Pith. sign in

REVIEW 3 major objections 5 minor 43 references

The Illusion of Visual Tool-Use: A Causal Audit of Thinking with Images

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Most visual tool calls don't change the answer.

desk verdict A careful empirical audit of visual tool-use with a genuinely new step-level estimand and a stable ATE decomposition, but the headline 'no causal effect' language overstates what the local VEG probe can establish. read the letter →

arxiv 2608.06270 v1 pith:FS5UTUET submitted 2026-08-06 cs.AI

Zhiheng Wang , Bo Peng , Lai Wei , Chaochao Lu This is my paper · ORCID
classification cs.AI
keywords visualtool-usethinkingwithimagescausalauditEvidenceGainpolicymiscalibrationmultimodallargelanguagemodelscounterfactualinterventioncrop-and-zoom
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

The paper asks whether the images a vision-language model receives from its own crop-and-zoom tool calls actually cause the final answer. Auditing six open models on five fine-grained perception benchmarks through policy-, trajectory-, and step-level interventions, it finds that most tool-using rollouts are not causally using the returned image: either the observation has no measurable effect on the answer (Calling Without Looking), or it is informative but the tool-call schedule is incoherent (Looking Without Planning). Only a minority of rollouts, labeled Calibrated, use the visual evidence effectively, and this minority is what carries the aggregate accuracy gains of tool-use. The paper terms this the illusion of visual tool-use and argues that benchmark improvements should not be read as evidence that models are thinking with images.

What carries the argument

The central object is a causal graph of a visual tool-use rollout with three routes from input to answer: the direct path, the observation-mediated path $T \to O \to Y$, and the action-induced shortcut $T \to Y$. The paper's step-level estimand, Visual Evidence Gain (VEG), is the difference between the change in the model's probability gap under the real returned crop and the change under a same-shape counterfactual crop, with prefix and tool action held fixed; the probability gap is the model's next-token preference for the correct option minus its strongest distractor, so VEG cancels the shortcut and isolates the visual contribution. A deterministic trajectory-level diagnostic then sorts each rollout into one of four groups (No-call, Calling Without Looking, Looking Without Planning, Calibrated) using the pre-call probability gap, the peak per-call VEG, whether the tool-call limit was hit, and whether calls continued after saturation. An additive decomposition of the policy-level accuracy gain into per-group contributions shows that the Calibrated group is the only consistently positive contributor.

What would settle it

Re-run the four-group diagnostic and ATE decomposition on a held-out fine-grained benchmark with the same thresholds; the central claim fails if the Calibrated subset is not the dominant positive contributor, or if trajectories labeled Calling Without Looking show mean |VEG| comparable to Calibrated trajectories when the counterfactual crop is drawn from a different image rather than a different location in the same image.

Watch

Extended reading notes

Core claim

The central discovery is a structural decoupling between the tool-use policy and the utility of the observations it elicits, which the paper calls policy miscalibration. Using a causal graph of a rollout, the paper separates the intended observation-mediated path (action to observation to answer) from an action-induced shortcut (action to answer) that can shift the answer without any visual evidence. The step-level estimand Visual Evidence Gain (VEG) measures, for each call, the change in the model's preference for the correct option under the real returned crop minus the same change under a counterfactual random crop, holding prefix and action fixed; this cancels the shortcut and isolates the visual contribution. On the V* benchmark, a deterministic four-group diagnostic (No-call, Calling Without Looking, Looking Without Planning, Calibrated) and an additive decomposition of the policy-level accuracy gain show that the Calibrated group is the only consistently positive contributor, while the majority of tool calls are either causally inert or scheduled incoherently. The paper concludes that aggregate gains are real but give an illusion of broad causal effectiveness.

Load-bearing premise

The audit assumes that Visual Evidence Gain — the difference in next-token probability-gap change between a real and a counterfactual crop under a fixed prefix and tool action — captures whether the returned observation causally affects the model's answer; the authors themselves note it is a local probe under a fixed readout, not the full causal effect on an unconstrained final-answer rollout.

