{"id":"65991207-0b02-4b9d-bd95-ad3c47d3380f","arxiv_id":"2610.37666","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"ThinkWorld grounds VLA action reasoning in discrete motion tokens and uses imagined-versus-observed future agreement as an RL reward, enabling unlabeled human videos to improve robot policies.","lead":"ThinkWorld trains a robot-control model to reason with short sequences of motion tokens, called a Chain of Dynamics, and rewards reasoning whose imagined one-second-later scene matches what the video actually shows. This lets ordinary human videos, which contain no robot action labels, contribute to training robot policies.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"The load-bearing link is Eq. (3)'s future-agreement reward, but Table 2 shows the reward signal is thin: grounded versus empty CoD differ by only 0.02 cosine, so non-gameability of the unlabeled-video supervision is not established.","rationale":"The reader's weakest_assumption identifies the same load-bearing concern, and I find no separate objection that changes the verdict. The central claim depends on Eq. (3) as the channel through which unlabeled video supervises action reasoning, so the reward's discriminative power is the most security-critical assumption. Table 2's small margin between grounded and empty CoD predictions is internal evidence that the reward may be dominated by content that is not action-specific. This does not disprove the method: the real-robot trial denominators (Table 6), the matched-scale robot-only control (Fig. 6), the cumulative ablation (Table 3), and the explicit single-seed caveat (Appendix D) are real supporting evidence. But those results do not isolate whether the RL reward, rather than the SFT-stage future-prediction objective and CoD grounding, is what makes unlabeled video useful. The paper's own numbers show a modest +1.2 percentage-point effect from 50% human video and +1.2 points from RL, both within single-seed uncertainty. Because the reader already placed the paper at CONDITIONAL for this reason, the appropriate verdict is unchanged; the condition should be resolved by an independent reward-discrimination test before acceptance.","tokens_in":27634,"tokens_out":6343,"duration_ms":56726,"concrete_test":"Take a held-out set of transitions with known action labels (robot trajectories plus EgoDex/EgoVerse hand-trajectory annotations, excluded from SFT and RL). For each transition, compute r_world from Eq. (6) with the frozen projector for: (i) the ground-truth CoD, (ii) token-shuffled CoDs, (iii) empty CoDs, and (iv) a camera-motion-only baseline CoD that predicts the future from current-frame ego-motion. Measure the rank separation: the fraction of shuffled, empty, and camera-motion CoDs whose reward exceeds the ground-truth CoD's reward, separately for robot and human-ego videos. If that fraction is large, or if the camera-motion baseline's reward distribution overlaps the ground-truth distribution in human videos, Eq. (3)'s reward cannot support the claimed supervision. Repeating the check before and after RL would also show whether RL improves the discriminative margin.","verdict_should_be":"UNCHANGED","load_bearing_attack":"ThinkWorld's central claim requires that r_world in Eq. (3) reward action-correct CoDs on unlabeled human video. That requires the frozen future projector, trained on the same CoD tokens during SFT (Eqs. 1-2), to be sensitive to CoD content, and requires the one-second DINO change in egocentric video to be dominated by demonstrated motion rather than camera or appearance change. Table 2 is the paper's own evidence on this point, and it is the weakest link: with CoD on, cosine similarity to the true future is 0.74; with CoD off it is 0.72, and empty/shuffle CoDs reach 0.72/0.71. The RL reward in Eq. (6) averages global and motion-weighted patch agreement, so the effective margin between a grounded and an ungrounded CoD is only about 0.02-0.03 cosine. The synthetic 0.05 penalty for malformed CoDs (Appendix A.6) does not repair this margin. If r_world mostly scores static scene structure or ego-motion changes, the RL stage can reinforce CoDs that predict appearance or camera shifts without grounding in controllable behavior; that is precisely the failure mode the paper assigns to latent action models in Section 2. The attention analysis is explicitly descriptive and non-causal, so it does not close the gap.