{"id":"99ef2e76-0b30-4410-a91f-fd1881688e20","arxiv_id":"2608.08960","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Vision-text compressed agents underperform because of an agentic policy gap, not OCR alone, and cross-modal self-distillation from the model's own text policy substantially closes this gap.","lead":"This paper shows that language-model agents using compressed visual histories fail not because they cannot read the history, but because they reason and decide differently from text-history agents. The authors introduce a two-stage self-distillation method that transfers the stronger text-based policy to the visual-based policy, improving task performance while cutting context cost.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The agentic-policy-gap diagnosis depends on an unmeasured information-preservation assumption; an OCR-perfect matched-state subset test can settle whether the gap is perceptual or policy-level.","rationale":"The paper's headline contribution is a causal diagnosis: the vision-text compression gap is not only perceptual but an agentic policy gap, and CAPS exploits this by transferring text-policy behavior. The hinge is Eq. (4): x^T_t and x^I_t are assumed to be semantically corresponding, so any divergence in decisions can be attributed to policy rather than perception. The paper's evidence for this is indirect. History recovery is measured on a separate set of generic extraction questions and only in aggregate; even there image EM lags text EM by 7-15 points. The task gains of RL and CAPS could be achieved by learning which readable fragments matter, while the remaining mismatch on matched states could be concentrated in states where the rendered image is hard to read. The matched-state measurement does not condition on readability, so the agentic-gap conclusion is underdetermined. The proposed OCR-perfect subset test is a direct and inexpensive check: if mismatches vanish where full OCR succeeds, the diagnosis collapses to a perceptual bottleneck; if they persist, the diagnosis is solid. The paper's Limitations already concede renderer sensitivity, so the authors have the infrastructure to run this test. A secondary concern is that CAPS gains over AgentOCR may partly reflect extra SFT/RL compute or a stronger Qwen2.5-VL teacher; this is worth a same-compute visual-only GRPO baseline, but it does not threaten the diagnostic claim as directly. Overall, the reader's CONDITIONAL verdict remains appropriate, and this stress test strengthens rather than changes it.","tokens_in":22405,"tokens_out":7619,"duration_ms":81094,"concrete_test":"On the 10,000 matched-state prefixes from Section B.2, run a high-accuracy OCR system on each rendered history image and retain only states where OCR exactly recovers the full text (queries, document titles, years). Recompute action-type agreement, Text-Search/Visual-Answer rate, and query Jaccard on this OCR-perfect subset. If the policy mismatch persists at roughly the levels in Table 5, the agentic-policy-gap diagnosis is supported; if agreement becomes near-perfect on the subset, the gap is primarily perceptual and the equivalence assumed in Eqs. (4) and (13)-(15) needs revision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on Eq. (4)'s paired contexts x^T_t and x^I_t being semantically equivalent, but the paper never verifies that the rendered visual history preserves all task-relevant information. Its own numbers show it does not fully: History QA image EM is 82.78 vs. 90.03 text for the base VLM, and CAPS drops to 73.70 image vs. 89.35 text (Tables 4 and 8). The argument that RL improves SearchQA while recovery does not only shows task gains do not require better generic recovery; it does not show that the specific matched-state divergences, e.g., 5.22% query Jaccard match and 51.67% Text-Search/Visual-Answer, occur when the relevant content was actually readable. If the renderer's font, layout, or token-budget downsampling hides exactly the queries and entities that drive teacher decisions, then the observed behavioral drift is perceptual, not agentic, and the online distillation in Eqs. (13)-(15) asks the visual student to imitate text it cannot see. The paper acknowledges renderer sensitivity in the Limitations but provides no conditional analysis.