acceptodds
Under review as a conference paper at ICLR 2027

Reading is not Reasoning: Bridging the Agentic Policy Gap in Vision–Text Compression

Abstract

Multi-step language-model agents repeatedly process growing interaction histories, leading to substantial context costs. Vision-text compression reduces these costs by rendering history as images, but the resulting modality shift creates a marked capability gap. Through controlled evaluations of history recovery, matched-state decisions, and complete trajectories, we show that this gap cannot be explained by OCR quality alone. Visual-history agents exhibit systematic drift in action selection, query formulation, stopping, and evidence use, revealing an agentic policy gap. We introduce CAPS, a two-stage Cross-modal Agentic Policy Self-distillation framework that uses the same model's stronger text-history policy to supervise its visual-history counterpart. Offline trajectory self-distillation transfers successful text-policy behavior to visual-history inputs, while online policy self-distillation provides dense supervision on states visited by the visual-history policy during reinforcement learning. On SearchQA, CAPS improves over AgentOCR by 5.0 and 3.4 percentage points with 3B and 7B backbones, respectively. On full-history ALFWorld, the corresponding gains are 15.6 and 14.5 percentage points. Across settings, CAPS reduces average memory-context cost by up to 63.3% and peak cost by up to 83.4% relative to matched text-history policies. These results show that explicit cross-modal policy self-distillation can preserve agent capability under vision-text compression. Our code will be made publicly available.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
agent, rl, compression

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_iXm5OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper investigates the performance gap between text-history and image-history agents when interaction histories are compressed into images. The authors find that this gap cannot be explained by OCR accuracy alone, but also involves differences in reasoning and action selection. To address this, they propose CAPS, a two-stage cross-modal self-distillation framework that transfers the stronger text-history policy to a visual-history agent through offline trajectory distillation and online policy distillation during RL. Experiments on SearchQA and ALFWorld show improvements over prior visual-history baselines while reducing context token costs and achieving performance comparable to text-history agents.

Critical strengths
  • Interesting and practical problem. Using OCR-style image compression to reduce the context cost of agent interaction histories is an interesting direction, with clear potential for improving the efficiency of long-horizon agents.

  • Insightful motivation and analysis. The paper provides a useful distinction between history information recovery and downstream reasoning/decision-making. The diagnostic experiments offer interesting evidence that the performance gap cannot be explained by OCR accuracy alone.

Critical weaknesses
  • Limited evaluation scope. The experiments are restricted to two benchmarks and relatively small, outdated Qwen2.5-VL models (3B/7B), which have limited long-context reasoning and tool-use capabilities compared with recent models. It is unclear whether the findings generalize to stronger agents and more challenging long-horizon tasks. I would recommend evaluating at least a 30B+ MoE model from the Qwen3.5 series, along with more diverse long-context agent benchmarks.

  • Lack of statistical reliability analysis. On SearchQA, the 7B CAPS model only outperforms its text-history teacher by 0.5 percentage points (43.5% vs. 43.0%). This margin is quite small, and the paper does not report results across multiple random seeds or statistical significance tests. It is therefore unclear whether the improvement over the text-history baseline is robust or simply due to experimental variance.

Questions for authors
  1. Generalization to stronger models and longer-horizon tasks: Have the authors explored whether the identified policy gap and the effectiveness of CAPS persist with more capable, recent models (e.g., Qwen3.5 30B+ MoE) or more challenging long-context agent benchmarks? Given the limited evaluation with Qwen2.5-VL-3B/7B, additional evidence would be important for assessing the generalizability of the findings.

  2. Statistical reliability: On SearchQA, the 7B CAPS model only outperforms its text-history teacher by 0.5 percentage points (43.5% vs. 43.0%). Are results from multiple random seeds or statistical significance tests available to establish whether this improvement is robust?

3 comments

We thank the reviewer for the constructive feedback and address your points below.

