acceptodds
Under review as a conference paper at ICLR 2027

ThymeV2: Towards Infinite Visual Exploration with Finite Context

Abstract

Complex perception tasks often require multimodal models to explore and refine visual evidence over multiple interactions. Recent multimodal agents have begun to demonstrate such behavior, yet scaling this capability toward sustained exploration remains challenging. Current systems often operate within limited exploration horizons, and extending exploration naturally causes visual contexts to grow rapidly. In this work we introduce ThymeV2, a multimodal model for sustained active perception in images. ThymeV2 combines adaptive multi-turn visual exploration with an evolving visual memory, allowing the model to acquire new evidence while selectively maintaining useful observations. To unlock this capability, we design visual tools for multi-region image acquisition and removing unnecessary observations, construct large-scale multi-turn exploration dataset and develop a two-stage post-training recipe. We first perform supervised fine-tuning (SFT) on turn-level examples using the visual context available at each turn. Following SFT, our reinforcement learning (RL) recipe combines outcome and process signals with fine-grained control over policy updates for stable optimization. Together, these designs support deep exploration, self-correction, and stable learning over extended interactions. We also introduce StratBenchFinal, a challenging visual exploration benchmark with 893 manually reviewed examples across five difficulty levels. ThymeV2 achieves performance comparable to Gemini 3.1 Pro with the same tools and outperforms Gemini 3.1 Pro without tools by up to 12.76 percentage points on average across six challenging perception benchmarks.We will release our models, data, and code to facilitate future research.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
think with image

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_t2zqOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

ThymeV2 is a multimodal framework for sustained visual exploration that combines adaptive image cropping and zooming with an evolving visual memory, selectively removing unnecessary observations while preserving textual reasoning history. Its main contributions include ThymeData, a high-quality training dataset, and a two-stage training pipeline combining context-aligned supervised fine-tuning with reinforcement learning using outcome and process feedback. It also introduces StratBenchFinal, a challenging benchmark consisting of 893 manually verified visual questions. Across six perception benchmarks, ThymeV2 improves over its base models and approaches the performance of Gemini 3.1 Pro equipped with the same tools, while its memory mechanism reduces cumulative visual-token consumption compared with retaining all observations.

Critical strengths
  1. Effective integration of visual exploration and memory management. ThymeV2 selectively removes obsolete visual observations while preserving reasoning history. Memory ablations show both higher accuracy and lower cumulative visual-token consumption than retaining all observations, providing concrete evidence that the design improves the effectiveness and efficiency of visual exploration.

  2. A coherent training framework with strong empirical results. Context-aligned supervision and outcome- and process-based RL jointly address learning under changing visual contexts and recovering from tool errors. The substantial improvements over the base models across six benchmarks, together with training-stability ablations, support the practical value of this framework.

Critical weaknesses
  1. The experiments provide limited support for long-horizon exploration under finite context. Section 6.4 reports average interaction lengths of only 2.09–2.98 turns and an observed maximum of 11 turns, under a 15-turn budget. Appendix E further specifies a 262,144-token context limit, while Section 3.1 states that textual history is preserved when visual observations are removed. Thus, the experiments demonstrate adaptive short-to-moderate exploration, but do not establish sustained performance as interaction length grows under a fixed total-context budget. This gap is consequential because scaling exploration while controlling context growth is a central motivation of the paper.
  2. The memory ablation does not fully isolate the benefit of learned observation removal. Section 6.4 compares the full model against an “inference-time Append-All variant of the same checkpoint,” which disables visual-state removal. Accuracy falls from 71.69% to 64.15% on VisualProbe-Hard and from 62.48% to 59.57% on StratBenchFinal. Although these results support the usefulness of removal for the trained model, Sections 4.1 and 5 explicitly train it on dynamically maintained visual contexts, so the ablation also changes the training–inference alignment. Without an append-all model trained consistently under that setting, or budget-matched sliding-window baselines, the results cannot clearly distinguish the advantage of learned memory selection from distribution mismatch or generic context reduction.
Questions for authors
  1. Exploration scalability. Section 6.4 and Figure 3(a) report average interaction lengths of 2.09–2.98 turns and an observed maximum of 11 turns under a 15-turn budget. Moreover, Section 3.1 states that textual history continues to accumulate. These results appear insufficient to support the title’s claim of “infinite visual exploration” with “finite context.” Can the authors provide experiments with substantially larger interaction budgets, reporting accuracy, context growth, and computational cost, or narrow the claim to match the demonstrated exploration horizon?
  2. Memory management. Section 6.4 compares the full model with an inference-time Append-All variant of the same checkpoint, reducing VisualProbe-Hard accuracy from 71.69% to 64.15%. However, Appendix A.1 states that training inputs follow the dynamically maintained visual memory. Can the authors compare against a model trained to retain all context or a sliding-window baseline with the same visual-context budget? These comparisons would help distinguish the benefit of learned memory selection from adaptation to the training context and the benefits of simply limiting context size.
  3. Effectiveness of process rewards. In Section 6.3 and Table 2, the Pure Outcome baseline removes both process feedback and tool-error masking and exhibits tool-use collapse. This comparison does not isolate whether process rewards are responsible for the improvement. Can the authors report a controlled comparison that removes process rewards while retaining the same masking scheme and training budget, including benchmark accuracy and tool-error rates? This would clarify the contribution of process rewards to task performance and training stability.
  4. Reliability of the LLM judge. Section 5.2 uses an LLM judge to assign trajectory-level process scores and local penalties, which directly influence policy updates. Can the authors report agreement between these judgments and human assessments? Would replacing the judge with another reward model yield comparable training stability and final performance? These results would help establish the reliability and robustness of the reward signals.
2 comments

