Pith. sign in

REVIEW 3 major objections 6 minor 46 references

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

T0 review · 3 major / 6 minor · reviewed 2026-08-14 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read 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.

desk verdict 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. read the letter →

arxiv 2608.08960 v1 pith:RTXNDOKK submitted 2026-08-09 cs.AI

Cheng Fan , Junyi Zhou , Tingzhang Luo , RongJian Xu , Qiyanhui Lu , Mingjian Zhu , Hanting Chen , Jianyuan Guo This is my paper · ORCID
classification cs.AI
keywords vision–textcompressionagenticpolicygapcross-modalself-distillationLLMagentsreinforcementlearninghistorySearchQAALFWorld
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

  • 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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • 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.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

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.

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 (3)
  1. [Diagnosing the Visual-History Policy Gap; Eq. (4)] 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.
  2. [Tables 1–3] 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.
  3. [Ablation Studies; Table 3] 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.
minor comments (6)
  1. [Eq. (12)] 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).
  2. [Introduction and Method] Several cross-references are empty: “detailed in Sec. ” and “As shown in Sec. ” appear without section numbers or names; please fill these in.
  3. [Contributions paragraph] There is a typo in “our propsoed CAPS”; it should read “our proposed CAPS.”
  4. [Table 1] 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.
  5. [Appendix B.3] 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.
  6. [Abstract / Reproducibility] 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.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: CAPS is an empirical distillation method, and the 'agentic policy gap' diagnosis is an experimentally supported inference, not a renaming of its inputs.

full rationale

The paper's derivation chain is empirical rather than definitional. It establishes, via History QA, matched-state decision comparisons, and a blind trajectory rubric, that visual-history policies differ from text-history policies in action selection, query formulation, stopping, and evidence use even when RL does not improve generic OCR recovery. The proposed method then uses an independently trained text-history policy as a frozen teacher (Eq. 5) and transfers behavior with standard next-token SFT (Eq. 10), GRPO (Eqs. 11-12), and forward-KL distillation (Eqs. 13-15). No quantity reported as a prediction is fitted from the target result: the teacher is trained on text histories and outperforms the visual policy on the same benchmarks before distillation, and improvements are measured on held-out test splits with five out-of-domain datasets. The main fragility is the semantic-equivalence assumption in Eq. (4): if the deterministic renderer R loses task-relevant content, part of the observed gap could be perceptual rather than policy-level. However, that is a validity risk, not circularity, because the conclusion is falsifiable and the paper's own data (History QA image EM 73.70-82.78 vs. text EM 89.35-90.03) show imperfect recovery without making the target statement true by construction. The Limitations section also acknowledges renderer sensitivity, teacher dependence, and LLM-judge dependence. The cited baselines (AgentOCR, SKILL0, GiGPO, GRPO) are external works, no uniqueness theorem or ansatz is imported from the authors' prior work, and no load-bearing self-citation chain is present. Therefore, no circular step can be exhibited.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central method rests on the premise that rendered images and text histories are semantically interchangeable, and that the text policy is a valid teacher. The quantitative conclusions depend on self-constructed diagnostics and an LLM judge; these are domain assumptions rather than mathematical axioms.

free parameters (5)
  • Online distillation weight lambda = 0.05 (SearchQA), 0.01 (ALFWorld)
    Chosen by hand to balance RL reward and KL distillation; affects final performance.
  • Per-token KL clipping threshold tau = 0.05
    Caps large KL values to stabilize training; chosen without a stated selection rule.
  • Compression factor c = 1.0 default
    Controls image downsampling; the paper shows performance drops sharply for c>1, so c=1 is a favorable choice.
  • Teacher temperature = 1.1
    Used for online distillation teacher scoring; standard but ad hoc.
  • LoRA rank/alpha = 32/64
    Standard LoRA configuration; no ablation.
assumptions (5)
  • domain assumption Rendered visual history is semantically equivalent to the text history for the same interaction state (Eq. 4).
    The entire cross-modal distillation relies on paired contexts being interchangeable except for modality.
  • domain assumption The text-history policy is a stronger and reliable teacher for the visual-history policy.
    The method assumes the text policy's behavior is worth imitating; the paper demonstrates this on benchmarks but it is a premise.
  • domain assumption History QA benchmark measures text recovery ability from rendered images.
    The conclusion that the gap is not OCR rests on this self-constructed benchmark.
  • domain assumption GRPO with LoRA on Qwen2.5-VL is a stable RL training setup inherited from AgentOCR.
    The paper follows AgentOCR's configuration and does not verify stability across seeds.
  • domain assumption The LLM judge DeepSeek-V4-Flash provides valid trajectory quality scores under the given rubric.
    Trajectory quality conclusions rely on a proprietary judge; the paper acknowledges human evaluation would strengthen it.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Reading is not Reasoning: Bridging the Agentic Policy Gap in Vision-Text Compression." pith.science (2026). https://pith.science/paper/RTXNDOKK

