acceptodds
Under review as a conference paper at ICLR 2027

-0.6: A Unified Vision-Language World Model with Adaptive Streams-of-Thought

Abstract

Intelligence in the physical world requires self-directed, task-adaptive reasoning through different representations of the world as a problem unfolds. We introduce Ψ-0.6, a unified vision-language world model that adaptively selects the representations through which it reasons. Ψ-0.6 reasons through streams-of-thought: sequences interleaving images and text with geometrically-grounded, interpretable visual abstractions such as camera pose, depth, optical flow, and point tracks. A shared token interface supports any-to-any prompting, allowing the same model to generate, interpret, and compose these representations as intermediate reasoning steps. Ψ-0.6 matches state-of-the-art zero-shot performance on MindCube, a challenging benchmark of multi-view spatial reasoning. Composing these capabilities — estimating camera motion, selecting the relevant observed view, and answering from it — further improves performance. World modeling enables counterfactual imagination: the model translates a specified object motion into a flow prediction, generates the resulting scene, and answers from it. On FlowBench-Real, which tests this ability, Ψ-0.6 shows better generalization than Qwen. Through reinforcement learning from task feedback, Ψ-0.6 learns adaptive streams-of-thought that select visual representations for each input and task. This improves visual question-answering accuracy, with gains transferring to unseen benchmarks. By unifying perception and imagination, geometry and language, as operations of the same model, Ψ-0.6 uses them for representation-adaptive reasoning.

Primary area:
foundation or frontier models, including LLMs
Keywords:
computer vision, vision-language model

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_XgZrOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents Ψ-0.6, which is a unified vision-language world model that reasons through streams-of-thought. Streams-of-thought are sequences that interleave images and text with grounded and interpretable abstractions, including camera pose, depth, optical flow, and point tracks. Ψ-0.6 enables compositional reasoning over these capabilities. It also uses reinforcement learning (RL) to learn adaptive streams-of-thought that select visual representations for each task. Experiments on challenging spatial reasoning benchmarks like MindCube show strong zero-shot performance. It also shows better generalization than Qwen model on FlowBench-Real. Finally, Ψ-0.6 also demonstrates transfer to unseen benchmarks.

Critical strengths
  1. Unifying vision-language understanding and world modeling with different representations as intermediate reasoning steps is novel. The demonstration of compositional reasoning shows strong generalizability. This is also aligned with human visual reasoning, which selects different representations for different tasks at hand. The part on using reinforcement learning to select representations is particularly interesting. Training an adaptive reasoning model using RL that operates at the level of representations is crucial to efficient and flexible visual reasoning.

  2. The paper's figures are well-made and helpful for understanding the paper's core methodology and results. Readers can grasp the paper's idea, method, and results before reading it in detail.

Critical weaknesses
  1. The reinforcement learning part for training the policy to select the representation, despite being interesting, has limited performance gain compared to random selection. In Figure 5, most of the policy results are on par with random selection. Best-of-16 performs much better, but you should also compare it with random best-of-16. Also, this will have a much higher average inference cost.
Questions for authors
  1. In Figure 3, GeoThinker-8B has the highest performance, better than all of your model variants. Could you explain why it performs so well? Since both models are 8B, it is a fair comparison at the model-size level.
2 comments
AuthorsOpenReview ↗

Response to Reviewer XgZr

We thank the reviewer for the careful assessment and are glad that they found the unified design "novel", the RL-based selection of representations "particularly interesting", and the figures "well-made and helpful". Their main concern was that learned selection gains little over random selection in Fig. 5, and that policy best-of-16 should be compared with random best-of-16. As requested, we added random best-of-16 to Fig. 5 and best-of-k for every k from 1 to 16 to the appendix (W1). We also explain GeoThinker's result (Q1).

Weaknesses

W1. Learned vs random selection. The reviewer asks for random best-of-16 next to the policy's best-of-16, and notes that best-of-16 has a much higher inference cost.

Learned selection stays ahead of random selection in pooled accuracy at every k from 1 to 16, and its lead is clearest at a single rollout, which is how the policy is used.

  • Single rollout (averaged over 16 rollouts): above random on all eight benchmarks. Pooled: +1.13 points on held-out (p < 0.001) and +0.34 on unseen benchmarks (p = 0.004).
  • Best-of-16 (revised Fig. 5): above random on six of the eight benchmarks.
  • Across k, the policy leads on all eight benchmarks at k = 1–2, seven at k = 3, six at k = 4–5, five at k = 6–7, four at k = 8–10, five at k = 11, and six from k = 12 to 16 (Appendix C, p. 25).