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies vision–text compression (VTC) for multi-step LLM agents, where interaction history is rendered as images to save context tokens. It claims that the performance gap between visual-history and text-history agents is not primarily an OCR or perceptual limitation, but an “agentic policy gap”: even when the relevant content is readable, visual-history agents systematically drift in action selection, query formulation, stopping, and evidence use. The paper supports this diagnosis with three analyses: a self-constructed History QA benchmark, a matched-state decision comparison, and a blind trajectory rubric. It then proposes CAPS, a two-stage cross-modal self-distillation method. The offline stage trains the visual-history policy on successful text-history trajectories rendered as images, and the online stage uses a frozen text-history policy as a dense full-vocabulary teacher during GRPO on student-visited states. On SearchQA, CAPS improves over AgentOCR by 5.0 and 3.4 EM points with 3B and 7B backbones; on full-history ALFWorld the gains are 15.6 and 14.5 points, with substantial context-cost reductions relative to matched text-history policies.","tokens_in":22623,"tokens_out":6377,"duration_ms":65210,"significance":"If the diagnosis holds, the paper makes a useful conceptual contribution: it separates perceptual recovery from decision-level policy behavior in VTC agents and provides a concrete training recipe for recovering agentic capability. The matched-state analysis, the trajectory rubric with inverse-probability weighting and blind labels, the two-stage ablation, and the compression-factor sweep are well designed and go beyond simple end-task comparisons. The method itself is simple and plausible, and the reported cost reductions are meaningful for practical agent deployments. The main caveats are that the central diagnosis relies on an unverified information-preservation assumption for the renderer, and that all headline numbers come from single runs without uncertainty estimates. These issues are fixable in revision rather than fatal.","major_comments":[{"comment":"The central diagnosis assumes that the deterministic renderer R preserves all task-relevant information, so that paired contexts x^T_t and x^I_t in Eq. (4) are semantically equivalent. This is never directly measured, and the paper's own numbers show that the visual history is not fully recoverable: the strict History QA image EM is 82.78 vs. 90.03 for the base VLM and 73.70 vs. 89.35 for CAPS (Tables 4 and 8). The matched-state divergences in Table 5 (51.67% Text-Search/Visual-Answer, 5.22% query Jaccard match) could therefore be concentrated in states where the rendered image does not contain the relevant query, title, or year in readable form, which would make the gap perceptual rather than policy-level. The observation that RL improves SearchQA without improving generic History QA rules out “better OCR explains the task gain,” but it does not localize the divergent decisions to states where the content was readable. I request a conditional analysis: restrict the matched-state sample to states where the ground-truth relevant fields (last search query, document title, answer-bearing passage, year) are correctly extracted from the rendered image, and recompute action-type agreement, Text-Search/Visual-Answer rate, and query Jaccard. If the divergence survives on this OCR-perfect subset, the agentic-policy-gap interpretation is supported; if it collapses, the gap should be reclassified as perceptual. This test is feasible with the existing History QA pipeline.","section":"Diagnosing the Visual-History Policy Gap; Eq. (4)"},{"comment":"All main quantitative claims rest on single training runs without standard deviations, significance tests, or repeated seeds. The headline SearchQA gains are modest (+5.0 and +3.4 EM points in Table 1), and the ablation differences in Table 3 are 1–3 points, which could easily be within run-to-run noise for RL training of this kind. Several baseline rows are taken from prior papers (marked † and ‡), so those comparisons may not share identical training budgets, seeds, or evaluation pipelines. Please report at least three seeds for CAPS and the key ablations, with means and standard deviations, and clarify exactly which comparisons were run with matched pipelines.","section":"Tables 1–3"},{"comment":"The ablation isolates the two distillation stages but does not include a compute-matched or teacher-identity control. CAPS receives offline SFT, GRPO, and online distillation, so it is not clear how much of the gain comes from the cross-modal pairing specifically rather than from the additional supervised signal or the extra optimization budget. I ask for a control that spends the same additional optimization on GRPO alone (without any distillation) and a control with a mismatched or non-paired teacher (e.g., a teacher conditioned on a different task or a randomly paired text history). Without such controls, the contribution of cross-modal policy self-distillation, as opposed to more training