step toward general world foundation models. We discuss its current boundaries together with the research directions they suggest f…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0065", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Conclusion"}}}
{"evidence_id": "Orca.ev_0066", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "1) Limited multimodal world signals.", "quote": "Orca currently learns mainly from vision and language, which cover only a subset of the multimodal world signals. However, many state transitions are expressed through other sensory or physical signals. For example, whether water is boilin…", "summary": "Orca currently learns mainly from vision and language, which cover only a subset of the multimodal world signals. However, many state transitions are expressed through other senso…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0066", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "1) Limited multimodal world signals."}}}
{"evidence_id": "Orca.ev_0067", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "1) Limited multimodal world signals.", "quote": "2) ViT space supervision. Orca aimed to provide a new learning paradigm that, in all other respects, employed a naive setup, thus using a pre-trained VLM and supervising visual state prediction within a frozen vision encoder. This design s…", "summary": "2) ViT space supervision. Orca aimed to provide a new learning paradigm that, in all other respects, employed a naive setup, thus using a pre-trained VLM and supervising visual st…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0067", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "1) Limited multimodal world signals."}}}
{"evidence_id": "Orca.ev_0068", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "1) Limited multimodal world signals.", "quote": "3) Model size limited. Due to resource constraints, our current experiments are mainly conducted at the 4B and 0.8B scale. The current scale is insufficient to fully integrate greater world knowledge, more modalities, and more data. We fou…", "summary": "3) Model size limited. Due to resource constraints, our current experiments are mainly conducted at the 4B and 0.8B scale. The current scale is insufficient to fully integrate gre…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0068", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "1) Limited multimodal world signals."}}}
{"evidence_id": "Orca.ev_0069", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "4) Vision benchmark limited.", "quote": "Although the proposed PRICE-V0.1 covers multiple real-world data sources, its scale, diversity, and interaction richness are still limited. We hope it can serve as an initial step toward a more comprehensive evaluation of real-world state…", "summary": "Although the proposed PRICE-V0.1 covers multiple real-world data sources, its scale, diversity, and interaction richness are still limited. We hope it can serve as an initial step…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0069", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "4) Vision benchmark limited."}}}
{"evidence_id": "Orca.ev_0070", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "5) Short-horizon transition supervision.", "quote": "The current state-transition supervision is constrained by the event annotation. Most event annotations describe short-horizon, minute-level state transitions, which are suitable for learning local transitions but insufficient for modeling…", "summary": "The current state-transition supervision is constrained by the event annotation. Most event annotations describe short-horizon, minute-level state transitions, which are suitable…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0070", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "5) Short-horizon transition supervision."}}}
{"evidence_id": "Orca.ev_0071", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "5) Short-horizon transition supervision.", "quote": "6) Downstream readout limited. Currently, we have verified that the world latent we have learned is readout language, vision, and action. However, this is far from enough, as information from other fields such as hearing, quantum circuits,…", "summary": "6) Downstream readout limited. Currently, we have verified that the world latent we have learned is readout language, vision, and action. However, this is far from enough, as info…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0071", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "5) Short-horizon transition supervision."}}}