Editorial extensions

If this is right

  • A positive policy-level accuracy gain from visual tool-use is compatible with most tool calls being causally inert, so aggregate accuracy alone is not evidence that the model is using the returned images.
  • Mode 1 (Calling Without Looking) rollouts can skip tool calls without losing accuracy, since their observations carry no causal evidence; Mode 2 (Looking Without Planning) rollouts call for an early-stopping rule that commits before late harmful calls.
  • Step-level VEG provides per-step credit that outcome-only reinforcement learning cannot supply, because rewarding only final correctness treats a Calibrated trajectory and a wasted trajectory that both reach the correct answer as identical.
  • An intervention on tool-use should be graded by whether it increases the Calibrated fraction and decreases Mode 1 plus Mode 2 mass, regardless of whether the aggregate ATE moves.

Reading between the lines

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

  • A testable extension of the audit to other tools (OCR, segmentation, video frame selection) would replace the returned text, box, or frame with a plausible alternative; the paper gives this recipe in its limitations but does not run it, so whether CWL and LWP recur beyond crop-and-zoom is open.
  • If the illusion is widespread, leaderboards that reward tool-calling behavior may be selecting for performative tool use, since a model can gain aggregate accuracy by calling tools without using their content; a calibration-style diagnostic would then be a useful public evaluation metric.
  • Because step-level VEG needs token-level access, a black-box analogue that samples multiple free-form answers under real versus counterfactual crops and compares answer distributions could extend the audit to closed models the paper lists as out of scope.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper formulates visual tool-use as a structural causal model over image, query, tool actions, returned observations, and final answer, and audits it with three interventions: a policy-level toggle between tool-use and direct inference, a trajectory-level corruption of all returned observations, and a step-level counterfactual replacement of a single observation. The step-level estimand, Visual Evidence Gain (VEG, Eq. 4), is the difference between next-token probability-gap changes under a real and a counterfactual crop with the same prefix and action. Across six open models and five benchmarks, the authors report that aggregate accuracy gains from tool-use are real but concentrated in a small 'Calibrated' subset, while most tool-using rollouts are either 'Calling Without Looking' (returned observations have no causal effect on the answer) or 'Looking Without Planning' (observations are informative but the call schedule is incoherent).

Significance. If the claims held, this would be a valuable causal audit protocol for the thinking-with-images paradigm, with a falsifiable trajectory-level diagnostic (Algorithm 1) and an operational decomposition of the policy-level ATE (Eq. 6). The empirical care is real: dynamic corruption avoids leakage from clean observations, the force-answer ablation separates truncation from evidence loss, repeated seeds and threshold sweeps appear in appendices, and code is released. The central reservation is that VEG is an immediate next-token readout, not an estimate of the effect of an observation on the final free-form answer; the manuscript itself concedes this in App. A.3, yet the headline CWL claim is phrased as a causal claim about the answer. This gap does not invalidate the diagnostic as a local probe, but it does mean that the paper's strongest conclusion currently outruns the evidence.