signal, is not fully identified.","section":"Ablation Studies; Table 3"}],"minor_comments":[{"comment":"The importance-sampling ratio uses π^S_θold, but the student policy is elsewhere denoted π^I_θ; either introduce S explicitly or change the symbol to π^I_θold for consistency with Eq. (11).","section":"Eq. (12)"},{"comment":"Several cross-references are empty: “detailed in Sec. ” and “As shown in Sec. ” appear without section numbers or names; please fill these in.","section":"Introduction and Method"},{"comment":"There is a typo in “our propsoed CAPS”; it should read “our proposed CAPS.”","section":"Contributions paragraph"},{"comment":"The SKILL0 7B row has a formatting issue (“44.40.34k”), and the header “Avg.↑” is misaligned for the token-cost columns; please fix the table layout.","section":"Table 1"},{"comment":"The trajectory rubric relies entirely on an LLM judge with no human agreement study; the paper acknowledges this in the Limitations, but a small human-annotated sample would substantially strengthen the diagnostic claims.","section":"Appendix B.3"},{"comment":"The abstract states that code will be released “in a future release”; please provide a concrete public code or checkpoint link at publication time to support reproducibility.","section":"Abstract / Reproducibility"}],"recommendation":"major_revision","confidential_remarks":"The paper is potentially a good fit for a top AI conference, and I am sympathetic to the central thesis. My main concern is that the agentic-policy-gap diagnosis hinges on an untested semantic-equivalence assumption in Eq. (4), and the suggested conditional OCR-perfect subset test would determine whether the gap is truly policy-level. If that test supports the diagnosis and the statistical-rigor concerns are addressed, I would be happy to accept a revised version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe quick take: this paper is worth engaging. It identifies a real problem—vision-text compression saves tokens but degrades agent behavior—and it provides the cleanest decomposition I've seen of why: generic history recovery stays flat while decision-level behavior drifts. The CAPS method, two-stage cross-modal self-distillation (offline SFT on successful text-policy trajectories plus online KL distillation during GRPO), is well-motivated and the ablations do isolate the two stages and show that the visual input matters. The gains on SearchQA and ALFWorld are consistent across 3B and 7B, and the memory-cost reductions are meaningful.\n\nThe main soft spot: the central claim that the gap is 'agentic' rather than 'perceptual' rests on the assumption that the rendered image preserves all task-relevant information. That assumption is never directly checked. Their own History QA numbers show a persistent image-minus-text EM gap—82.78 vs 90.03 for the base VLM, 73.70 vs 89.35 after CAPS. The matched-state query mismatch rate (only 5.22% token Jaccard agreement) could be exactly the perceptual loss of precise query strings, not a reasoning failure. If the font or token-budget downsampling hides the entities and queries that drive teacher decisions, then the online distillation stage asks the student to imitate text it cannot see. The Limitations section acknowledges renderer sensitivity but doesn't provide the conditional analysis that would settle it. This is the one thing I'd want a referee to push on.\n\nOther issues are more standard: single runs without variance, no code yet, a self-built QA benchmark, and a proprietary judge for the trajectory rubric. Each is addressable, and none is fatal.\n\nBottom line: the paper overreaches slightly in its title and abstract, but the core result—that policy-level distillation across modalities helps and that OCR alone is not the whole story—likely holds in some form. It deserves a serious referee; I'd ask for a matched-state analysis conditioned on whether the relevant text was actually readable, plus code and uncertainty numbers. I'd bring it to reading group.","headline":"A useful diagnosis and a sensible distillation method, but the paper never directly measures whether the renderer preserves the exact information that drives the text policy, so the 'agentic gap' claim is not as clean as the title implies.","tokens_in":23169,"tokens_out":3111,"would_cite":true,"duration_ms":30098,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Visual-history agents can read the compressed context yet reason and act worse; CAPS distills the text policy into the visual one to close the gap.","keywords":["vision–text compression","agentic policy gap","cross-modal self-distillation","LLM agents","reinforcement learning","history compression","SearchQA","ALFWorld"],"falsifier":"If