@misc{pith2026260808960,
  author       = {Pith},
  title        = {Pith review of: Reading is not Reasoning: Bridging the Agentic Policy Gap in Vision-Text Compression},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RTXNDOKK}},
  note         = {Machine review of arXiv:2608.08960}
}
read the original 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 \textbf{CAPS}, a two-stage \textbf{C}ross-modal \textbf{A}gentic \textbf{P}olicy \textbf{S}elf-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\% with 3B and 7B backbones, respectively. On full-history ALFWorld, the corresponding gains are 15.6\% and 14.5\%. 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 in a future release.

Figures

Figures reproduced from arXiv: 2608.08960 by the authors.

Figure 1
Figure 1. Textual-history vs. visual-history policy. (a) Textual and visual history representations under the same interaction tra￾jectory. (b) Task improvement is not explained by improved history recovery. (c) Visual-history policies make different decisions from text-history policies under matched states. (d) Text-history policies produce higher-quality trajectories across multiple dimensions. Vision–text compression (VTC)… view at source ↗
Figure 2
Figure 2. Overview of CAPS. Offline distillation transfers successful text-policy behavior to visual histories, GRPO optimizes task rewards, and online distillation supervises student-visited states. where D(·∥·) denotes a discrepancy between response dis￾tributions and ρtr is the interaction-state distribution used for cross-modal self-distillation. Offline Trajectory Self-Distillation Trajectory collection and filtering. Fo… view at source ↗
Figure 3
Figure 3. SearchQA performance, memory-context cost, and search behavior under different image compression factors. All points evaluate the same 7B CAPS checkpoint. therefore do not reduce token cost through additional image downsampling in the default setting. Instead, CAPS lowers context cost through a better decision policy that completes tasks with fewer interactions. Policy-Gap Analysis We repeat the diagnostic evaluatio… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Blind trajectory-rubric scores on SearchQA. Each dimension is scored from 0 to 2, and higher is better. Conclusion In this paper, we identify an agentic policy gap when multi￾step histories are represented as images: visual-history agents can recover content yet use it…
Figure 5
Figure 5. Figure 5: Actual visual histories reconstructed from saved successful trajectories with the runtime renderer and no additional [PITH_FULL_IMAGE:figures/full_fig_p017_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

46 extracted references · 4 canonical work pages

  1. [1]

    arXiv preprint arXiv:2112.09332 , year=

    Webgpt: Browser-assisted question-answering with human feedback , author=. arXiv preprint arXiv:2112.09332 , year=

    arXiv

  2. [2]

    NeurIPS 2022 Foundation Models for Decision Making Workshop , year=

    React: Synergizing reasoning and acting in language models , author=. NeurIPS 2022 Foundation Models for Decision Making Workshop , year=

    2022

  3. [3]

    Advances in neural information processing systems , volume=

    Toolformer: Language models can teach themselves to use tools , author=. Advances in neural information processing systems , volume=

  4. [4]

    arXiv preprint arXiv:2510.00615 , year=

    Acon: Optimizing context compression for long-horizon llm agents , author=. arXiv preprint arXiv:2510.00615 , year=

    arXiv

  5. [5]

    arXiv preprint arXiv:2605.05191 , year=

    LongSeeker: Elastic Context Orchestration for Long-Horizon Search Agents , author=. arXiv preprint arXiv:2605.05191 , year=

    arXiv

  6. [6]

    arXiv preprint arXiv:2510.18234 , year=

    Deepseek-ocr: Contexts optical compression , author=. arXiv preprint arXiv:2510.18234 , year=

    arXiv

  7. [7]