{"evidence_id": "Orca.ev_0072", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "7) Loss function limited.", "quote": "We use three losses to fully train Orca, but this is not consistent enough for the Next-State-Prediction modeling. A simpler loss and supervision need to be proposed.", "summary": "We use three losses to fully train Orca, but this is not consistent enough for the Next-State-Prediction modeling. A simpler loss and supervision need to be proposed.", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0072", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "7) Loss function limited."}}}
{"evidence_id": "Orca.ev_0073", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "8)", "quote": "Embodied task difficulty limited. Our settings are quite stringent, resulting in lower performance. However, it's undeniable that the current embodiment tasks are still relatively short and easy.", "summary": "Embodied task difficulty limited. Our settings are quite stringent, resulting in lower performance. However, it's undeniable that the current embodiment tasks are still relatively…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0073", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "8)"}}}
{"evidence_id": "Orca.ev_0074", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "8)", "quote": "Future Works. We also provided some inspiration for the community, including:", "summary": "Future Works. We also provided some inspiration for the community, including:", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0074", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "8)"}}}
{"evidence_id": "Orca.ev_0075", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "8)", "quote": "1) More modalities input. The crucial next step is not simply to add more modalities, but to align them to the same underlying state to better constrain state transitions with the laws of physics.", "summary": "1) More modalities input. The crucial next step is not simply to add more modalities, but to align them to the same underlying state to better constrain state transitions with the…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0075", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "8)"}}}
{"evidence_id": "Orca.ev_0076", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "8)", "quote": "2) Toward native world-state modeling. Native world foundation models can be pre-trained from scratch. To overcome the constraints imposed by a certain existing ViT space or other embedding model spaces, a unified world latent space can be…", "summary": "2) Toward native world-state modeling. Native world foundation models can be pre-trained from scratch. To overcome the constraints imposed by a certain existing ViT space or other…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0076", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "8)"}}}
{"evidence_id": "Orca.ev_0077", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "8)", "quote": "3) A world model state transition evaluation system. This system constructs a unified evaluation framework for state prediction, intervention response, physical quantifiability, and counterfactual inference, preventing world models from re…", "summary": "3) A world model state transition evaluation system. This system constructs a unified evaluation framework for state prediction, intervention response, physical quantifiability, a…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0077", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "8)"}}}
{"evidence_id": "Orca.ev_0078", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "8)", "quote": "4) Model-Data-Evaluation self-evolutionary closed loop. The model autonomously generates interaction trajectories and counterfactual samples, which are automatically evaluated and value-filtered before being fed back into the training syst…", "summary": "4) Model-Data-Evaluation self-evolutionary closed loop. The model autonomously generates interaction trajectories and counterfactual samples, which are automatically evaluated and…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0078", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "8)"}}}
{"evidence_id": "Orca.ev_0079", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "5) Expanding the boundaries of human cognition.", "quote": "Gradually extending from embodied intelligence to complex systems such as AI for science, microscopic quantum mechanics, macroscopic universe, and life sciences, using a unified state transition world representation to support scientific d…", "summary": "Gradually extending from embodied intelligence to complex systems such as AI for science, microscopic quantum mechanics, macroscopic universe, and life sciences, using a unified s…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0079", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "5) Expanding the boundaries of human cognition."}}}