Response to Weakness 1: Evaluation on Newer Models and WebShop

We made every effort to evaluate our approach on newer model architectures and benchmarks. Due to GPU resource constraints, Qwen3.5-9B is the largest advanced model we can train. The results are detailed below:

SearchQA

MethodModelNQTriviaQAPopQAHotpotQA2WikiMuSiQueBamboogleMacro %Tok. Avg.Tok. Max.
Text TeacherQwen3.5-9B47.067.049.448.244.620.449.646.6567.042,515
AgentOCRQwen3.5-9B42.664.847.442.040.617.641.642.4228.041,050
CAPSQwen3.5-9B44.866.846.847.244.621.052.846.3250.01906
MethodModelNQTriviaQAPopQAHotpotQA2WikiMuSiQueBamboogleMacro %Tok. Avg.Tok. Max.
Text TeacherQwen3-VL-4B44.464.346.044.339.717.144.042.80.58k13.83k
AgentOCRQwen3-VL-4B35.458.040.834.833.210.526.434.10.25k3.28k
CAPSQwen3-VL-4B41.362.245.240.638.215.640.040.40.27k1.12k

ALFWorld

MethodModelPickLookCleanHeatCoolPick2Overall Success RateTok. Avg.Tok. Max.
Text TeacherQwen3-VL-4B98.3100.082.970.773.994.787.50.66k2.45k
AgentOCRQwen3-VL-4B95.365.887.958.363.556.774.60.20k0.69k
CAPSQwen3-VL-4B100.0100.0100.085.894.196.796.50.18k0.64k

WebShop Benchmark Evaluation:

WebShop is a complex, web-based interactive environment designed to evaluate agents in realistic online shopping scenarios, requiring agents to interact with realistic web interfaces to find and purchase products.

MethodModelAcc %ScoreTokens
AgentOCRQwen2.5-VL-3B56.375.20.42k
SKILL0Qwen2.5-VL-3B66.478.60.49k
Text TeacherQwen2.5-VL-3B67.1981.881,263.26
CAPSQwen2.5-VL-3B69.5382.57290.69

These results demonstrate that across newer models and challenging benchmarks like WebShop, prior history-compression methods still exhibit a noticeable agentic performance gap compared to uncompressed text policies. In contrast, our CAPS method remains highly effective, showing clear improvements over prior undisstilled methods, successfully narrowing the performance gap while maintaining compression efficiency, and even surpassing text history policies on select tasks.

Response to Weakness 2: Statistical Reliability and Robustness

To rigorously address experimental variance and statistical reliability, we conducted two additional independent training runs (totaling 3 runs) and report the mean and standard deviation:

SearchQA

SizeNQTriviaQAPopQAHotpotQA2WikiMuSiQueBamboogleAvg.Tok. Avg.Tok. Max.
3B43.40 ± 0.9160.34 ± 0.5043.94 ± 1.2840.43 ± 0.3638.87 ± 0.9614.93 ± 0.6234.93 ± 1.6739.55 ± 0.28346.19 ± 22.151806.67 ± 352.33
7B44.32 ± 0.3664.50 ± 0.7147.50 ± 1.0243.19 ± 0.5242.78 ± 0.4818.41 ± 0.3943.20 ± 3.4943.42 ± 0.12316.91 ± 27.311500.00 ± 207.85

ALFWorld

SizePickLookCleanHeatCoolPick2SuccessTok. Avg.Tok. Max.
3B94.42 ± 0.4291.85 ± 5.2899.07 ± 1.6097.06 ± 5.0991.62 ± 9.1682.95 ± 6.1492.97 ± 2.07211.49 ± 11.12914.67 ± 42.77
7B98.38 ± 1.5491.80 ± 8.22100.00 ± 0.0091.66 ± 0.5697.55 ± 4.2486.10 ± 4.6194.92 ± 2.44213.00 ± 12.85882.00 ± 50.48

The small standard deviations across three independent training runs confirm that our method's performance improvements are reliable and effective.

We fully acknowledge the reviewer's concern about the small improvement of 7B CAPS over the text-history teacher. Our core work focuses on vision-text compression (VTC) application and optimization for agentic tasks, aiming to reduce the agentic performance gap relative to uncompressed text policies while achieving compression, rather than substantially outperforming text histories. We will adjust our description in Section 5.2 of the revised manuscript to rigorously emphasize statistically comparable performance with uncompressed text policies alongside substantial token cost reductions, while maintaining direct comparisons against competing VTC methods.