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"ThinkWorld is a VLA training framework whose central proposal is to represent an action as a Chain-of-Dynamics (CoD), a sequence of discrete motion tokens shared across human and robot embodiments. In stage (a), supervised fine-tuning grounds CoD tokens in robot actions and annotated human hand trajectories, and trains a future projector that, conditioned on the current observation and a CoD, predicts DINOv3 features of the observed future one second ahead. In stage (b), the model samples alternative CoDs and uses the similarity between the imagined future and the observed future as a reward in GRPO, with a small supervised replay term. Because the reward needs only an observed state transition, the paper claims that action-unlabeled human videos can directly supervise action reasoning. At deployment the CoD and future prediction are discarded and only a continuous action expert is used. Experiments report real-robot success on 120 trials across three bimanual tasks (macro 94.4%, micro 93.3%), qualitative human-to-robot skill composition on oven and drawer tasks, 97.3% average on LIBERO, 55.6% on LIBERO-PRO, and 71.7% on LIBERO-Plus, together with ablations on training components and on the fraction of human video.","tokens_in":27929,"tokens_out":7669,"duration_ms":63455,"significance":"If the proposed mechanism works as stated, the paper makes a significant contribution: it offers a concrete route for large-scale human video to provide training signal for robot action reasoning without action annotations, while keeping inference cost unchanged. The empirical base is real and unusually well disclosed: trial denominators are given for all 120 physical trials (Table 6), a matched-scale robot-only control separates human-data benefit from training-scale effects (Figure 6), a cumulative component ablation is provided (Table 3), and LIBERO-PRO/LIBERO-Plus results are reported with baselines. The core mechanism, however, rests on the assumption that the world-imagination reward in Eq. (3) is a faithful and non-gameable measure of action correctness on unlabeled transitions, and that assumption is not yet established by the evidence in the manuscript. The headline numbers are strong, but they do not by themselves demonstrate the unlabeled-video supervision pathway that distinguishes ThinkWorld from prior work.","major_comments":[{"comment":"The faithfulness of the world-imagination reward is the load-bearing assumption of the paper, and it remains unresolved. Table 2 shows all-patch cosine similarity of 0.71–0.72 for shuffled, empty, and no-CoD conditions versus 0.74 for CoD on; the top-25% motion-patch margin is larger (0.54–0.55 versus 0.62), which partly mitigates the raw 0.02 all-patch gap, but Eq. (6) averages the global term with the motion-weighted term and the paper reports no distribution of r_world over sampled CoD candidates in either robot or human transitions. Because the future projector is frozen during RL and was trained on the same CoD tokens and the same observed futures (Eqs. (1)–(2)), a projector that relies on static scene structure or generic ego-motion would assign high reward to trivial or empty CoDs, and the malformed-CoD penalty in Appendix A.6 does not repair that failure mode. I would like to see reward-level sensitivity statistics (mean, standard deviation, and margin between grounded and ungrounded candidates), a correlation of the reward with an independent measure of action correctness, and an RL ablation in which the reward is replaced by a static-future baseline or by a projector deliberately blinded to CoD tokens.","section":"§3.3, Eq. (3), Table 2, Appendix A.6","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The claim that videos without action annotations provide direct supervision is not isolated by any experiment. The human data used during SFT appear to carry motion annotations (Section 3.2 describes an 'annotated human hand trajectory', and Table 4 lists human motion tokens), so the human-data scaling results in Figure 6 test labeled human motion rather than action-unlabeled video. The RL stage is the only place where action-free transitions enter, but its measured contribution is +1.2 points on LIBERO-PRO (Table 3, 54.4% to 55.6%), and no variant varies the amount of action-unlabeled human video in RL while holding the SFT data fixed. The skill-composition result is also reported only qualitatively, without trial denominators for the composed oven and drawer rollouts. I would request a controlled experiment that trains with and without RL on human transitions that have no CoD targets, plus quantitative success counts for the composition tasks.","section":"§3.3, Figure 6, Table 3, Appendix B.3","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The causal role of CoD in future prediction is demonstrated at the level of conditioning, but the causal role of the future-imagination reward in improving action reasoning is not demonstrated. Table 2 and Figure 7 measure future-prediction quality under input manipulations, while the attention and PCA analyses are explicitly descriptive and non-causal by the authors' own statement in Appendix C.2. Since the