Best-of-16 is an analysis of headroom, not the inference procedure we propose.

Questions

Q1. GeoThinker. The reviewer asks why GeoThinker-8B (83.0) outperforms Ψ-0.6 (76.3) in Fig. 3, given that both models are 8B.

GeoThinker-8B is a specialized model for MindCube-like tasks:

  • GeoThinker-8B is Qwen3-VL-8B with an external 3D model (VGGT) fused into its layers, fine-tuned on 1.8M spatial samples that include MindCube's training split: about 180 times the 10k MindCube samples of our fine-tune.
  • Given to Qwen3-VL-8B, Ψ-0.6's own predicted camera poses are nearly as useful as VGGT's (48.3 vs 50.0, Appendix F, p. 31).
  • GeoThinker's authors report that removing the external encoder and predicting geometry inside the model, with identical training data, raises MindCube from 83.0 to 91.0 (GeoThinker-F, arXiv 2610.20592), which in fact supports our argument for unified design.

As discussed in the general response, GeoThinker also builds on a production VLM, while Ψ-0.6 is trained from scratch with far more limited resources.

Thanks for the rebuttal. I maintain my original positive evaluation of the paper.

Reviewer_ZrFiOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Ψ-0.6, a unified model that generates visual and geometric predictions as intermediate steps for answering questions. It also trains a policy to select which intermediate representations to use. Experiments cover spatial reasoning, counterfactual motion prediction, and video question answering.

Critical strengths

Using the same model to generate intermediate predictions and answer from them is an interesting direction.

The paper compares learned representation selection with random selection. It also tests the policy on unseen benchmarks.

Critical weaknesses

The benefit over direct SFT is unclear. The paper includes task-specific fine-tuning, but does not clearly compare the learned selection policy with direct question-answering SFT on the same model under a matched training budget. This makes the value of the additional policy training difficult to assess.

The empirical gains are uneven. On EgoDex, learned selection improves accuracy from 50.2% with random selection to 53.5%. On FlowBench-Real, the fine-tuned model exceeds fine-tuned Qwen by only 0.9 percentage points, while trailing it on FlowBench-Sim. These results provide some support for the approach, but do not yet show a consistent advantage that justifies its added complexity.

Questions for authors

How does the learned selection policy compare with direct question-answering SFT on the same model, using the same data and a similar training budget?

Under a matched inference budget, how does it compare with fixed intermediate steps or repeated RGB-only reasoning?

3 comments
AuthorsOpenReview ↗

Response to Reviewer ZrFi

We thank the reviewer for the careful review, and for commending the paper's comparison between learned selection with random selection, and the paper's testing of the policy on unseen benchmarks. The main concerns are whether learned selection adds value beyond direct question answering, whether its gains are consistent, and how it compares with other choices at the same inference budget. We address each below, with new best-of-k results (Appendix C, p. 25).

Weaknesses

W1 / Q1. Direct question-answering SFT. The reviewer asks how the learned selection policy compares with direct question-answering SFT on the same model, using the same data and a similar training budget.

Fig. 5's RGB-only bar measures this control:

  • We now make it clear that, prior to learning stream-of-thought, Ψ-0.6's instruction training already includes direct question answering on SSv2, PE and OpenVid, three of the four sources the policy is trained on (Appendix D, p. 27).
  • The policy is above RGB-only on all eight benchmarks in Fig. 5, by 1.2 to 6.0 points.

W2. Uneven gains. The reviewer finds the gains uneven: learned selection improves EgoDex from 50.2% (random selection) to 53.5%, and on FlowBench fine-tuned Ψ-0.6 exceeds fine-tuned Qwen by 0.9 points on Real.

Learned selection's gains are consistent in direction and statistically reliable.

  • EgoDex, 53.5 vs 50.2: significant (paired McNemar exact test, p = 0.014).
  • Averaged over 16 rollouts, the policy is above random selection on all eight benchmarks. Pooled: +1.13 points [+0.83, +1.43] on held-out (p < 0.001) and +0.34 [+0.11, +0.58] on unseen benchmarks (p = 0.004) (Appendix C, p. 25).