Thank you for the detailed response and additional experiments. The new results provide useful evidence, although my concern about generalization to more challenging long-horizon tasks and more advanced agentic models is not fully resolved. Nevertheless, I continue to appreciate the contributions highlighted in my original review, particularly the practical value of compressing agent histories and the analysis distinguishing information recovery from downstream reasoning and decision-making. I will therefore maintain my borderline-positive assessment.

Thank you for your thoughtful follow-up and for recognizing the practical value of history compression and our analysis of information recovery versus downstream decision-making. We agree that further evaluation on more challenging long-horizon tasks and stronger agentic models is an important next step.

We plan to extend CAPS to SWE-bench Verified using larger, code-capable multimodal models. SWE-bench Verified is a human-validated subset of 500 issue-resolution instances from SWE-bench. Given a real GitHub repository and an issue description, a system must generate a code patch, which is evaluated by running tests [1].

Existing evaluations illustrate the interaction horizons relevant to this extension. SWE-Gym reports 30.1 and 31.6 average interaction turns per task on SWE-bench Verified for its fine-tuned Qwen2.5-Coder-14B-Instruct and Qwen2.5-Coder-32B-Instruct agents using OpenHands [2]. In the end-to-end latency experiments we added during this rebuttal, our text-history teachers averaged 13.83 steps per task on ALFWorld (3B; 140 tasks, each evaluated three times) and 2.97 on SearchQA (7B; 1,025 tasks, each evaluated three times).

Separately, our aggregation of all 500 public trajectories for the Devstral Small (2512) submission hosted in the official SWE-bench evaluation repository yields a mean per-episode peak input length of 37.4K tokens [3]. The corresponding values in our ALFWorld and SearchQA latency experiments are 1.10K and 1.26K tokens, respectively.

Given the limited rebuttal time and available GPU resources, we plan to complete this extension after the discussion period and incorporate the results into a subsequent manuscript revision.

[1] SWE-bench Verified: official dataset card and official repository.

[2] Training Software Engineering Agents and Verifiers with SWE-Gym, Table 3, SWE-bench Verified, fine-tuned columns.

[3] Devstral Small (2512): public submission and evaluation records.

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

The paper investigates the performance degradation that occurs when multi-step agents receive their interaction histories as rendered images instead of text. It introduces Cross-modal Agentic Policy Self-distillation (CAPS), which first trains a visual-policy on successful text-history trajectories and then combines GRPO with token level online self-distillation using a same sized text-history teacher. Experiments use Qwen2.5-VL-3B and 7B on a seven-dataset SearchQA suite and ALFWorld. CAPS improves over AgentOCR by 5.0/3.4 percentage points on SearchQA and 15.6/14.5 points on full-history ALFWorld, reaching 93.8%/95.7% ALFWorld success. The authors also perform ablations showing the contributions of offline and online distillation, alongside diagnostic comparisons demonstrating that task success can improve even as history-recovery accuracy declines.

Critical strengths
  1. The paper introduces a well-motivated cross-modal training pipeline that combines offline SFT with GRPO and online self-distillation. Successful text-policy trajectories provide an initial learning signal, while online distillation supplies dense token-level guidance on states visited by the student. This extracts richer supervision from the existing training tasks without introducing additional task instances.

  2. CAPS improves over AgentOCR at both model scales on SearchQA and ALFWorld, with particularly large gains of ~15 points on ALFWorld. The evaluation includes both search and embodied agent tasks, making the empirical contribution broader than a single benchmark result.

Critical weaknesses
  1. The main History QA experiment in Figure 1 does not sufficiently support the claim that OCR is not a major bottleneck. Task success can improve despite declining generic extraction accuracy if the model becomes worse at recovering irrelevant details while preserving or improving recovery of task-relevant evidence. A stronger diagnostic would test whether the model correctly recovers the specific facts needed for the next decision. Alternatively, the authors could compare the same frozen text-history policy using the original history versus a history reconstructed from its rendered image, holding the prompt and decoding settings fixed on actual downstream evals. This would help quantify how much recovery errors affect downstream decisions.