major comments (3)
  1. [§3.3.3, Eq. (4); Abstract; §4.1] The headline claim that in CWL trajectories 'returned observations have no causal effect on the answer' rests on VEG, which is a controlled contrast in next-token option-restricted probability gaps after appending '<answer>' to a fixed prefix. Appendix A.3 explicitly states that VEG is 'a local counterfactual estimate ... not a claim that we recover the full effect of the tool call on unconstrained final-answer formation.' The manuscript's own trajectory-level results (Table 2 and the force-answer ablation in App. B.4, Table 11) show that corrupting all observations produces large accuracy drops even when truncation is controlled (Mini-o3: 87.8 to 36.6; Qwen3-VL-8B: 91.1 to 73.0). If most individual calls have near-zero VEG while trajectory-level corruption strongly affects answers, the observation effect must either accumulate across calls or propagate through subsequent reasoning; either way, per-call VEG is not validated as a measure of causal effect on the final answer. Please either validate VEG against an answer-level counterfactual (e.g., ablating a single real observation in a full rollout and measuring final-answer changes, where feasible) or reframe the CWL and Calibrated claims as statements about the local next-token readout, not about causal effect on the answer.
  2. [§4.2, Algorithm 1; §4.3, Eq. (6)] The diagnostic defines 'Calibrated' as trajectories with V_max >= epsilon and no saturation, and Eq. (6) then shows that this Calibrated group carries most of the positive ATE. Because VEG is the very quantity used to define the group, part of the association between 'Calibrated' and positive evidence contribution is definitional. The independent part is the per-group accuracy gap, which is genuinely empirical, but its causal interpretation inherits the unvalidated VEG-to-answer link from the previous comment. The paper should state this explicitly and either supply an external validation of the group labels (the behavioral audit in App. E.3, such as pre-call commitment rates, is a useful start) or weaken the causal wording to 'groups defined by the local VEG readout.'
  3. [§4.3, Table 5] The central quantitative claim that the Calibrated subset contributes the majority of the positive ATE is reported without uncertainty quantification. V* has only 191 items, and some group sizes are small (e.g., DeepEyes Calibrated is 7.9% of trajectories), so the per-group contributions f_b * Delta_policy,b in Table 5 may be noisy. Since the authors already report repeated-seed results for VEG (App. C.1) and trajectory corruption (App. B.5), please add a similar repeated-seed or bootstrap analysis for the ATE decomposition, or at least report per-group confidence intervals, before treating the decomposition as a quantitative finding.
minor comments (5)
  1. [Abstract; §3.3.3] The paper says 'six representative models' and 'five benchmarks,' but the step-level VEG analysis (Table 3) and the trajectory diagnostic (Tables 4–5) cover only a subset; please clarify early which models and benchmarks are used at each intervention level.
  2. [Table 2] The row for DeepEyes in Table 2 is visually corrupted (accuracy values and Delta run together); please reformat the table.
  3. [Eq. (5); App. C.3] Notation is inconsistent between POER in Eq. (5), which is defined for k>=2 and g0 <= tau_sat, and the 'Post-saturation Over-Extension Rate' description in §3.3.3; please align the definition with the per-call breakdown in App. C.2.
  4. [Figure 4] The caption reads 'sat = 0.95' rather than 'tau_sat = 0.95'; please fix for consistency.
  5. [§3.3.3] The sentence 'By saturation conditions VEG on pre-call confidence' is ungrammatical; it should read 'By saturation status, VEG is conditioned on pre-call confidence.'

Circularity Check

1 steps flagged · score 4.0 of 10

The 'no causal effect' wording for Calling Without Looking is built into the VEG threshold that defines the label, while the underlying frequency measurements and ATE localization are empirical.

  1. self definitional [Abstract; §3.3.3 Eq. (4); §4.2 Algorithm 1]
    "In Calling Without Looking, returned observations have no causal effect on the answer. ... if g0 > τsat or Vmax < ϵ then return MODE1 (CWL) ▷ saturated prior or inactive call"

    Algorithm 1 assigns the CWL label precisely when the pre-first-call gap exceeds τsat or the peak per-call VEG is below ϵ = 0.01, and VEG (Eq. 4) is the paper's own operational measure of whether the observation's visual content contributes to the answer. The abstract's statement that in CWL 'returned observations have no causal effect on the answer' is therefore entailed by the labeling rule rather than derived from it. The genuinely empirical discoveries are the frequencies of the label across models (e.g., 72.8% for DeepEyes) and the independent behavioral signatures, which the paper does provide. The causal wording is a definitional consequence of thresholding VEG, not an independently established property of the so-labeled trajectories.

full rationale

The paper's quantitative core is not fitted or self-referential: VEG is a measured counterfactual contrast between real and counterfactual crops under a fixed prefix and action, and the policy-level ATE is an independent accuracy comparison. The trajectory-level corruption and force-answer experiments are additional external checks, and the threshold sensitivity sweeps show that the Calibrated-concentration result is not an artifact of one cutoff. There is no load-bearing self-citation: the Wei et al. (2026) marginal-improvement claim is corroborated by the paper's own Table 1, and the other co-authored references are contextual. The main circularity is semantic: the CWL failure mode is defined by a near-zero VEG threshold and then reported as 'no causal effect on the answer,' so that headline phrasing reduces to the definition. The paper partially mitigates this by validating the labels against independent think-block commitment and by explicitly conceding in App. A.3 that VEG is a local counterfactual estimate under a fixed readout, 'not a claim that we recover the full effect of the tool call on unconstrained final-answer formation.' That concession limits the reach of the causal wording but does not repair the definitional circularity in the abstract. Overall, the empirical prevalence numbers and the ATE decomposition carry independent content, so the circularity is partial rather than total.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical or ontological entities; its inventions are estimands, diagnostic labels, and experimental interventions. The main free quantities are the two thresholds of the trajectory diagnostic, both checked by sensitivity analysis. The load-bearing assumptions concern the causal graph, the local logit readout as a proxy for answer causality, and the prompt-level direct baseline.