On FlowBench, Ψ-0.6 transfers better than fine-tuned Qwen3-VL-8B from simulation to real robots.

  • On FlowBench-Real, real robot videos that neither model saw in training, Ψ-0.6 matches it (57.5 vs 56.6 overall, 56.3 vs 53.6 category average, Table 2). Ψ-0.6 loses 11.3 points from Sim to Real; Qwen loses 17.0.

Questions

Q2. Matched inference budget. The reviewer asks how learned selection compares, at a matched inference budget, with fixed intermediate steps or repeated RGB-only reasoning.

At a matched inference budget, learned selection beats both a question-agnostic choice of streams and RGB-only.

  • Fig. 5's random bar is the matched-budget control for the choice of streams: it draws one notation per question. Averaged over 16 rollouts, the policy beats it on all eight benchmarks (see W2).
  • Repeating RGB-only reasoning reproduces the same answer, because answers are scored by the likelihood of each option. The policy is above RGB-only on all eight benchmarks.
  • With k samples from each, the policy stays ahead of random selection in pooled accuracy for every k from 1 to 16 (Appendix C, p. 25).

Thank you for the new tests across sampling budgets. They show that learned selection performs better than random selection. My main concern remains the comparison with direct question-answering SFT. The RGB-only result is helpful, but it does not match the training data and budget. I would also like to see a direct comparison with a fixed choice of intermediate representation. Given the small gain on unseen benchmarks, I prefer to keep my score of 2.

AuthorsOpenReview ↗

Clarification of Misunderstanding

Thank you for the follow-up. We want to clarify that our experiments already address both comparisons, and apologize that our earlier response was not sufficiently clear.

Direct QA SFT: The RGB-only baseline in Fig. 5 is our direct QA-SFT control, with a matched 1.2M-step training budget.

  • Both are evaluated using the same frozen answerer, so the only difference is the selected intermediate representations.
  • The answerer was trained on direct QA from SSv2, PE-Video, and OpenVid (three of the four policy-training sources). The policy is initialized at 1.16M steps and trained for another 40k steps, matching the 1.2M-step budget (Appendix D).
  • Therefore, the direct QA SFT control you asked for is exactly the RGB-only baseline we report.
  • Result: Learned selection outperforms RGB-only by 1.2–6.0 points across all eight benchmarks.

Fixed representations: Our random-selection baseline uniformly samples from the same set of intermediate representations, so its expected accuracy equals the average performance of the fixed strategies. Learned selection outperforms random selection across all eight benchmarks.

We hope this resolves the two concerns.

Reviewer_yFxoOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents PSI-0.6, an any-to-any multimodal model that accepts and generates images, depth maps, point clouds, and text. The model is trained in three stages, each with its own datasets (pretraining, mid-training and post-training). One central contribution is “streams of thought,” a reasoning mechanism in which the model can generate non-textual representations (such as depth or flow maps) as intermediate steps within its chain of thought. The authors further optimize these reasoning traces with reinforcement learning. Experiments show that PSI-0.6 outperforms baselines on MindCube when combined with structured prompting.

Critical strengths

1.Unifying multiple visual and geometric modalities in a single any-to-any model is an interesting and compelling direction, particularly for spatial reasoning. 2.Allowing the model to reason in output modalities beyond text, rather than only consuming them as inputs, is a notable idea and a meaningful step beyond text-only chain of thought. 3.Using RL to optimize multimodal reasoning traces appears novel and could be broadly useful.