RL update optimizes the frozen projector's similarity score, the +1.2 point gain after RL in Table 3 could in principle come from improved sequence regularities or language priors rather than from better action-grounded CoDs. I would like to see optimization-dynamics evidence: whether average r_world on held-out human transitions increases after RL, whether rewards of grounded versus trivial sampled CoDs diverge, and whether the improvement is concentrated in motion-relevant patches rather than in static scene regions.","section":"§4.5, Table 2, Figure 7, Appendix C.2","authors_acknowledge":"PCA and attention are descriptive analyses and are not used as causal evidence.","authors_acknowledge_p":0.73}],"minor_comments":[{"comment":"The footnote says 'LIBERO-PRO axis values are arithmetic means over Goal, Spatial, LIBERO-10, and Object,' but the LIBERO-PRO columns are Object, Position, Semantic, Task, and Environment; the footnote should be corrected or clarified.","section":"Table 1, footnote","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"Three separate figures in Appendix B.3/C.1 are all numbered 'Figure 10' with identical captions for color matching, cup hanging, and sorting; they should be renumbered and given distinct captions.","section":"Figures 10–12","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The sentence 'revealing a shared action-level structure shared across human and robot behaviors' contains a duplicated 'shared' and should be reworded.","section":"§4.5, Cross-Embodiment Structure of CoD","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The trainable-modules row lists 'Language model + future projector' for VLM SFT, but Section 3.2 and Eq. (2) describe the future projector as a separate decoder; it would help to state explicitly which components are frozen versus updated in each phase, including the vision tower.","section":"Appendix A.4, Table 5","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"major_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"Two things before you read it. The new mechanism is real: ThinkWorld tokenizes human and robot end-effector motion into one discrete motion space (Chain-of-Dynamics), trains a VLM to emit CoDs and predict the DINO-space future each CoD induces, then runs GRPO with a reward equal to imagined-versus-observed future agreement. That combination — a shared motion-token reasoning interface plus future-consistency RL over unlabeled video — is not in the cited literature, and it attacks the action-label bottleneck rather than just scaling robot data. The reporting is better than the field average: 120 real-robot trials with denominators (Table 6), a matched-scale robot-only control (Fig. 6), a cumulative ablation (Table 3), and an explicit single-seed caveat (Appendix D). Real-robot results: 94.4 macro / 93.3 micro success, versus 62.2/60.0 for the best baseline.\n\nThe load-bearing claim is that the world-imagination reward turns unlabeled human video into action supervision. The evidence is the paper's thinnest link, and it's the paper's own table that shows it: an empty CoD reaches 0.72 all-patch cosine to the true future, a grounded CoD 0.74. The stress-test note is right to flag this, though partly overstated. In the paper's favor: the motion-patch margin is larger (0.62 vs 0.54–0.55), the reward weights exactly those patches, and GRPO standardizes within a group of four candidates, so relative ranking matters more than the raw 0.02. Against it: the projector is frozen during RL but was trained on the same CoD tokens and futures it now scores, so the reward's independence is not established — the failure mode the paper assigns to latent action models in Section 2 is a live risk for its own signal. This is the question to push hardest on.\n\nSecond, the skill-composition demo (Sec. 4.3), the most distinctive result, is qualitative only: successful rollouts, no trial counts. A claim about executing a complete sequence never demonstrated in robot data needs a denominator.\n\nThird, the headline benchmark rests on one training seed (acknowledged), and the two headline mechanisms have small measured effects: +1.2 points from RL (54.4 → 55.6) and +1.2 points over the matched robot-only control. Modest, but the monotonic human-video trend is suggestive. LIBERO-Plus is mixed: strong on layout and initial-state shifts, weak on camera and noise.