{"evidence_id": "Orca.ev_0080", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "B. Related Work", "quote": "We organize related work according to the primary learning objective at the center of each paradigm, rather than by all downstream capabilities a model may exhibit. Under this view, some unified models span multiple capability domains, but…", "summary": "We organize related work according to the primary learning objective at the center of each paradigm, rather than by all downstream capabilities a model may exhibit. Under this vie…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0080", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "B. Related Work"}}}
{"evidence_id": "Orca.ev_0081", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "B.1. Self-Supervised Learning", "quote": "Latent World Models. Latent world models shift predictive learning from reconstructing high-entropy observations to modeling task-related latent. Joint Embedding Predictive Architecture (JEPA)-style work established this path by predicting…", "summary": "Latent World Models. Latent world models shift predictive learning from reconstructing high-entropy observations to modeling task-related latent. Joint Embedding Predictive Archit…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0081", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "B.1. Self-Supervised Learning"}}}
{"evidence_id": "Orca.ev_0082", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "B.1. Self-Supervised Learning", "quote": "V-JEPA (Bardes et al., 2024) demonstrates that robust video representations can be learned solely from feature predictions, without pixel reconstruction. V-JEPA 2 (Assran et al., 2025) combines large-scale internet video pre-training with…", "summary": "V-JEPA (Bardes et al., 2024) demonstrates that robust video representations can be learned solely from feature predictions, without pixel reconstruction. V-JEPA 2 (Assran et al.,…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0082", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "B.1. Self-Supervised Learning"}}}
{"evidence_id": "Orca.ev_0083", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "B.1. Self-Supervised Learning", "quote": "How Orca differs. JEPA-style models demonstrate the effectiveness of latent prediction for self-supervised visual representation learning, which is closely related to Orca's observation-only state transition. Orca starts from a broader wor…", "summary": "How Orca differs. JEPA-style models demonstrate the effectiveness of latent prediction for self-supervised visual representation learning, which is closely related to Orca's obser…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0083", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "B.1. Self-Supervised Learning"}}}
{"evidence_id": "Orca.ev_0084", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "B.2. Next Token Prediction", "quote": "Large Language Models. Since the development of autoregressive large language models (Guo et al., 2025;Touvron et al., 2023), recent representative works have further pushed the paradigm along several directions. LLaMA 3.1 (Meta-AI, 2024)…", "summary": "Large Language Models. Since the development of autoregressive large language models (Guo et al., 2025;Touvron et al., 2023), recent representative works have further pushed the p…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0084", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "B.2. Next Token Prediction"}}}
{"evidence_id": "Orca.ev_0085", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "B.2. Next Token Prediction", "quote": "Multimodal Large Language Models. In the context of multimodal environments, many works have emerged. These include instruction-tuned visual language models, native multimodal foundation models, and agents. LLaVA (Liu et al., 2023) pioneer…", "summary": "Multimodal Large Language Models. In the context of multimodal environments, many works have emerged. These include instruction-tuned visual language models, native multimodal fou…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0085", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "B.2. Next Token Prediction"}}}
{"evidence_id": "Orca.ev_0086", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "B.2. Next Token Prediction", "quote": "The unified multimodal models also begin to blur the boundary between token prediction and frame prediction. For example, Emu3 (Wang et al., 2026a) and Emu3.5 (Cui et al., 2025) unify multimodality under the \"next-token-prediction\" paradig…", "summary": "The unified multimodal models also begin to blur the boundary between token prediction and frame prediction. For example, Emu3 (Wang et al., 2026a) and Emu3.5 (Cui et al., 2025) u…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0086", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "B.2. Next Token Prediction"}}}