Critical weaknesses
  1. Insufficient detail on training data. The paper does not adequately describe the data used in each training stage, such as the types of questions, the sources and characteristics of the videos, and the dataset sizes. This makes it difficult to assess the source of the gains and limits reproducibility.
  2. Insufficient detail on tokenization. It is unclear how images, depth maps, point clouds, and other modalities are tokenized. Details such as tokenizer architecture, vocabulary or codebook size, resolution, and whether tokenizers are shared or modality-specific are important for understanding and reproducing the approach.
  3. Missing baselines on MindCube. The evaluation lacks zero-shot results for off-the-shelf MLLMs given the same auxiliary signals (e.g., depth maps or camera parameters as additional input). Without these, it is unclear whether the gains come from the model’s ability to generate these representations or simply from having access to them.
  4. Missing comparison with tool-use approaches. A natural alternative to generating depth or 3D representations internally is to have an MLLM call external expert models (e.g., a depth estimator or 3D reconstruction model) as tools. A comparison against such a pipeline is needed to justify the unified any-to-any design.
  5. Limited benchmark coverage. The model is not evaluated on several established spatial reasoning benchmarks, such as SAT (SAT: Dynamic Spatial Aptitude Training for Multimodal Language Models), IPT( Imaginative Perception Tokens Enhance Spatial Reasoning in Multimodal Language Models), All-Angles (https://danielchyeh.github.io/All-Angles-Bench/) and VSI-BENCH. Results on these would substantially strengthen the claims of improved spatial reasoning. 6.Efficiency on FlowBench. Qwen fintuned (56) and PSI-0.6 (57) perform similarly on FlowBench. Given this, reporting token usage and inference latency would help clarify whether the multimodal reasoning offers practical advantages or incurs additional cost. 7.figures are overcrowded and difficult to read. Simplifying them or splitting them into multiple panels or moving numerical results to tables would improve clarity.
Questions for authors

Please address the weaknesses above.

How does PSI-0.6 compare with latent visual reasoning approaches such as Mirage (Machine Mental Imagery: Empower Multimodal Reasoning with Latent Visual Tokens), and with methods that use intermediate perception tokens, such as Imaginative Perception Tokens Enhance Spatial Reasoning in Multimodal Language Models?

can you make contributions compared to PSI more clear?

2 comments
AuthorsOpenReview ↗

Response to Reviewer yFxo (1/2)

We thank the reviewer for a detailed and constructive review, and for finding the any-to-any design "an interesting and compelling direction" and reasoning in output modalities "a meaningful step beyond text-only chain of thought". The main concerns are missing details (training data, tokenization), missing baselines on MindCube (the same auxiliary signals, tool use) and benchmark coverage. We address each below with new experiments and new appendix sections.

Weaknesses

W1. Training data. The reviewer asks what data each training stage uses (question types, video sources, dataset sizes), to judge where the gains come from and to allow reproduction.

Ψ-0.6 learns its geometry from expert pseudo-labels on ≈12.3M videos, and its language from standard public text, caption and question-answering data. Appendix D (p. 27) lists every source, its size and its supervision for each training stage. In short:

  • Video: ≈12.3M clips from 24 sources covering natural scenes, human activity, egocentric and robot video. The largest are a hand-selected web-video collection (4.3M), SpatialVid (2.7M) and YouTube (1.6M). Others include PE, Kinetics-700, Egocentric-10k, ScanNet++, DROID and EgoDex. Depth and camera pose are Depth Anything 3 pseudo-labels; optical flow comes from AllTracker.
  • Text: FineWeb-Edu and DCLM (≈350B training token positions).
  • Captions: global and time-indexed captions of images and videos, either provided by the dataset or generated by Qwen3-VL-8B.
  • Video QA (1.29M items): descriptive questions rewritten from OpenVid and PE captions, and action questions built from SSv2 labels.
  • Image QA (2.9M items): VQAv2, LLaVA-Instruct-150K, PixMo and Cambrian.
  • Multi-view spatial QA (659K items): SenseNova-SI.
  • Camera control: text-to-camera trajectories from CO3D, MVImgNet and SpatialVid.

W2. Tokenization. The reviewer asks how each modality is tokenized: tokenizer architecture, codebook size, resolution, and whether tokenizers are shared or modality-specific.

Tokenizers are modality-specific, but all modalities share one 65,536-token vocabulary, so the transformer reads and writes every modality the same way. Appendix E (p. 30) gives the details. In short:

  • RGB, depth and optical flow: each has a separate, frozen local patch quantizer that maps every 16×16-pixel patch to a few discrete codes (4 for RGB, 2 for depth and for flow).
  • Camera pose: the relative pose between two frames, discretized into 8 integers (field of view, rotation, translation).
  • Text: GPT-2 BPE.
  • Pointer tokens: each patch is preceded by one, saying where and when the patch belongs. This lets a prompt condition on, or generate, any subset of patches.
  • A channel embedding tells the model which modality each token belongs to.

The appendix gives the quantizer architecture and training, codebook sizes, input resolution, tokens per frame, and the transformer architecture.

W3. Same auxiliary signals. The reviewer asks for off-the-shelf MLLMs given the same auxiliary signals (depth maps or camera parameters) zero-shot, to tell whether the gains come from generating these representations or simply from having them.