\n\nFor whom: people building VLAs, learning from human video, or using world models as reward sources. It deserves a serious referee, and the fixes are concrete — composition trial counts, a second seed, and a reward-gameability check (oracle vs malformed CoD rewards on held-out human video). I'd send it out.","headline":"A genuinely new cross-embodiment motion-token reasoning interface with honest, thorough evaluation — but the future-agreement reward carrying the unlabeled-video claim has thin, partly self-referential support, and the skill-composition demo lacks trial counts.","tokens_in":28629,"tokens_out":10601,"duration_ms":71776},"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek/deepseek-v4-flash","headline":"ThinkWorld claims that unlabeled human video can refine robot action reasoning through future imagination, with action thoughts rewarded by how well their imagined futures match the observed next frame.","keywords":["vision-language-action models","robotic manipulation","chain-of-dynamics","world imagination","reinforcement learning","human video learning","future-state prediction","cross-embodiment transfer"],"falsifier":"Run the same two-stage training on egocentric human clips in which the camera moves or the scene's appearance changes but no task-relevant manipulation occurs within the one-second horizon; if the world-imagination reward still ranks such Chains-of-Dynamics highly and downstream generalization gains persist, the signal is crediting appearance or camera alignment rather than demonstrated control. A direct check would compare reward-ranked Chains-of-Dynamics against ground-truth wrist motion on a held-out annotated set and ask whether the top-rewarded candidates are also the ones closest to the demonstrated motion.","tokens_in":28974,"feed_emoji":"🤖","tokens_out":12073,"duration_ms":90518,"temperature":0.7,"pith_summary":"ThinkWorld claims that a robot policy can learn to reason about an action as a Chain-of-Dynamics: a short sequence of discrete motion tokens describing how the world changes under that action, shared between human and robot embodiments. The paper grounds this token space in action-labeled demonstrations, then adds a world-imagination stage in which the model predicts the future state its Chain-of-Dynamics would produce and is rewarded when that imagined future matches the observed next frame. Because the reward needs only temporally ordered video, ordinary human footage becomes a training signal for action reasoning even though it carries no executable actions. On that basis the paper reports 97.3% average success on the standard LIBERO benchmark, 55.6% on the harder LIBERO-PRO variant, 94.4% macro and 93.3% micro success rates across real-robot trials, and composition of complete oven and drawer tasks whose full sequences were never shown in robot data.","feed_headline":"Imagined futures turn unlabeled human video into robot training signal","feed_subtitle":"Rewarding action chains for correctly imagining the next second lets human video teach robot skills.","key_machinery":"The central object is the Chain-of-Dynamics: a sequence of discrete motion tokens produced by a motion tokenizer over normalized one-second delta end-effector trajectories, with one shared vocabulary used for human wrist motion and robot motion. Around it, the machinery is a future-imagination loop. Learnable future queries are appended after the Chain-of-Dynamics, and a decoder maps their contextualized hidden states to a predicted one-second-ahead feature map in the frozen DINO feature space, while the observed future frame, encoded by the same frozen encoder, supplies the target. The world-imagination reward is a motion-weighted cosine similarity between the imagined and observed future, placing extra weight on DINO patches whose features change most between the current and future frames. This reward is standardized within groups of sampled candidates and used in a group-relative policy-optimization update of the vision-language model, while the action expert and future projector stay frozen and a small supervised replay preserves motion grounding. The mechanism's role is to turn observed state transitions in unlabeled video into a gradient signal for the intermediate reasoning tokens, and through the shared vision-language model, into the representations used for continuous control.","core_discovery":"ThinkWorld's central claim is that the world itself can supervise an action thought. An action is represented as a Chain-of-Dynamics, a discrete sequence of motion tokens computed from one-second hand or end-effector deltas, which gives human motion and robot motion a common vocabulary without assuming kinematic equivalence. A supervised stage teaches the vision-language model to emit the Chain-of-Dynamics from the instruction and observations and, conditioned on that Chain-of-Dynamics, to predict a one-second-ahead future representation in a frozen self-supervised DINO feature space. A reinforcement-learning stage then samples several Chains-of-Dynamics for the same context, imagines each candidate's future, and rewards candidates whose imagined future agrees with the observed future, weighting agreement more heavily in the image patches that actually