{"evidence_id": "Orca.ev_0087", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "B.2. Next Token Prediction", "quote": "How Orca differs. Next-token models organize knowledge and reasoning through autoregressive language modeling. Orca uses language as an explicit semantic condition for state transition: language can specify events, task intentions, and cau…", "summary": "How Orca differs. Next-token models organize knowledge and reasoning through autoregressive language modeling. Orca uses language as an explicit semantic condition for state trans…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0087", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "B.2. Next Token Prediction"}}}
{"evidence_id": "Orca.ev_0088", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "B.3. Next Frame Prediction", "quote": "Image Generation Models. From a macro perspective, the representative image generation models can be viewed as frame-level prediction models. These models map language or multimodal conditions to target visual observations, thus extending…", "summary": "Image Generation Models. From a macro perspective, the representative image generation models can be viewed as frame-level prediction models. These models map language or multimod…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0088", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "B.3. Next Frame Prediction"}}}
{"evidence_id": "Orca.ev_0089", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "B.4. Next Action Prediction", "quote": "Vision Language Action Models. Vision Language Action (VLA) models, an architecture in embodied intelligence, provide a feasible path to improve generalization and multi-task learning capabilities (Bai et al., 2025b(Bai et al., , 2026;;Lyu…", "summary": "Vision Language Action Models. Vision Language Action (VLA) models, an architecture in embodied intelligence, provide a feasible path to improve generalization and multi-task lear…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0089", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "B.4. Next Action Prediction"}}}
{"evidence_id": "Orca.ev_0090", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "B.4. Next Action Prediction", "quote": "How Orca differs. VLA and world-action models usually organize embodied learning around action prediction, policy learning, or joint video-action modeling. Orca follows a world-learning-first philosophy: it first learns how scenes and obje…", "summary": "How Orca differs. VLA and world-action models usually organize embodied learning around action prediction, policy learning, or joint video-action modeling. Orca follows a world-le…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0090", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "B.4. Next Action Prediction"}}}
{"evidence_id": "Orca.ev_0091", "paper_id": "orca_2026_07", "source_type": "paper_text", "page": null, "section": "B.4. Next Action Prediction", "quote": "Table E3. Detailed rule-based results under real-robot OOD settings.", "summary": "Table E3. Detailed rule-based results under real-robot OOD settings.", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0091", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "B.4. Next Action Prediction"}}}
{"evidence_id": "Orca.ev_0092", "paper_id": "orca_2026_07", "source_type": "figure_caption", "page": null, "section": "", "quote": "Intro | Sec 2: Orca | Sec 3: Training | Sec 4: Evaluation | Sec 5: Conclusion | Sec 6: Authors | References Appendix A: Conception | B: Related Work | C: Train Settings | D: Infra | E: Eval Settings | F: Visualization", "summary": "Intro | Sec 2: Orca | Sec 3: Training | Sec 4: Evaluation | Sec 5: Conclusion | Sec 6: Authors | References Appendix A: Conception | B: Related Work | C: Train Settings | D: Infra…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0092", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"figure": "Sec 1 :"}}}
{"evidence_id": "Orca.ev_0093", "paper_id": "orca_2026_07", "source_type": "figure_caption", "page": null, "section": "", "quote": "Figure 3. Overview of pre-training data. Orca's pre-training data includes video, event, and VQA data. A. Video Data supports 1) Observation-only state transition, A. Video Data and B. Event Data support 2) Event-conditioned state transiti…", "summary": "Figure 3. Overview of pre-training data. Orca's pre-training data includes video, event, and VQA data. A. Video Data supports 1) Observation-only state transition, A. Video Data a…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0093", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"figure": "Figure 3 ."}}}