Dear Reviewer t2zq,

We hope you are doing well. As the discussion period nears its close, we wanted to kindly follow up and thank you again for your encouraging evaluation and deeply insightful expert feedback. Your suggestions regarding claim calibration, controlled memory-eviction baselines, isolated process-reward ablations, and judge reliability audits helped us bring out the strongest rigor and clarity of ThymeV2, all of which we have detailed in our response above.

Thank you once again for your thoughtful guidance and support. If our responses have addressed your questions to your satisfaction, we would sincerely appreciate it if you would consider raising your rating in support of our paper. Please let us know if there is anything else we can clarify!

Best regards, The Authors

Thank you for the detailed responses and additional experiments, which address most of my concerns. I encourage the authors to include these experiments and their methodological details in the appendix.

I think the 15-turn budget in the experiments remains relatively limited and may constrain the model’s potential. I hope future work can scale to substantially longer horizons, moving closer to the goal of “infinite exploration.” My overall assessment remains positive, and I will maintain my original rating.

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

This paper proposes ThymeV2, a multimodal model for sustained visual exploration with evolving memory. Its memory-aligned SFT and RL framework enables stable multi-turn exploration while removing redundant observations. Experiments show strong gains over prior active-perception models with lower visual-token consumption.

Critical strengths
  • This paper propose an interesting mechanism, evolving memory.
  • This paper is well-written and easy to follow.
  • The authors construct high quality data and achieve good performance on benchmarks.
Critical weaknesses
  • Limited methodological novelty in evolving-memory learning. The capability is mainly distilled from Gemini-generated trajectories, while memory-wise RL primarily adapts standard GRPO to reconstructed visual contexts. Beyond generic trajectory-level process feedback, the framework introduces no memory-specific objective or explicit exploration-state representation. This leaves key challenges specific to sustained visual exploration insufficiently studied, including how to preserve potentially useful evidence under uncertainty and how to avoid repeatedly exploring previously discarded regions.

  • Restricted task scope. ThymeV2 reduces the broader capabilities of Thyme v1 to cropping, zooming, and observation removal, and is evaluated only on static-image perception tasks. Its applicability to broader agent settings with heterogeneous or irreversible observations remains unclear.

  • Potentially confounded experimental comparison. ThymeV2 uses domain-specific data generated by frontier Gemini models, yet its headline gains compare against Gemini without matched tools; with the same tools, Gemini 3.1 Pro scores 78.14 versus 75.96/77.02 for ThymeV2. Without a standard distillation baseline trained on identical data, the gains cannot be clearly attributed to the proposed method rather than teacher data and task specialization.

Questions for authors

Please kindly refer to the weakness section.

5 comments

Dear Reviewer Kkiz,

We hope this message finds you well. As the author–reviewer discussion period approaches its end, we wanted to gently follow up and thank you once again for your thoughtful review and for recognizing our evolving visual memory mechanism and dataset quality. Your insightful comments regarding design philosophy, broader cross-task and long-video transfer, and identical-data attribution helped us substantially strengthen the paper, and we have provided detailed analyses and new experimental results in our response above to address each point.

We truly appreciate your valuable time and guidance. If our responses and new results have satisfactorily resolved your concerns, we would be deeply grateful if you would consider raising your score. Please also let us know if any further clarification would be helpful.