  2. The online self-distillation stage's gains are not uniform across model scales. At the 7B scale, adding online distillation to offline distillation plus GRPO improves average SearchQA accuracy by only 0.6 points on average, but reduces accuracy on five of seven datasets. Also, the authors don't report the breakdown of different stages on the ALFWorld benchmark, where most of the gains occur.

Questions for authors
  1. Could the authors report end-to-end wall-clock time per episode for CAPS, AgentOCR, and the text-history baseline under matched hardware and evaluation settings, including image rendering, visual encoding, generation, and tool/environment interactions? The generation-time comparison in Appendix A.2 is useful, but a complete latency breakdown would clarify whether the reported token savings translate into practical speedups.

  2. How are histories handled when the stacked image exceeds the supported image dimensions or visual-token budget?

6 comments
AuthorsOpenReview ↗

Official Comment by Authors (Part 1/4)

We thank the reviewer for the positive assessment and insightful comments. We particularly appreciate your constructive suggestions, which have greatly sharpened our core analysis.

Response to Weakness 1: OCR vs. Policy Reasoning Diagnostics [1/2]

We thank the reviewer for the valuable comments. Based on your suggestions, we designed two controlled diagnostic experiments to further demonstrate that OCR is not the primary performance bottleneck when employing visual compression for agentic tasks.

Experiment 1: Full Downstream Comparison of Original Text and Image-Reconstructed Histories

We randomly sampled 1,025 questions from the SearchQA test data (150 each from NQ, TriviaQA, PopQA, HotpotQA, 2Wiki, MuSiQue, and 125 from Bamboogle). We evaluated our trained text-history model (the text-history teacher, 7B) on actual tasks where the history input was set to one of four conditions: original text, raw un-trained Qwen2.5-VL-7B base model transcripts, AgentOCR transcripts, and CAPS transcripts extracted verbatim from history images (all evaluated using 7B). Each condition performed retrieval and interaction using its own policy actions. The results are presented below alongside CAPS operating directly on visual history:

ConditionNQTriviaQAPopQAHotpotQA2WikiMuSiQueBamboogleMacro EM
Original Text43.3366.0054.0048.0040.6726.6739.2045.410
Base Transcript43.3365.3353.3348.0038.6726.0036.8044.495
AgentOCR Transcript45.3365.3350.6740.0036.0028.0035.2042.933
CAPS Transcript34.6764.6749.3338.6728.0020.6731.2038.171
CAPS (Visual History)41.3365.3357.3341.3344.0026.6747.2046.171

Consistent with Figure 1(b) in our paper, CAPS does not show clear improvement in generic OCR capability for recovering specific decision facts; instead, its transcript score decreases compared to AgentOCR. However, when CAPS answers directly from visual histories, it slightly outperforms the text-history teacher using textual history. This further proves that OCR capability is not the performance bottleneck for vision-text compression (VTC) methods in agentic tasks; rather, our method successfully improves the model's policy, delivering significant gains in agentic performance.

AuthorsOpenReview ↗

Official Comment by Authors (Part 2/4)

Response to Weakness 1: OCR vs. Policy Reasoning Diagnostics [2/2]

Experiment 2: Recovery of Facts Relevant to the Next Step

We randomly sampled 500 actual states from the text-history teacher's evaluation trajectories on the SearchQA test set (all evaluated using the 7B model). Using DeepSeek (deepseek-v4.1-flash-1017) conditioned solely on the original question and prior history, we synthesized 2,211 questions across three distinct categories:

  • Task-relevant fact questions (872 items): Assess whether the model can accurately read and extract historical information that is directly useful for answering the original question (e.g., core entities, attributes, and relationships).

  • Irrelevant-fact control questions (839 items): Assess whether the model can correctly extract other background details from the history that are present in the text/images but not required for the current reasoning step, checking for perceptual distraction or noise vulnerability.

  • Absent-information control questions (500 items): Check whether the model correctly returns NOT_IN_HISTORY as instructed when the specified information is genuinely absent from the history, rather than fabricating an answer (evaluating hallucination resistance under visual compression).