a visual-history policy with near-perfect recovery of every rendered item still made different matched-state decisions from the text policy, the agentic gap would be confirmed; if instead OCR-transcribing the rendered history and feeding that text to the same model erased the behavioral gap, the loss would be perceptual rather than policy-level. A concrete test: run the text-history policy on the OCR output of the rendered image on the matched-state benchmark, and if its action agreement with the true text policy jumps to near 100%, rendering has lost information and the paired-context equivalence fails.","tokens_in":22214,"feed_emoji":"🧠","tokens_out":6891,"duration_ms":61841,"temperature":0.7,"pith_summary":"This paper establishes that the capability loss induced by vision–text compression in multi-step language-model agents is not primarily a reading problem. It argues that visual-history agents may recover the content of a rendered history yet still drift in action selection, query formulation, stopping, and evidence use, a gap it calls the agentic policy gap. The evidence comes from controlled comparisons: reinforcement learning improves task success without improving history-recovery accuracy, matched states yield different decisions under text versus image histories, and text-conditioned trajectories score higher on a blind rubric. To close the gap, the paper introduces CAPS, which uses the same model's stronger text-history policy to supervise its visual-history counterpart, first by imitating successful text trajectories rendered as images and then by online distillation on states the visual policy actually visits during RL. The reported result is that CAPS improves search and embodied-agent performance while cutting memory-context cost substantially.","feed_headline":"They can read, not reason: CAPS closes the agentic gap","feed_subtitle":"Distilling the stronger text-history policy into the visual one lifts SearchQA by 5.0 and ALFWorld by 15.6 points at 3B scale.","key_machinery":"The load-bearing object is the paired-context construction of Eq. (4): for each canonical interaction state, the text history and its deterministic render produce semantically corresponding contexts, so the same model's text-history policy can supervise the visual policy. The first stage, offline trajectory self-distillation, renders successful text-teacher histories as images and trains the visual policy to reproduce the complete teacher response using next-token prediction. The second stage, online policy self-distillation, runs during GRPO: on states the visual policy visits, the frozen text teacher reads the same history as text and the student's own sampled response prefixes, and the student minimizes capped teacher-to-student forward KL along with the GRPO objective. The mechanism's job is to align the visual policy's action distribution with the text policy's on states the visual policy actually visits, rather than only on teacher-visited states.","core_discovery":"The central discovery is that the modality gap in visual-history agents is a decision-level divergence, not an OCR failure: on matched interaction states, visual-history policies answer when text policies would search, differ in query content, and exhibit lower-quality trajectories on evidence uptake, entity tracking, query progression, stopping, and answer grounding. CAPS transfers reasoning and action behavior from the frozen text-history policy to the visual-history policy through two stages of distillation, offline and online. On SearchQA it improves over AgentOCR by 5.0 points with a 3B backbone and 3.4 points with a 7B backbone; on full-history ALFWorld the gains are 15.6 and 14.5 points, and it reduces average memory-context cost by up to 63.3% and peak cost by up to 83.4% relative to matched text-history policies.","pith_inferences":["A testable extension is to measure the agentic gap under different renderers and languages: if the gap size tracks layout and font legibility rather than information content, the policy gap may be partly perceptual in some regimes, and CAPS would need renderer-aware supervision.","Because offline distillation relies on successful text-teacher trajectories, CAPS may transfer poorly to tasks where successful trajectories are rare; a natural extension is to add filtered failure trajectories or reward-relabeled subgoals so the visual policy can learn from near-misses.","The paper's evidence that CAPS lowers generic history-QA while raising task success suggests distillation is task-specific; one could test whether CAPS generalizes to unseen benchmarks or action spaces without retraining the teacher.","The paired-context assumption