An off-the-shelf MLLM needs camera poses supplied to reach the accuracy that Ψ-0.6 reaches from RGB alone. We ran this new baseline on MindCube-Tiny with Qwen3-VL-8B, using the same prompt and likelihood scoring as for Qwen3-VL-8B in Fig. 3 (Appendix F, p. 31):

  • Qwen3-VL-8B, RGB: 39.3.
    • camera poses from dedicated 3D models (VGGT, Depth Anything 3): 50.0.
    • camera poses predicted by Ψ-0.6: 48.3.
    • depth maps: 37.0 (no gain).
  • For reference, Ψ-0.6 from RGB alone: 49.7 (Fig. 3).

Ψ-0.6 can therefore supply the signal itself, at near-expert quality.

W4. Tool use. The reviewer asks for a comparison with an MLLM that calls external expert models (a depth estimator, a 3D reconstruction model) as tools, to justify the unified design.

Expert 3D tools help Qwen3-VL-8B, but it still trails Ψ-0.6 answering from RGB alone. We built this new pipeline on MindCube-Tiny: Qwen3-VL-8B gets two VGGT-backed tools, one returning the relative camera pose between two views and one returning a depth map, and decides itself when to call them (Appendix F, p. 31):

  • Qwen3-VL-8B, RGB: 30.3 (Tool use requires Qwen3-VL-8B to write out its answer rather than pick the most likely option, which lowers its RGB-only score from 39.3 (Fig. 3) to 30.3.).
    • VGGT tools, called agentically: 36.8.
    • VGGT camera poses and depth, always provided: 43.0.
  • Ψ-0.6, RGB, zero-shot: 49.7.
AuthorsOpenReview ↗

Response to Reviewer yFxo (2/2)

W5. More spatial benchmarks. The reviewer asks for results on more spatial reasoning benchmarks.

Zero-shot, Ψ-0.6 is on par with Qwen3-VL-8B on All-Angles' EgoHumans scenes and ahead on its camera-pose and manipulation questions, but trails it on SAT and IPT (Appendix G, p. 33):

  • All-Angles: on par on the EgoHumans part (44.1 vs 46.5), and ahead on the camera-pose (64.8 vs 47.2, p = 0.009) and manipulation (54.9 vs 39.4, p = 0.001) questions of the Ego-Exo4D part.
  • SAT-Real: behind (56.7 vs 70.0), at about GPT-4o's level (57.5).
  • IPT: behind (e.g. perspective taking 50.0 vs 56.8).

As discussed in the general response, gaps to Qwen3-VL-8B mix our contribution with differences in training scale and language pretraining.

W6. Cost on FlowBench. The reviewer notes that fine-tuned Qwen (56.6) and Ψ-0.6 (57.5) perform similarly on FlowBench, and asks for token usage and inference latency, to see whether multimodal reasoning offers practical advantages or adds cost.

Ψ-0.6 matches fine-tuned Qwen3-VL-8B (trained on 30+ Trillion tokens) on FlowBench-Real with at least 12 times fewer training tokens (Appendix D, p. 27), at a higher but still sub-second cost per question.

Inference, per question on an A40 (batch 1, bf16; Appendix H, p. 33): Qwen3-VL-8B (fine-tuned) reads 479 input tokens and takes 0.17 s. Ψ-0.6 (fine-tuned) reads ≈2,870 input tokens, almost all of them for the image, and takes ≈0.61 s (estimated from its measured per-token latency). The reported rows answer directly, so the extra inference cost comes from Ψ-0.6's denser image tokens, not from streams-of-thought.

W7. Figures. The reviewer finds the figures overcrowded.

We have remade several figures in the paper, by significantly reducing the amount of text in figures and adding white space. They now look very clear.

Questions

Q1. Mirage and IPT. The reviewer asks how Ψ-0.6 compares with latent visual reasoning (Mirage) and with intermediate perception tokens (IPT).

Mirage and IPT also reason through visual intermediates, but theirs are defined and supervised per task, whereas Ψ-0.6's are general physical quantities learned once in pretraining.

  • Mirage's intermediates are latent tokens, distilled from helper images built for each task, and cannot be inspected directly.
  • IPT trains a unified model to generate a task-specific imagined image (a novel view, a side view along a path, or a top-down map), mainly as training supervision. Its headline results generate no image at inference.
  • Ψ-0.6's intermediates (camera pose, depth, optical flow, imagined frames) are shared across tasks and generated at inference. They can be inspected, and the choice among them can itself be learned (§5).