We evaluated AgentOCR, CAPS, and the untrained Qwen2.5-VL base model using visual histories (all 7B):

ModelRelevant Fact Answers (%)Irrelevant Fact Answers (%)Absent-Info Control Questions (%)
AgentOCR91.34590.263100.00
CAPS89.03785.848100.00
Base VLM90.84187.740100.00

Furthermore, 414 of the 500 states had at least one task-relevant fact question that passed filtering. From these 414 states, we retained the 299 states where both AgentOCR and CAPS successfully answered all corresponding task-relevant fact questions. We then prompted models to execute the actual next step (search or answer) on these same 299 states. Response validity was evaluated via DeepSeek (deepseek-v4.1-flash-1017) using binary scoring (acceptable = 1, unacceptable = 0). "Action acceptability" evaluates whether the model's next step is reasonable given the current question and history. In our agentic framework, the action space consists of strictly two choices: continuing to search or providing the final answer.

The practical criteria for an action to be judged as "acceptable" are:

  • For Answer actions: Acceptable if the history already supports the answer (with correct entities, relations, dates, etc.), making it appropriate to stop and answer.
  • For Search action: Acceptable if there is still an active information gap and the formulated search query effectively retrieves missing facts or resolves necessary ambiguity.
ModelHistory InputAcceptable-Action Rate (Macro-Avg)
AgentOCRVisual History (7B)47.306%
CAPSVisual History (7B)54.409%
Text TeacherTextual History (7B)53.833%

CAPS shows no clear OCR capability enhancement over AgentOCR or the base model, yet achieves superior performance on agentic tasks. This confirms that OCR capability is not the primary performance bottleneck for visual-history compression in agentic tasks.

AuthorsOpenReview ↗

Official Comment by Authors (Part 3/4)

Response to Weakness 2: Stage-Wise Breakdown & Ablation Analysis

Our additional evaluation results indicate that for the 7B model, offline SFT distillation (Off-SD) already yields substantial improvements, bringing the model much closer to the text policy. This suggests that for 7B models, our offline distillation trajectories are of exceptionally high quality, enabling the model to internalize the strong text policy primarily through SFT. Subsequent gains from GRPO and online distillation are smaller but remain positive. Specifically, online distillation (On-SD) acts as a precise policy regularizer on student-visited states, contributing targeted accuracy gains on hard multi-hop instances while further optimizing context token costs.

Both offline and online distillation share the objective of distilling the high-quality text-history policy, but their improvement directions and magnitudes vary across model scales. The 3B model gains more in task performance without further token optimization, whereas the 7B model gains moderately in performance while achieving further token reduction. Exploring the specific mechanisms behind these distillation variants is a compelling direction for future work. We will expand Section 6 of the revised manuscript to include the following discussion:

Future Work. While offline trajectory self-distillation initializes the visual-history policy from successful text-history trajectories, online policy self-distillation provides dense supervision on student-visited states. Building on our stage-wise findings, an important next step is to examine how model scale and task characteristics influence the distribution of states encountered after offline initialization, as well as the complementary role of online supervision. Such an analysis will help clarify the distinct performance gains and context-token costs observed across different settings.

SearchQA Stage-Wise Results (3B and 7B):

SizeMethod VariantNQTriviaQAPopQAHotpotQA2WikiMuSiQueBamboogleAvg. EMTok. Avg.Tok. Max.
3BOff-SD39.857.841.129.129.05.611.230.50.17k0.76k
3BOff-SD → GRPO40.659.543.239.636.214.028.837.40.27k3.38k
3BFull CAPS42.560.144.440.137.815.434.439.20.37k1.76k
7BOff-SD42.661.344.639.436.815.432.839.00.25k1.86k
7BOff-SD → GRPO45.764.747.843.741.118.738.442.90.35k1.68k
7BFull CAPS44.464.247.142.942.318.345.643.50.29k1.74k

For ALFWorld, we evaluated stage-wise ablations primarily on the 3B model. On ALFWorld, applying either distillation method alone yields more noticeable gains compared to SearchQA, which may relate to ALFWorld's specific task characteristics. Nonetheless, full CAPS achieves the largest overall improvement and optimal trade-off in token cost and performance.