    Advances in Neural Information Processing Systems , volume=

    Vision-centric token compression in large language model , author=. Advances in Neural Information Processing Systems , volume=

  8. [8]

    Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Glyph: Scaling context windows via visual-text compression , author=. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    work page

Show all 46 references
  1. [9]

    arXiv preprint arXiv:2601.04786 , year=

    AgentOCR: Reimagining Agent History via Optical Self-Compression , author=. arXiv preprint arXiv:2601.04786 , year=

  2. [10]

    arXiv preprint arXiv:2402.03300 , year=

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models , author=. arXiv preprint arXiv:2402.03300 , year=

    arXiv

  3. [11]

    arXiv preprint arXiv:2601.18734 , year=

    Self-Distilled Reasoner: On-Policy Self-Distillation for Large Language Models , author=. arXiv preprint arXiv:2601.18734 , year=

    arXiv

  4. [12]

    arXiv preprint arXiv:2604.13016 , year=

    Rethinking on-policy distillation of large language models: Phenomenology, mechanism, and recipe , author=. arXiv preprint arXiv:2604.13016 , year=

    arXiv

  5. [13]

    arXiv preprint arXiv:2605.15155 , year=

    Self-distilled agentic reinforcement learning , author=. arXiv preprint arXiv:2605.15155 , year=

    arXiv

  6. [14]

    arXiv preprint arXiv:2604.02268 , year=

    Skill0: In-context agentic reinforcement learning for skill internalization , author=. arXiv preprint arXiv:2604.02268 , year=

    arXiv

  7. [15]

    Advances in Neural Information Processing Systems , volume=

    Group-in-group policy optimization for llm agent training , author=. Advances in Neural Information Processing Systems , volume=

  8. [16]

    arXiv preprint arXiv:2603.09095 , year=

    Reading, not thinking: Understanding and bridging the modality gap when text becomes pixels in multimodal llms , author=. arXiv preprint arXiv:2603.09095 , year=

    arXiv

  9. [17]

    International Conference on Learning Representations , volume=

    Swe-bench: Can language models resolve real-world github issues? , author=. International Conference on Learning Representations , volume=

  10. [18]

    arXiv preprint arXiv:2305.16291 , year=

    Voyager: An open-ended embodied agent with large language models , author=. arXiv preprint arXiv:2305.16291 , year=

    arXiv

  11. [19]

    arXiv preprint arXiv:2010.03768 , year=

    Alfworld: Aligning text and embodied environments for interactive learning , author=. arXiv preprint arXiv:2010.03768 , year=

    2010 arXiv

  12. [20]

    arXiv preprint arXiv:2503.09516 , year=

    Search-r1: Training llms to reason and leverage search engines with reinforcement learning , author=. arXiv preprint arXiv:2503.09516 , year=

    arXiv

  13. [21]

    arXiv preprint arXiv:2305.18290 , year=

    Direct preference optimization: Your language model is secretly a reward model , author=. arXiv preprint arXiv:2305.18290 , year=

    arXiv

  14. [22]

    arXiv preprint arXiv:2409.19256 , year=

    Hybridflow: A flexible and efficient rlhf framework , author=. arXiv preprint arXiv:2409.19256 , year=

    arXiv

  15. [23]

    arXiv preprint arXiv:2506.06122 , year=

    Reinforcement learning optimization for large-scale learning: An efficient and user-friendly scaling library , author=. arXiv preprint arXiv:2506.06122 , year=

    arXiv

  16. [24]

    The Eleventh International Conference on Learning Representations , pages=

    Language modelling with pixels , author=. The Eleventh International Conference on Learning Representations , pages=. 2023 , organization=

    2023

  17. [25]

    arXiv preprint arXiv:2510.18840 , year=

    See the text: From tokenization to visual reading , author=. arXiv preprint arXiv:2510.18840 , year=