Warm regards, The Authors

Dear Reviewer Kkiz,

We wanted to send a brief, polite follow-up to thank you again for your thoughtful review. We truly hope our controlled memory-eviction ablations, cross-task and long-video transfer results, and identical-data SFT attribution baselines have satisfactorily answered your concerns. If our responses have addressed your questions, we would sincerely appreciate it if you would consider updating your rating before the deadline.

Thank you once again for your valuable time and support!

Warm regards, The Authors

Dear Reviewer Kkiz,

With the author–reviewer discussion period wrapping up today (November 18 at 11:59pm AoE), we wanted to send one brief final follow-up to thank you again for your thoughtful review. We remain readily available before the deadline if any further clarification would be helpful, and we would deeply appreciate it if you would consider updating your rating should you find our responses and new experiments satisfactory.

Thank you once again for your valuable time and support!

Warm regards, The Authors

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

The paper presents ThymeV2, a Qwen3.5-based multimodal model (9B and 27B) for multi-turn active perception on high-resolution images. Tool and memory. The model uses a single tool that can request several zoomed crops per call. The same call can drop earlier crops from an “evolving visual memory.” Text history is kept, while dropped images leave the context. Training has two stages SFT and RL.

Critical strengths
  1. ThymeV2-9B improves its backbone by 14.2 points on average and by 34.9 on VisualProbe-Hard. It beats every reproduced active-perception baseline on all six benchmarks. It comes within about 2 points of Gemini 3.1 Pro given the same tools.

  2. The paper keeps well-documented multi-turn RL recipe with a concrete failure diagnosis.

  3. Turn-level replay of the active visual memory, in both SFT and RL, is a sensible way to keep training and inference contexts aligned.

Critical weaknesses

W1. The evaluation does not isolate the source of the gains, and train/test contamination is not ruled out.

(a) Harness confound.

  1. Gemini 3.1 Pro gains 13.88 points from the harness alone (64.26 → 78.14). That is almost exactly ThymeV2-9B’s gain over its backbone (14.23).

  2. The most informative baseline is missing: Qwen3.5-9B and -27B run zero-shot in the ThymeV2 harness.

  3. Baselines also run under much tighter budgets. ThymeV2 gets 262k context, 20k output tokens per call and 15 generations. Most baselines get 32k context, and Chain-of-Focus gets 512 output tokens (Table 5). Table 1 mixes training effects with harness and budget effects.

(b) Most of the gain comes from distilling the comparison model.

  1. The SFT trajectories are generated by Gemini-3-Pro inside the same harness (App. A.1), and Gemini 3.1 is also the comparison model.

  2. Rollout Mask (72.81) is only 0.8 above SFT.

  3. “Comparable to Gemini with the same tools” is therefore largely a student approaching its teacher. The framing should reflect that, and the RL contribution should be stated plainly.

(c) Possible contamination.

SFT and RL questions are drawn from, among others, Lai et al. 2026 (Mini-o3) and Wei et al. 2026 (Zooming without Zooming). Those same works supply two evaluation benchmarks: VisualProbe-Hard and ZoomBench. These are also the benchmarks with the largest gains (+34.9 and +16.7). No decontamination procedure is described. StratBenchFinal’s nine source benchmarks are never named, so overlap with the training sources, or with V*, HR-Bench and MME-RealWorld, cannot be checked.

(d) StratBenchFinal is selected on the backbone’s failures.

Difficulty is defined by Qwen3.5-9B no-think pass@8. In particular, 393 of the 893 items are ones the backbone never solved. Any model derived from that backbone will look like it gains on such a set, partly through selection effects and label noise.

(e) No statistical treatment, and one unexplained inconsistency.

ThymeV2 is evaluated with a single sample at temperature 1.0, with no seeds or confidence intervals. Claims such as “trails Gemini by 0.95 points” (one question), “matches on V-Hard”, and the +6.13 V-Hard margin at update 160 are within noise. App. H.1 reports ThymeV2 RL-step300 at 63.81% (67/105) on the same VisualProbe-Hard examples minus one. Table 1 reports 71.69% (76/106) for the same checkpoint. This discrepancy implies either large run-to-run variance or a different configuration. Either way it needs explaining.

W2. The central claims about long-horizon (“infinite”) exploration and evolving memory are not supported by targeted experiments, and several RL components are unablated.

(a) Episodes are short, and the title overclaims.

  1. Mean interaction length is 2.09–2.98 turns, including the final answer (Fig. 3a). The maximum observed is 11 turns, under a 15-turn budget and a 12-active-image cap.