{"evidence_id": "Orca.ev_0094", "paper_id": "orca_2026_07", "source_type": "figure_caption", "page": null, "section": "", "quote": "Intro | Sec 2: Orca | Sec 3: Training | Sec 4: Evaluation | Sec 5: Conclusion | Sec 6: Authors | References Appendix A: Conception | B: Related Work | C: Train Settings | D: Infra | E: Eval Settings | F: Visualization", "summary": "Intro | Sec 2: Orca | Sec 3: Training | Sec 4: Evaluation | Sec 5: Conclusion | Sec 6: Authors | References Appendix A: Conception | B: Related Work | C: Train Settings | D: Infra…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0094", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"figure": "Sec 1 :"}}}
{"evidence_id": "Orca.ev_0095", "paper_id": "orca_2026_07", "source_type": "figure_caption", "page": null, "section": "", "quote": "Figure 4. Downstream readout architectures. To language reuses the LM head for text readout. To vision only trains an MLP adaptor and LoRA on top of a frozen SD3.5 to readout images. To action trains an MLP adaptor and a DiT-based Action E…", "summary": "Figure 4. Downstream readout architectures. To language reuses the LM head for text readout. To vision only trains an MLP adaptor and LoRA on top of a frozen SD3.5 to readout imag…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0095", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"figure": "Figure 4 ."}}}
{"evidence_id": "Orca.ev_0096", "paper_id": "orca_2026_07", "source_type": "figure_caption", "page": null, "section": "", "quote": "Intro | Sec 2: Orca | Sec 3: Training | Sec 4: Evaluation | Sec 5: Conclusion | Sec 6: Authors | References Appendix A: Conception | B: Related Work | C: Train Settings | D: Infra | E: Eval Settings | F: Visualization 3.2.3. To Action: Act…", "summary": "Intro | Sec 2: Orca | Sec 3: Training | Sec 4: Evaluation | Sec 5: Conclusion | Sec 6: Authors | References Appendix A: Conception | B: Related Work | C: Train Settings | D: Infra…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0096", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"figure": "Sec 1 :"}}}
{"evidence_id": "Orca.ev_0097", "paper_id": "orca_2026_07", "source_type": "figure_caption", "page": null, "section": "", "quote": "With these optimizations, training throughput increases from 0.66 to 2.91 Samples/Sec/GPU, achieving approximately a 4.4× acceleration compared to the StarVLA (StarVLA Community, 2026) commonly used in the embodied community. Optimization…", "summary": "With these optimizations, training throughput increases from 0.66 to 2.91 Samples/Sec/GPU, achieving approximately a 4.4× acceleration compared to the StarVLA (StarVLA Community,…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0097", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"figure": "fig_6"}}}
{"evidence_id": "Orca.ev_0098", "paper_id": "orca_2026_07", "source_type": "figure_caption", "page": null, "section": "", "quote": "Figure 5. Loss of model and data scaling.", "summary": "Figure 5. Loss of model and data scaling.", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0098", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"figure": "Figure 5 ."}}}
{"evidence_id": "Orca.ev_0099", "paper_id": "orca_2026_07", "source_type": "figure_caption", "page": null, "section": "", "quote": "Figure 6. Scaling behavior on downstream readouts performance.", "summary": "Figure 6. Scaling behavior on downstream readouts performance.", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0099", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"figure": "Figure 6 ."}}}
{"evidence_id": "Orca.ev_0100", "paper_id": "orca_2026_07", "source_type": "figure_caption", "page": null, "section": "", "quote": "The downstream readout performance across text generation, image prediction, and action generation. The text generation performance is the average performance on four benchmarks: TemporalBench, MVBench, SWITCH, and 3DRSBench. The details a…", "summary": "The downstream readout performance across text generation, image prediction, and action generation. The text generation performance is the average performance on four benchmarks:…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0100", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"figure": "fig_9"}}}