implies that any renderer change that preserves text content should preserve the transfer; a further test is whether CAPS still closes the gap when the visual history is rendered with different fonts, colors, or widths while keeping the same information."],"forward_implications":["Vision–text compression for agents should be evaluated on policy decisions such as action agreement, stopping, and query content, and on trajectory quality, not just OCR or history-recovery accuracy, since the residual gap after compression is decision-level.","CAPS can match or exceed the text-history teacher on full-history ALFWorld while cutting average context cost by over 60%, so the token savings of visual history need not come at the cost of agent capability.","Aggressive image downsampling backfires: increasing the compression factor from 1.0 to 2.5 drops SearchQA from 43.54% to 35.50% and raises redundant searches, so CAPS obtains its savings from better policies that finish tasks in fewer steps, not from smaller images.","The two distillation stages are complementary: offline-only reaches 37.4/42.9 and online-only 37.1/41.4 average SearchQA EM, while combining them reaches 39.2/43.5 for the 3B/7B backbones.","After CAPS, action-type agreement with the text-history policy rises from 70.87% to 79.53%, and the problematic text-Search-to-visual-Answer mismatch drops by 19.25 points, showing the distillation changes real decisions, not just token statistics."],"supporting_citations":[{"why":"Supplies the AgentOCR visual-history agent and rendering pipeline that CAPS builds on and outperforms.","marker":"Feng et al. 2026b"},{"why":"Provides SKILL0, the strongest visual-history baseline that CAPS compares against on SearchQA and ALFWorld.","marker":"Lu et al. 2026b"},{"why":"Supplies the GRPO reinforcement-learning algorithm used to train both the text-teacher and visual-student policies.","marker":"Shao et al. 2024"},{"why":"Provides the online policy self-distillation formulation that CAPS adapts for cross-modal distillation during RL.","marker":"Zhao et al. 2026"},{"why":"Attributes part of the text-to-image modality gap to degraded reasoning rather than recognition, the claim CAPS extends to agentic policies.","marker":"Sun et al. 2026"},{"why":"Defines the Qwen2.5-VL backbone family on which both teacher and student policies are built.","marker":"Bai et al. 2025"},{"why":"Supplies the Search-R1-style SearchQA training configuration inherited by the baseline and CAPS.","marker":"Jin et al. 2025"},{"why":"Provides the ALFWorld embodied benchmark used for the full-history and short-history evaluations.","marker":"Shridhar et al. 2020"}],"fun_headline_variants":["CAPS closes the agentic gap: reading ≠ reasoning","Vision-text agents reason worse—CAPS distills it back","Distilling text policy into vision: +5.0 on SearchQA","Not OCR: CAPS fixes visual-agent decision drift"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The renderer is assumed to preserve all task-relevant information, so the text and visual histories are semantically equivalent for the same interaction state; if rendering loses or distorts content through layout, truncation, or visual-token limits, the measured gap could be partly perceptual rather than policy-level, and the text teacher's supervision would not transfer cleanly.","fun_headline_variants_meta":{"raw":{"variants":["CAPS closes the agentic gap: reading ≠ reasoning","Vision-text agents reason worse—CAPS distills it back","Distilling text policy into vision: +5.0 on SearchQA","Not OCR: CAPS fixes visual-agent decision drift"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000252,"raw_usage":{"total_tokens":1599,"prompt_tokens":1022,"completion_tokens":577,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":638,"completion_tokens_details":{"reasoning_tokens":507}},"tokens_in":638,"tokens_out":577,"duration_ms":6556,"temperature":1.0,"reasoning_tokens":507,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:18:44.848825+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"If a visual-history policy with near-perfect recovery of every rendered item still made different matched-state decisions from the text policy, the agentic gap would be confirmed; if instead OCR-transcribing the rendered history and feeding that text to the same model erased the behavioral gap, the loss would be perceptual rather than policy-level. A concrete test: run the text-history policy on the OCR output of the rendered image on the matched-state benchmark, and if its action agreement with the true text policy jumps to near 100%, rendering has lost information and the paired-context equivalence fails.","supporting_citations":[],"review_version":1}