change. This reward does not require action labels, so unlabeled human video and robot demonstrations update the same reasoning space; the future predictor and the Chain-of-Dynamics are used only during training and dropped at deployment, leaving a continuous action expert that consumes the vision-language representations shaped by this process.","pith_inferences":["A natural extension the paper does not test is varying the imagination horizon: if one second is what makes human ego-video usable, then longer or shorter horizons should change the reward's discriminative power, and searching over that horizon could tune how much of a skill the Chain-of-Dynamics must encode.","As unlabeled video grows, the group-relative reward may saturate if most sampled Chains-of-Dynamics predict similar futures; tracking reward diversity across sampled candidates as data scale increases would reveal whether the signal stays discriminative.","Because the Chain-of-Dynamics vocabulary is normalized delta motion rather than joint positions, the same two-stage recipe should in principle transfer to other morphologies, such as legged or mobile manipulators, where the human-data gain could be measured directly."],"forward_implications":["Action-unlabeled human video becomes a directly usable training signal, so behavioral coverage can be scaled without collecting more embodiment-specific robot trajectories; the paper shows gains increase monotonically as the human-video fraction rises from 0% to 50% of the data.","Because Chain-of-Dynamics structure clusters by motion direction rather than by embodiment, compositional knowledge transfers across embodiments: a robot that has only learned atomic skills can execute a full human-demonstrated sequence such as open, put-left, close, never shown in robot data.","Since future imagination is discarded at deployment, the reasoning-time cost of world modeling is zero at inference; the deployed policy is not slowed by the Chain-of-Dynamics or future-prediction computations used during training.","The reward comes from ordinary state transitions, so any source of temporally ordered video, including human ego-video and robot teleoperation, can feed the same reasoning space, which is what the human-data scaling study demonstrates."],"supporting_citations":[{"why":"Supplies the motion-tokenizer architecture used to turn one-second human wrist and robot end-effector deltas into the discrete Chain-of-Dynamics vocabulary.","marker":"[57]"},{"why":"Supplies the action-expert initialization and the bimanual robot demonstration dataset, and names the baseline against which real-robot gains are measured.","marker":"[25]"},{"why":"Provides the frozen DINOv3 encoder whose patch features define the future state the model is trained to imagine.","marker":"[63]"},{"why":"Establishes the use of DINO feature spaces for future-state comparison, the representational choice carried into the imagination loss.","marker":"[84]"},{"why":"Is the distribution-shift benchmark whose changed objects, positions, task semantics, and environments carry the paper's main generalization claims.","marker":"[85]"},{"why":"Is the base manipulation benchmark whose standard training set is used to post-train the model before zero-shot evaluation on shift benchmarks.","marker":"[49]"},{"why":"Exemplifies the latent-action approach whose underconstrained variables, able to capture appearance or camera motion rather than behavior, frame the motivation for consequence-grounded supervision.","marker":"[76]"},{"why":"Supplies the large egocentric human-video corpus used in supervised fine-tuning and reinforcement learning, the data source behind the unlabeled-video scaling claims.","marker":"[58]"},{"why":"Supplies the second egocentric human-video corpus used to build the shared Chain-of-Dynamics space across embodiments.","marker":"[30]"}],"fun_headline_variants":["World imagination reinforces chain-of-dynamics reasoning for robot control","Rewarding action chains for correct imagined futures lets human video teach robots","Imagining one-second futures turns unlabeled human video into robot training signal","Chain-of-Dynamics: a shared motion token space lets human video refine robot skills","World imagination reinforces chain-of-dynamics reasoning from unlabeled video"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The load-bearing premise is that agreement, in a frozen self-supervised visual feature space, between a future imagined from a sampled Chain-of-Dynamics and the observed one-second-later frame is a faithful and non-gameable proxy for action correctness on unlabeled video; under that premise, the paper's own ablation shows the reward operates on a thin margin, with an empty Chain-of-Dynamics reaching 0.72 cosine similarity against 0.74 for a grounded