{"evidence_id": "Orca.ev_0101", "paper_id": "orca_2026_07", "source_type": "figure_caption", "page": null, "section": "", "quote": "Text Generation. It demonstrates the model's out-of-distribution (OOD) commonsense reasoning, comprehension capabilities, and high-level cognitive abilities. 2. Image Prediction. It visualizes this latent cognitive capability through OOD s…", "summary": "Text Generation. It demonstrates the model's out-of-distribution (OOD) commonsense reasoning, comprehension capabilities, and high-level cognitive abilities. 2. Image Prediction.…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0101", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"figure": "1."}}}
{"evidence_id": "Orca.ev_0102", "paper_id": "orca_2026_07", "source_type": "figure_caption", "page": null, "section": "", "quote": "Intro | Sec 2: Orca | Sec 3: Training | Sec 4: Evaluation | Sec 5: Conclusion | Sec 6: Authors | References Appendix A: Conception | B: Related Work | C: Train Settings | D: Infra | E: Eval Settings | F: Visualization", "summary": "Intro | Sec 2: Orca | Sec 3: Training | Sec 4: Evaluation | Sec 5: Conclusion | Sec 6: Authors | References Appendix A: Conception | B: Related Work | C: Train Settings | D: Infra…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0102", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"figure": "Sec 1 :"}}}
{"evidence_id": "Orca.ev_0103", "paper_id": "orca_2026_07", "source_type": "figure_caption", "page": null, "section": "", "quote": "Figure 7. Visual comparison of image prediction in the real world.", "summary": "Figure 7. Visual comparison of image prediction in the real world.", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0103", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"figure": "Sec 1 :"}}}
{"evidence_id": "Orca.ev_0104", "paper_id": "orca_2026_07", "source_type": "figure_caption", "page": null, "section": "", "quote": "Intro | Sec 2: Orca | Sec 3: Training | Sec 4: Evaluation | Sec 5: Conclusion | Sec 6: Authors | References Appendix A: Conception | B: Related Work | C: Train Settings | D: Infra | E: Eval Settings | F: Visualization", "summary": "Intro | Sec 2: Orca | Sec 3: Training | Sec 4: Evaluation | Sec 5: Conclusion | Sec 6: Authors | References Appendix A: Conception | B: Related Work | C: Train Settings | D: Infra…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0104", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"figure": "Sec 1 :"}}}
{"evidence_id": "Orca.ev_0105", "paper_id": "orca_2026_07", "source_type": "figure_caption", "page": null, "section": "", "quote": "Figure 8. Recovery after repeated grasp failures. Orca recovers from early spoon-grasp failures and eventually makes progress, while 𝜋 0.5 remains unstable with repeated failed attempts.", "summary": "Figure 8. Recovery after repeated grasp failures. Orca recovers from early spoon-grasp failures and eventually makes progress, while 𝜋 0.5 remains unstable with repeated failed at…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0105", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"figure": "Figure 8 ."}}}
{"evidence_id": "Orca.ev_0106", "paper_id": "orca_2026_07", "source_type": "figure_caption", "page": null, "section": "", "quote": "Intro | Sec 2: Orca | Sec 3: Training | Sec 4: Evaluation | Sec 5: Conclusion | Sec 6: Authors | References Appendix A: Conception | B: Related Work | C: Train Settings | D: Infra | E: Eval Settings | F: Visualization", "summary": "Intro | Sec 2: Orca | Sec 3: Training | Sec 4: Evaluation | Sec 5: Conclusion | Sec 6: Authors | References Appendix A: Conception | B: Related Work | C: Train Settings | D: Infra…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0106", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"figure": "Sec 1 :"}}}
{"evidence_id": "Orca.ev_0107", "paper_id": "orca_2026_07", "source_type": "figure_caption", "page": null, "section": "", "quote": "Intro | Sec 2: Orca | Sec 3: Training | Sec 4: Evaluation | Sec 5: Conclusion | Sec 6: Authors | References Appendix A: Conception | B: Related Work | C: Train Settings | D: Infra | E: Eval Settings | F: Visualization Sec 1: Intro | Sec 2:…", "summary": "Intro | Sec 2: Orca | Sec 3: Training | Sec 4: Evaluation | Sec 5: Conclusion | Sec 6: Authors | References Appendix A: Conception | B: Related Work | C: Train Settings | D: Infra…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0107", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"figure": "Sec 1 :"}}}
{"evidence_id": "Orca.ev_0108", "paper_id": "orca_2026_07", "source_type": "figure_caption", "page": null, "section": "", "quote": "Intro | Sec 2: Orca | Sec 3: Training | Sec 4: Evaluation | Sec 5: Conclusion | Sec 6: Authors | References Appendix A: Conception | B: Related Work | C: Train Settings | D: Infra | E: Eval Settings | F: Visualization", "summary": "Intro | Sec 2: Orca | Sec 3: Training | Sec 4: Evaluation | Sec 5: Conclusion | Sec 6: Authors | References Appendix A: Conception | B: Related Work | C: Train Settings | D: Infra…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0108", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"figure": "Sec 1 :"}}}