  2. No experiment scales the turn budget for ThymeV2. The only scaling evidence is Mini-o3’s own table, reproduced in App. G. At these lengths, append-all contexts would not approach the 262k limit. Nothing shows that memory pruning is what makes longer exploration possible. “Towards Infinite Visual Exploration with Finite Context” is not substantiated.

(b) The memory ablation is confounded.

  1. “Append-All” is an inference-time change to a checkpoint trained with removal. The 7.54-point V-Hard drop may reflect train/test mismatch rather than a benefit of memory.

  2. The right comparison is a model trained with append-all versus one trained with evolving memory, run across increasing turn budgets. The token savings are modest. Per example, they work out to about 2.8k visual tokens on V-Hard (300,592/106) and about 1.0k on Strat (917,586/893). Read off Fig. 3b, that is roughly 5% and 3–4% of cumulative visual tokens.

  3. Dropping images from the middle of the context also breaks prefix/KV caching. Real prefill cost and latency may therefore not improve, and none are reported.

(c) Several RL components lack ablations.

  1. Process feedback. It is ablated only through the standard deviation of logged reward (0.0387 vs 0.0439). No final accuracy is reported for “w/o process feedback.”

  2. “Pure Outcome.” This variant removes process feedback and masking at the same time, so the two effects are conflated. Bounded step penalties. These are never ablated. As formulated, the per-token correction is Δ/N with |Δ| ≤ 0.125, while the rollout advantage of roughly unit scale is applied to every token. The penalty’s relative strength therefore shrinks roughly as 1/N and may be negligible for typical turn lengths.

  3. The 1/K_i weighting (Eq. 14). It down-weights trajectories that perform more memory reconstructions. This seems to work against the stated goal, and it is unjustified and unablated.

  4. Span Mask Only is reported at update 140 while Full is reported at update 300. A matched-update comparison is needed.

  5. Turn-level assessment. The variant is compared at a single checkpoint (update 160), while the mean over shared checkpoints differs by only 0.47 points.

Questions for authors
  1. What do Qwen3.5-9B and Qwen3.5-27B score zero-shot inside the ThymeV2 harness (same tool, 15 turns, same context)? Ideally, also give the strongest baselines matched context and turn budgets.

  2. Please describe the decontamination between ThymeData’s question sources and all evaluation sets, especially VisualProbe-Hard and ZoomBench, whose source works also supply training questions.

  3. Which nine benchmarks does StratBenchFinal draw from, and in what proportions? Do any items overlap with the other five evaluation sets or with ThymeData?

  4. Please report mean ± std over several seeds or samples for Table 1.

  5. Why does App. H.1 report 63.81% for RL-step300 on VisualProbe-Hard while Table 1 reports 71.69%?

4 comments

Dear Reviewer qqgk,

We hope this message finds you well. As the author–reviewer discussion period approaches its end, we wanted to gently follow up and express our deepest gratitude once again for your exceptionally thorough and rigorous review. Your thoughtful insights regarding evaluation confounds, decontamination, statistical stability, and targeted RL and memory ablations genuinely elevated the rigor of our work, and we have incorporated comprehensive controlled experiments and clarifications in our response above to address every point you raised.

We truly appreciate your valuable time and expertise. If our responses and new experimental results have satisfactorily resolved your concerns, we would be deeply grateful if you would consider raising your score. Please also let us know if any further clarification would be helpful.

Warm regards, The Authors

Dear Reviewer qqgk,

We wanted to send a brief, polite follow-up to thank you once again for your extraordinarily thorough and constructive review. We sincerely hope our new controlled baselines, multi-seed statistics, decontamination audit, and matched-update RL ablations have addressed all of your concerns. If you find that our responses have resolved the issues raised, we would be deeply grateful if you would consider raising your score before the discussion closes.

Thank you very much once again for your time and effort!

Warm regards, The Authors

Dear Reviewer qqgk,

As the discussion period closes today (November 18 at 11:59pm AoE), we wanted to send a brief final note of thanks for your exceptionally thorough and helpful review. We are standing by until the deadline in case any final questions arise regarding our new controlled baselines, decontamination audit, or statistical results, and we would be sincerely grateful if you would consider raising your score if your concerns have been resolved.

Thank you very much once again for your time and effort!

Warm regards, The Authors

Dear Reviewer qqgk,

With roughly 24 hours left before the discussion window closes (November 18 at 11:59pm AoE), we wanted to send one brief final thank-you for your extraordinarily thorough review. If our controlled baselines, decontamination audit, and statistical analyses have resolved your concerns, we would be sincerely grateful if you would consider updating your score before the portal closes.

Thank you very much once again for your time!

Warm regards, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.