  18. [26]

    Nature , pages=

    Multimodal learning with next-token prediction for large multimodal models , author=. Nature , pages=. 2026 , publisher=

    2026

  19. [27]

    arXiv preprint arXiv:2505.09388 , year=

    Qwen3 technical report , author=. arXiv preprint arXiv:2505.09388 , year=

    arXiv

  20. [28]

    The twelfth international conference on learning representations , year=

    Minillm: Knowledge distillation of large language models , author=. The twelfth international conference on learning representations , year=

  21. [29]

    International Conference on Learning Representations , volume=

    Speculative knowledge distillation: Bridging the teacher-student gap through interleaved sampling , author=. International Conference on Learning Representations , volume=

  22. [30]

    International Conference on Learning Representations , volume=

    On-policy distillation of language models: Learning from self-generated mistakes , author=. International Conference on Learning Representations , volume=

  23. [31]

    arXiv preprint arXiv:2601.02780 , year=

    Mimo-v2-flash technical report , author=. arXiv preprint arXiv:2601.02780 , year=

    arXiv

  24. [32]

    Advances in Neural Information Processing Systems , volume=

    Leveraging visual tokens for extended text contexts in multi-modal learning , author=. Advances in Neural Information Processing Systems , volume=

  25. [33]

    arXiv preprint arXiv:2601.21468 , year=

    MemOCR: Layout-Aware Visual Memory for Efficient Long-Horizon Reasoning , author=. arXiv preprint arXiv:2601.21468 , year=

    arXiv

  26. [34]

    arXiv preprint arXiv:2504.19413 , year=

    Mem0: Building production-ready ai agents with scalable long-term memory , author=. arXiv preprint arXiv:2504.19413 , year=

    arXiv

  27. [35]

    Advances in Neural Information Processing Systems , volume=

    A-mem: Agentic memory for llm agents , author=. Advances in Neural Information Processing Systems , volume=

  28. [36]

    arXiv preprint arXiv:2506.15841 , year=

    Mem1: Learning to synergize memory and reasoning for efficient long-horizon agents , author=. arXiv preprint arXiv:2506.15841 , year=

    arXiv

  29. [37]

    Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Triviaqa: A large scale distantly supervised challenge dataset for reading comprehension , author=. Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  30. [38]

    Proceedings of the 61st annual meeting of the association for computational linguistics (volume 1: Long papers) , pages=

    When not to trust language models: Investigating effectiveness of parametric and non-parametric memories , author=. Proceedings of the 61st annual meeting of the association for computational linguistics (volume 1: Long papers) , pages=

  31. [39]

    Transactions of the Association for Computational Linguistics , volume=

    Natural questions: a benchmark for question answering research , author=. Transactions of the Association for Computational Linguistics , volume=. 2019 , publisher=

    2019

  32. [40]

    Proceedings of the 2018 conference on empirical methods in natural language processing , pages=

    HotpotQA: A dataset for diverse, explainable multi-hop question answering , author=. Proceedings of the 2018 conference on empirical methods in natural language processing , pages=

    2018

  33. [41]

    Proceedings of the 28th International Conference on Computational Linguistics , pages=

    Constructing a multi-hop qa dataset for comprehensive evaluation of reasoning steps , author=. Proceedings of the 28th International Conference on Computational Linguistics , pages=

  34. [42]

    Transactions of the Association for Computational Linguistics , volume=

    MuSiQue: Multihop Questions via Single-hop Question Composition , author=. Transactions of the Association for Computational Linguistics , volume=. 2022 , publisher=

    2022

  35. [43]

    Findings of the Association for Computational Linguistics: EMNLP 2023 , pages=

    Measuring and narrowing the compositionality gap in language models , author=. Findings of the Association for Computational Linguistics: EMNLP 2023 , pages=

    2023

  36. [44]

    2025 , eprint=

    Qwen2.5-VL Technical Report , author=. 2025 , eprint=

    2025

  37. [45]

    2025 , eprint=

    Qwen2.5 Technical Report , author=. 2025 , eprint=

    2025

  38. [46]

    DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence , author=

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.