{"evidence_id": "Orca.ev_0109", "paper_id": "orca_2026_07", "source_type": "figure_caption", "page": null, "section": "", "quote": "Intro | Sec 2: Orca | Sec 3: Training | Sec 4: Evaluation | Sec 5: Conclusion | Sec 6: Authors | References Appendix A: Conception | B: Related Work | C: Train Settings | D: Infra | E: Eval Settings | F: Visualization", "summary": "Intro | Sec 2: Orca | Sec 3: Training | Sec 4: Evaluation | Sec 5: Conclusion | Sec 6: Authors | References Appendix A: Conception | B: Related Work | C: Train Settings | D: Infra…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0109", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"figure": "Sec 1 :"}}}
{"evidence_id": "Orca.ev_0110", "paper_id": "orca_2026_07", "source_type": "figure_caption", "page": null, "section": "", "quote": "combines a video model and an action decoder, demonstrating strong generalization ability and sample efficiency. Cosmos-Policy (Kim et al., 2026) adapts the pre-trained Cosmos-Predict (NVIDIA et al., 2025a) into a robot policy through sing…", "summary": "combines a video model and an action decoder, demonstrating strong generalization ability and sample efficiency. Cosmos-Policy (Kim et al., 2026) adapts the pre-trained Cosmos-Pre…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0110", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"figure": "fig_19"}}}
{"evidence_id": "Orca.ev_0111", "paper_id": "orca_2026_07", "source_type": "figure_caption", "page": null, "section": "", "quote": "Intro | Sec 2: Orca | Sec 3: Training | Sec 4: Evaluation | Sec 5: Conclusion | Sec 6: Authors | References Appendix A: Conception | B: Related Work | C: Train Settings | D: Infra | E: Eval Settings | F: Visualization", "summary": "Intro | Sec 2: Orca | Sec 3: Training | Sec 4: Evaluation | Sec 5: Conclusion | Sec 6: Authors | References Appendix A: Conception | B: Related Work | C: Train Settings | D: Infra…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0111", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"figure": "Sec 1 :"}}}
{"evidence_id": "Orca.ev_0112", "paper_id": "orca_2026_07", "source_type": "table", "page": null, "section": "", "quote": "Noisy Action + Time", "summary": "Noisy Action + Time", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0112", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"table": "Image Xt+d Query1 Text Query2 MLP MLP Text Query2 Teacher Forcing Image Xt+1 Latent Vit Space Teacher Forcing Image Xt+d Latent Vit Space Latent Representation Latent Representation Orca Encoder 🔥 ❄ Frozen Trainable ❄ ❄ 🔥 🔥 🔥 🔥 World Latent Representation❄ World Latent Representation❄ Noise"}}}
{"evidence_id": "Orca.ev_0113", "paper_id": "orca_2026_07", "source_type": "table", "page": null, "section": "", "quote": "Orca achieves the best overall result among the same-size VLMs and the large-size world models, demonstrating the advantages of the proposed learning paradigm.", "summary": "Orca achieves the best overall result among the same-size VLMs and the large-size world models, demonstrating the advantages of the proposed learning paradigm.", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0113", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"table": "Table 1 ,"}}}
{"evidence_id": "Orca.ev_0114", "paper_id": "orca_2026_07", "source_type": "table", "page": null, "section": "", "quote": "The comparison of the text generation. ↑ represents the higher value, the better. World Models (Large size) V-JEPA 2.1 1 (+LLaMA3-8B) | 10 | 75.4 | 28.5 | / | / | / Emu3 2 | 8 | 35.2 | 9.5 | 39.1 | 38.0 | 30.4 Emu3.5 | 34 | 39.5 | 9.5 | 31…", "summary": "The comparison of the text generation. ↑ represents the higher value, the better. World Models (Large size) V-JEPA 2.1 1 (+LLaMA3-8B) | 10 | 75.4 | 28.5 | / | / | / Emu3 2 | 8 | 3…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0114", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"table": "Table 1 ."}}}
{"evidence_id": "Orca.ev_0115", "paper_id": "orca_2026_07", "source_type": "table", "page": null, "section": "", "quote": "The cross-benchmark general capability comparison of the text generation. Model | State Transition 1 Commonsense Reasoning 2 Spatial Relations 3 Dynamic Motion 4 Qwen3.5-4B | 51.86 | 57.76 | 54.68 | 57.03 Orca-4B | 64.13 (+12.27%) | 62.95…", "summary": "The cross-benchmark general capability comparison of the text generation. Model | State Transition 1 Commonsense Reasoning 2 Spatial Relations 3 Dynamic Motion 4 Qwen3.5-4B | 51.8…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0115", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"table": "Table 2 ."}}}