ALFWorld Stage-Wise Results (Qwen2.5-VL-3B):

Method VariantPickLookCleanHeatCoolPick2Success Rate (%)Tok. Avg.Tok. Max.
Full CAPS (Main Paper)94.4491.6797.22100.0093.0183.7793.75201.08952
Visual Off-SD → GRPO100.0088.6997.8382.7886.3460.8786.72227.22826
GRPO + On-SD100.00100.00100.0085.7186.6649.4387.11252.30994
Matched Text SFT → Visual GRPO92.1476.1994.3978.4675.0554.8780.47280.811,036
Reviewer_RDFxOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper first discusses the reasons for the capability loss induced by vision-text compression in agentic tasks. The results identify an agentic policy gap rather than OCR quality alone. Then, it proposes a two-stage cross-modal training framework, which includes SFT, GRPO, and OPSD, to transfer the text-history policy to the visual-history policy. It then evaluates the proposed method on Search-QA and ALFWorld, showing performance gains with substantially lower context costs.

Critical strengths
  1. The paper is well-written and first uses detailed analysis to disentangle the two reasons that lead to the performance degradation caused by vision-text compression: reading/OCR ability and the agentic policy itself.

  2. The paper uses a simple yet effective approach that transfers text-history capabilities to visual-history capabilities through offline and online trajectory distillation.

  3. Experiments show performance improvements with substantial context compression on two datasets.

Critical weaknesses
  1. The paper does not report variance across multiple training runs, particularly on ALFWorld. Given the relatively small improvements over the text-history teacher, it is unclear whether these gains are statistically significant。

  2. The experiments are limited to Qwen2.5-VL with 3B and 7B parameter. Although the results demonstrate effectiveness across different model sizes, it remains unclear whether the identified agentic policy gap and the proposed CAPS framework generalize to other VLM families.

  3. The experiments are limited to two relatively simple benchmarks, SearchQA and ALFWorld.

Questions for authors

Please see the weaknesses above, mainly points 2 and 3. I'm wondering whether the findings could be generalized to different VLM families and more challenging benchmarks.

4 comments
AuthorsOpenReview ↗

Official Comment by Authors (Part 1/2)

We sincerely thank the reviewer for the thorough evaluation and constructive feedback. We deeply appreciate your valuable suggestions, which have helped us significantly improve the completeness of our work.

Response to Weakness 1: Statistical Reliability & Teacher Baseline Clarification

1. Multi-Run Variance and Statistical Reliability

We sincerely thank the reviewer for the constructive suggestion. To rigorously demonstrate the statistical reliability and robustness of our method, we conducted two additional independent training runs with different random seeds (totaling 3 runs). The mean and standard deviation across these three independent runs are reported below:

SearchQA

SizeNQTriviaQAPopQAHotpotQA2WikiMuSiQueBamboogleAvg.Tok. Avg.Tok. Max.
3B43.40 ± 0.9160.34 ± 0.5043.94 ± 1.2840.43 ± 0.3638.87 ± 0.9614.93 ± 0.6234.93 ± 1.6739.55 ± 0.28346.19 ± 22.151806.67 ± 352.33
7B44.32 ± 0.3664.50 ± 0.7147.50 ± 1.0243.19 ± 0.5242.78 ± 0.4818.41 ± 0.3943.20 ± 3.4943.42 ± 0.12316.91 ± 27.311500.00 ± 207.85

ALFWorld

SizePickLookCleanHeatCoolPick2SuccessTok. Avg.Tok. Max.
3B94.42 ± 0.4291.85 ± 5.2899.07 ± 1.6097.06 ± 5.0991.62 ± 9.1682.95 ± 6.1492.97 ± 2.07211.49 ± 11.12914.67 ± 42.77
7B98.38 ± 1.5491.80 ± 8.22100.00 ± 0.0091.66 ± 0.5697.55 ± 4.2486.10 ± 4.6194.92 ± 2.44213.00 ± 12.85882.00 ± 50.48

The tightly bounded standard deviations across the three independent training runs confirm that the performance improvements achieved by our method are highly reliable and effective.