free parameters (2)
  • saturation threshold tau_sat = 0.95
    Hand-chosen probability gap above which a call is declared saturated; used in VEG analysis, POER, and the diagnostic (Algorithm 1). Sensitivity sweeps in App. E.1 show the qualitative conclusions are stable for tau_sat >= 0.90.
  • near-zero VEG cutoff epsilon = 0.01
    Threshold separating no-effect calls from informative calls in the diagnostic. Varying epsilon in App. E.1 does not change the model-level ordering or the conclusion that positive ATE is concentrated in the Calibrated subset.
assumptions (5)
  • standard math Pearl's structural causal model framework and its path-effect semantics are assumed as the formal foundation.
    Section 3.1 invokes Pearl's SCM, ATE, and NIE definitions; the causal graph in Figure 2 relies on these standard definitions.
  • domain assumption The tool-use rollout is generated by the SCM Ti <- f_pi(I,Q,T<i,O<i), Oi <- E_tool(I,Ti), Y <- f_pi(I,Q,T1:n,O1:n), with no unobserved confounders between observations and the answer.
    Section 3.2. The estimands and the claim that VEG isolates observation-mediated influence depend on this structural specification and on the absence of hidden confounding.
  • domain assumption The option-restricted next-token readout after appending '<answer>' tracks the model's answer preference, and VEG under a fixed prefix is a valid proxy for causal effect on final answers.
    Section 3.3.3 and App. A.3. The authors explicitly note VEG is a local probe, not a full free-form answer effect, which makes this assumption load-bearing for the failure-mode labels.
  • domain assumption Counterfactual same-shape random crops are exchangeable with real crops except for visual content, so the real-versus-counterfactual difference in probability gaps isolates visual evidence.
    Eq. 4 and the RANDOM-CROP construction in App. A.3. The counterfactual crop is sampled from elsewhere in the same image to preserve modality and crop format.
  • domain assumption Direct inference obtained by replacing the model's system prompt with a generic 'reason step by step and boxed answer' prompt is a valid do(pi=pi_direct) intervention that isolates tool availability.
    Section 3.3.1 and App. A.1. The prompt substitution may change instruction-following behavior beyond removing tools, so the policy-level ATE is partly confounded by prompt mismatch.

how reviews work

0 comments
Cite this review

Pith. "Pith review of The Illusion of Visual Tool-Use: A Causal Audit of Thinking with Images." pith.science (2026). https://pith.science/paper/FS5UTUET

@misc{pith2026260806270,
  author       = {Pith},
  title        = {Pith review of: The Illusion of Visual Tool-Use: A Causal Audit of Thinking with Images},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FS5UTUET}},
  note         = {Machine review of arXiv:2608.06270}
}
read the original abstract

The "thinking-with-images" paradigm equips multimodal LLMs with active visual operations such as crop-and-zoom. However, models using these operations often achieve only marginal or negative gains over direct inference at substantially higher token cost. They may also repeatedly crop irrelevant regions and fail on questions that direct inference answers correctly. We ask whether the returned visual evidence causally affects the answer. To answer this question, we formulate visual tool-use as a causal graph that separates observation-mediated paths from action-induced shortcuts. We then audit it through interventions at the three levels: policy (comparing tool-use with direct inference), trajectory (corrupting all observations during rollout), and step (counterfactually replacing one individual observation under a fixed prefix). Our step-level estimand, Visual Evidence Gain, isolates the contribution of each returned observation. Across six representative models and five fine-grained perception benchmarks, we uncover policy miscalibration with two failure modes. In Calling Without Looking, returned observations have no causal effect on the answer. In Looking Without Planning, observations are informative but the call schedule is incoherent. A trajectory-level diagnostic decomposes the policy-level accuracy gain and shows that the gain is concentrated in a Calibrated minority. We term this discrepancy the illusion of visual tool-use: despite aggregate accuracy gains, visual tool-use is not causally effective across a broad range of rollouts. The code is available at https://github.com/OpenCausaLab/CauAudit.