{"evidence_id": "Orca.ev_0116", "paper_id": "orca_2026_07", "source_type": "table", "page": null, "section": "", "quote": "The comparison of the PRICE-V0.1. ↑ represents the higher value, the better. In Avg., a±b is avg±std. A larger avg and a smaller std value represent a better result. Bold represents the best value. Model | Size (B) Gemini 3.1 Pro ↑ GPT 5.4…", "summary": "The comparison of the PRICE-V0.1. ↑ represents the higher value, the better. In Avg., a±b is avg±std. A larger avg and a smaller std value represent a better result. Bold represen…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0116", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"table": "Table 3 ."}}}
{"evidence_id": "Orca.ev_0117", "paper_id": "orca_2026_07", "source_type": "table", "page": null, "section": "", "quote": "Ablation results. \"-\" means the setting does not work. The first three rows average two metrics, while the last two average all three. Text, image, and action denote the average scores on text benchmarks, PRICE-V0.1, and overall rule-based…", "summary": "Ablation results. \"-\" means the setting does not work. The first three rows average two metrics, while the last two average all three. Text, image, and action denote the average s…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0117", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"table": "Table 5 ."}}}
{"evidence_id": "Orca.ev_0118", "paper_id": "orca_2026_07", "source_type": "table", "page": null, "section": "", "quote": "boasts powerful multilingual text rendering, infographic creation, and search-based visualization capabilities. GPT Image 2 (OpenAI, 2026a) extends the image generation product line to native text and image model interfaces, supporting hig…", "summary": "boasts powerful multilingual text rendering, infographic creation, and search-based visualization capabilities. GPT Image 2 (OpenAI, 2026a) extends the image generation product li…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0118", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"table": "tab_8"}}}
{"evidence_id": "Orca.ev_0119", "paper_id": "orca_2026_07", "source_type": "table", "page": null, "section": "", "quote": ".OpenVLA (Kim et al., 2024), which first presented an open-source VLA model built on a pre-trained VLM, showing strong cross-embodied general manipulation capabilities after training on large-scale robot data. 𝜋 0.5(Physical Intelligence e…", "summary": ".OpenVLA (Kim et al., 2024), which first presented an open-source VLA model built on a pre-trained VLM, showing strong cross-embodied general manipulation capabilities after train…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0119", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"table": "tab_9"}}}
{"evidence_id": "Orca.ev_0120", "paper_id": "orca_2026_07", "source_type": "table", "page": null, "section": "", "quote": "Scoring criteria for real-robot evaluation. Each task is evaluated within 60 seconds. If the robot becomes locked due to severe collision, or if the object falls and the task can no longer continue, evaluation is stopped. For each task, on…", "summary": "Scoring criteria for real-robot evaluation. Each task is evaluated within 60 seconds. If the robot becomes locked due to severe collision, or if the object falls and the task can…", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0120", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"table": "Table E2 ."}}}
{"evidence_id": "Orca.ev_0121", "paper_id": "orca_2026_07", "source_type": "formula", "page": null, "section": "Pre-Training Recipe", "quote": "L = 𝜆 obs L obs + 𝜆 evt L evt + 𝜆 vqa L vqa , (2)", "summary": "L = 𝜆 obs L obs + 𝜆 evt L evt + 𝜆 vqa L vqa , (2)", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0121", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Pre-Training Recipe"}}}
{"evidence_id": "Orca.ev_0122", "paper_id": "orca_2026_07", "source_type": "formula", "page": null, "section": "Results and Analysis. Based on Table 4, we obtained two conclusions:", "quote": "Environment OOD Rule-based ↑ M25 ↑ 1 M50 ↑ 1 SR ↑ 2 MaxP-F ↑ 3 FNS ↑ 4 DRR ↑ 5 SQS ↑ 6 V-JEPA", "summary": "Environment OOD Rule-based ↑ M25 ↑ 1 M50 ↑ 1 SR ↑ 2 MaxP-F ↑ 3 FNS ↑ 4 DRR ↑ 5 SQS ↑ 6 V-JEPA", "relevance": "source_evidence", "meta": {"original_ev_id": "ev_0122", "evidence_file": "Orca_The_World_is_in_Your_Mind.evidence_index.jsonl", "locator": {"section": "Results and Analysis. Based on Table 4, we obtained two conclusions:"}}}