one.","fun_headline_variants_meta":{"raw":{"variants":["World imagination reinforces chain-of-dynamics reasoning for robot control","Rewarding action chains for correct imagined futures lets human video teach robots","Imagining one-second futures turns unlabeled human video into robot training signal","Chain-of-Dynamics: a shared motion token space lets human video refine robot skills","World imagination reinforces chain-of-dynamics reasoning from unlabeled video"]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":0.000238,"raw_usage":{"cost":0.0002375646,"is_byok":false,"cost_details":{"upstream_inference_cost":0.0002375646,"upstream_inference_prompt_cost":2.66266e-05,"upstream_inference_completions_cost":0.000210938},"total_tokens":2174,"prompt_tokens":1015,"completion_tokens":1159,"prompt_tokens_details":{"audio_tokens":0,"video_tokens":0,"cached_tokens":903,"cache_write_tokens":0},"completion_tokens_details":{"audio_tokens":0,"image_tokens":0,"reasoning_tokens":1079}},"tokens_in":1015,"tokens_out":1159,"cost_source":"openrouter_usage","duration_ms":10938,"image_pages":[],"temperature":1.0,"billed_attempts":1,"reasoning_tokens":1079,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-10-28T04:01:15.744751+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-medium","input_kind":null,"input_chars":null,"input_total_chars":null,"passes_missing":null,"published_standing":null,"published_confidence":null,"objection_audit":null,"run_record":null,"basis":{"kind":"not_recorded","source":null,"chars_given":null,"chars_total":null},"receipt":null,"falsifier":"Run the same two-stage training on egocentric human clips in which the camera moves or the scene's appearance changes but no task-relevant manipulation occurs within the one-second horizon; if the world-imagination reward still ranks such Chains-of-Dynamics highly and downstream generalization gains persist, the signal is crediting appearance or camera alignment rather than demonstrated control. A direct check would compare reward-ranked Chains-of-Dynamics against ground-truth wrist motion on a held-out annotated set and ask whether the top-rewarded candidates are also the ones closest to the demonstrated motion.","supporting_citations":[{"cited_title":"FAST: Efficient Action Tokenization for Vision-Language-Action Models","cited_arxiv_id":"2501.09747","evidence_quote":"Supplies the motion-tokenizer architecture used to turn one-second human wrist and robot end-effector deltas into the discrete Chain-of-Dynamics vocabulary."},{"cited_title":"MolmoAct2: Action Reasoning Models for Real-world Deployment","cited_arxiv_id":"2605.02881","evidence_quote":"Supplies the action-expert initialization and the bimanual robot demonstration dataset, and names the baseline against which real-robot gains are measured."},{"cited_title":"DINOv3","cited_arxiv_id":"2508.10104","evidence_quote":"Provides the frozen DINOv3 encoder whose patch features define the future state the model is trained to imagine."},{"cited_title":"DINO-WM: World Models on Pre-trained Visual Features enable Zero-shot Planning","cited_arxiv_id":"2411.04983","evidence_quote":"Establishes the use of DINO feature spaces for future-state comparison, the representational choice carried into the imagination loss."},{"cited_title":"LIBERO-PRO: Towards Robust and Fair Evaluation of Vision-Language-Action Models Beyond Memorization","cited_arxiv_id":"2510.03827","evidence_quote":"Is the distribution-shift benchmark whose changed objects, positions, task semantics, and environments carry the paper's main generalization claims."},{"cited_title":"Libero: Bench- marking knowledge transfer for lifelong robot learning.Advances in Neural Information Processing Systems, 36:44776–44791, 2023","cited_arxiv_id":null,"evidence_quote":"Is the base manipulation benchmark whose standard training set is used to post-train the model before zero-shot evaluation on shift benchmarks."},{"cited_title":"Latent action pretraining from videos","cited_arxiv_id":null,"evidence_quote":"Exemplifies the latent-action approach whose underconstrained variables, able to capture appearance or camera motion rather than behavior, frame the motivation for consequence-grounded supervision."},{"cited_title":"EgoVerse: An Egocentric Human Dataset for Robot Learning from Around the World","cited_arxiv_id":"2604.07607","evidence_quote":"Supplies the large egocentric human-video corpus used in supervised fine-tuning and reinforcement learning, the data source behind the unlabeled-video scaling claims."},{"cited_title":"Egodex: Learning dexterous manipulation from large-scale egocentric video","cited_arxiv_id":null,"evidence_quote":"Supplies the second egocentric human-video corpus used to build the shared Chain-of-Dynamics space across embodiments."}],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}