The approaches are complementary: task-specific supervision of the kind Mirage and IPT use could be applied on top of Ψ-0.6's general intermediates. We added both to the related work (Appendix B.2, p. 22).

Q2. Contribution over PSI. The reviewer asks for a comparison of contribution relative to PSI.

PSI showed that structures such as flow and depth can be extracted from a visual world model; Ψ-0.6 shows what this makes possible for reasoning:

  • PSI integrates such structures back as new token types, but it is a vision-only model that cannot answer questions.
  • Ψ-0.6 is a new 8B model trained from scratch that adds language as a native modality. The integrated modalities (camera pose, depth, optical flow) can then serve as intermediate steps toward an answer: streams-of-thought.
  • We show that these streams compose with vision-language and world-model operations in a single reasoning trace (§3, §4).
  • We show that which stream to use can be learned to improve video reasoning (§5).

We made this explicit in Appendix B.1 (p. 22).

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers

We thank the reviewers for their careful reviews. We are glad that they found unifying vision-language understanding and world modeling, with different representations as intermediate reasoning steps, "novel" (XgZr) and "an interesting and compelling direction, particularly for spatial reasoning" (yFxo), and found reasoning in output modalities beyond text "a meaningful step beyond text-only chain of thought" (yFxo). Reviewers found using RL to select representations "particularly interesting" and "crucial to efficient and flexible visual reasoning" (XgZr), and said it "appears novel and could be broadly useful" (yFxo). They also noted that the paper compares learned with random selection and tests the policy on unseen benchmarks (ZrFi), that the compositional reasoning "shows strong generalizability", and that the figures are "well-made and helpful" (XgZr).

The main concerns are:

  • whether learning to select streams adds value beyond random selection and direct answering (ZrFi, XgZr);
  • baselines that give an off-the-shelf MLLM the same auxiliary signals or expert tools, and coverage of more spatial benchmarks (yFxo);
  • details of the training data and tokenization (yFxo).

We respond to every concern in the individual replies, with the following new experiments:

  • Best-of-k for every k from 1 to 16, for the policy and for random selection (XgZr W1, ZrFi Q2).
  • Qwen3-VL-8B given the same auxiliary signals as Ψ-0.6, and a Qwen3-VL-8B agent that calls expert 3D models as tools (yFxo W3, W4).
  • Three further spatial benchmarks, SAT, IPT and All-Angles, run zero-shot against Qwen3-VL-8B (yFxo W5).
  • FlowBench token usage and latency (yFxo W6).

The revised PDF adds:

  • new figures, including one comparing random vs policy's best-of-16 (XgZr W1);
  • appendices on the training data and the tokenizers (yFxo W1, W2), and on each new experiment above;
  • Mirage and IPT in the related work (yFxo Q1);
  • a paragraph on the contribution over PSI (yFxo Q2).

Contribution and comparisons with production VLMs. Ψ-0.6 presents a conceptual and technical advance, adaptive streams-of-thought, and the results, including the new experiments, support it. Ψ-0.6 is a unified model whose own geometric and world-model predictions (camera pose, depth, optical flow, imagined frames) serve as intermediate reasoning steps, and whose choice among them can be learned.

  • Streams-of-thought improve the same model: on MindCube, inverse dynamics raises Ψ-0.6 from 51.3 to 57.7 (§3).
  • Learned selection improves video question answering: the policy is above RGB-only on all eight benchmarks in Fig. 5, and, averaged over 16 rollouts, above random selection on all eight (pooled +1.13 points on held-out and +0.34 on unseen benchmarks, both p ≤ 0.005).
  • Ψ-0.6's own geometry is near expert quality: given to Qwen3-VL-8B, its predicted camera poses raise MindCube accuracy from 39.3 to 48.3, close to the 50.0 reached with VGGT's.

Many of the new comparisons are against Qwen3-VL-8B, a production VLM that builds on a mature language model and large-scale multimodal instruction tuning, developed with industry-scale compute and teams. Ψ-0.6 is trained from scratch by an academic team with far less data, compute, and human resource. Where Ψ-0.6 trails such models, the gap mixes our contribution with these differences in scale; the within-model comparisons above isolate the contribution.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.