2. Clarification on the Text-History Teacher Comparison & Comparative Framing

We fully acknowledge the reviewer's concern about the relatively small improvement of CAPS over the text-history teacher. Our core work focuses on the application and optimization of vision-text compression (VTC) for agentic tasks. Our primary objective is to achieve substantial context compression while narrowing the agentic performance gap relative to the uncompressed text-history policy, rather than aiming to substantially outperform text histories.

We will revise the wording and emphasis in Section 5.2 of the revised manuscript, modifying expressions such as “exceeds its teacher” to more precisely emphasize that CAPS achieves statistically comparable performance to the uncompressed text-history policy while substantially reducing average token costs. Our direct comparisons remain focused on alternative VTC-compression methods (such as AgentOCR and SKILL0). CAPS outperforms these baselines on most SearchQA datasets and achieves higher overall success rates on ALFWorld.

AuthorsOpenReview ↗

Official Comment by Authors (Part 2/2)

Response to Weakness 2: Evaluation on Newer Models and Benchmarks

We have evaluated our approach on newer model architectures and benchmarks. Due to GPU resource constraints, Qwen3.5-9B is the largest advanced model we can train. The results are detailed below:

SearchQA

MethodModelNQTriviaQAPopQAHotpotQA2WikiMuSiQueBamboogleMacro %Tok. Avg.Tok. Max.
Text TeacherQwen3.5-9B47.067.049.448.244.620.449.646.6567.042,515
AgentOCRQwen3.5-9B42.664.847.442.040.617.641.642.4228.041,050
CAPSQwen3.5-9B44.866.846.847.244.621.052.846.3250.01906
MethodModelNQTriviaQAPopQAHotpotQA2WikiMuSiQueBamboogleMacro %Tok. Avg.Tok. Max.
Text TeacherQwen3-VL-4B44.464.346.044.339.717.144.042.80.58k13.83k
AgentOCRQwen3-VL-4B35.458.040.834.833.210.526.434.10.25k3.28k
CAPSQwen3-VL-4B41.362.245.240.638.215.640.040.40.27k1.12k

ALFWorld

MethodModelPickLookCleanHeatCoolPick2Overall Success RateTok. Avg.Tok. Max.
Text TeacherQwen3-VL-4B98.3100.082.970.773.994.787.50.66k2.45k
AgentOCRQwen3-VL-4B95.365.887.958.363.556.774.60.20k0.69k
CAPSQwen3-VL-4B100.0100.0100.085.894.196.796.50.18k0.64k

These results indicate that while AgentOCR (which naively applies visual history compression) successfully reduces token overhead on newer models, it still suffers from a pronounced agentic performance gap relative to uncompressed text policies. In contrast, our CAPS method remains highly effective across newer architectures, showing substantial gains over AgentOCR without distillation while significantly closing the performance gap to uncompressed text policies, even outperforming text policies on certain tasks while preserving compression efficiency.

Response to Weakness 3: Evaluation on WebShop

WebShop is a complex, web-based interactive environment designed to evaluate agents in realistic online shopping scenarios. Agents must interact with realistic web interfaces to find and purchase products matching user requirements, representing a highly challenging task for VTC-compressed agents.

MethodModelAcc %ScoreTokens
AgentOCRQwen2.5-VL-3B56.375.20.42k
SKILL0Qwen2.5-VL-3B66.478.60.49k
Text TeacherQwen2.5-VL-3B67.1981.881,263.26
CAPSQwen2.5-VL-3B69.5382.57290.69

On the challenging WebShop benchmark, prior VTC compression baselines (AgentOCR and SKILL0) continue to exhibit an agentic performance gap compared to uncompressed text policies. Our method remains highly effective: compared to AgentOCR and SKILL0, it enhances agentic capabilities while maintaining strong compression performance, effectively eliminating the performance gap between text-history and visual-history policies.

Thank you again for your thoughtful questions and constructive feedback. As the discussion period draws to a close, we would appreciate hearing whether our responses and additional experiments have addressed your concerns. If any questions remain or any points would benefit from further clarification, please let us know so that we can address them before the discussion concludes. Thank you for your time and consideration.

Dear Reviewer RDFx,

As the discussion period closes tonight, we would greatly appreciate it if you could review our rebuttal and let us know whether our responses address your concerns. If any questions remain, we would be happy to clarify them before the deadline.

Thank you for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.