Figures

Figures reproduced from arXiv: 2608.06270 by the authors.

Figure 1
Figure 1. The paradox of visual tool-use. (a) Our experiments show that, despite far more tokens, tool-use policies yield marginal or no improvements over direct inference. (b) A qualitative failure on a V∗ instance (by Mini-o3): the direct prediction is correct, but tool￾augmented inference repeatedly crops irrelevant regions and leads to no valid answer or a wrong answer. Zhang et al., 2025b; Zheng et al., 2025; Wang et al.… view at source ↗
Figure 2
Figure 2. Causal graph for visual tool-use. (a) The single-step causal graph for a single tool-use step over (I, Q, T, O, Y ), highlighting the direct path (I, Q)→Y (blue), the observation-mediated path T →O→Y (green) and the action-induced shortcut T →Y (red). (b) A full trajectory composes n such atomic graphs. cess supervision can improve call quality, but does not decide whether a call is needed; ViRL still av￾erages abou… view at source ↗
Figure 3
Figure 3. Three-level causal intervention. (a) Policy-level intervention (§3.3.1) toggles the entire T →O →Y subgraph. (b) Trajectory-level intervention (§3.3.2) replaces every real observation with a corrupted one O˜ i during inference. The policy reacts to corrupted feedback and may produce a different action trace T˜ 1:˜n. (c) Step-level intervention (§3.3.3) replaces the real observation Oreal i with a counterfactual one … view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Per-call scatter of probability gap before call gi−1 vs. VEGi on V∗ . Each point is one tool call; color indicates the trajectory’s final correctness (C/I); dashed line marks τsat = 0.95. free-form answer-probing rollout. At each check￾point i, we take the generated pr…
Figure 5
Figure 5. Figure 5: Within-trajectory dynamics on the V∗ benchmark. (a) Mean VEG by call position k ∈ {1, 2, 3, ≥ 4}; Solid curves indicate mean on all calls and dashed curves indicate mean on non-saturated subset only. (b) Per-position Over-Extension Rate (POER, Eq. 5, per-call). Full pe…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

43 extracted references · 15 canonical work pages

  1. [1]

    Thinking with images , year =

    work page

  2. [2]

    2025 , howpublished =

    Introducing. 2025 , howpublished =

    2025

  3. [3]

    thinking with images

    Deepeyes: Incentivizing" thinking with images" via reinforcement learning , author=. arXiv preprint arXiv:2505.14362 , year=

    arXiv

  4. [4]

    arXiv preprint arXiv:2505.15966 , year=

    Pixel reasoner: Incentivizing pixel-space reasoning with curiosity-driven reinforcement learning , author=. arXiv preprint arXiv:2505.15966 , year=

    arXiv

  5. [5]

    arXiv preprint arXiv:2509.07969 , year=

    Mini-o3: Scaling up reasoning patterns and interaction turns for visual search , author=. arXiv preprint arXiv:2509.07969 , year=

    arXiv

  6. [6]

    arXiv preprint arXiv:2602.01334 , year=

    What Does Vision Tool-Use Reinforcement Learning Really Learn? Disentangling Tool-Induced and Intrinsic Effects for Crop-and-Zoom , author=. arXiv preprint arXiv:2602.01334 , year=

    arXiv

  7. [7]

    arXiv preprint arXiv:2509.20912 , year=

    DeFacto: Counterfactual Thinking with Images for Enforcing Evidence-Grounded and Faithful Reasoning , author=. arXiv preprint arXiv:2509.20912 , year=

    arXiv

  8. [8]

    arXiv preprint arXiv:2511.19661 , year=

    CodeV: Code with Images for Faithful Visual Reasoning via Tool-Aware Policy Optimization , author=. arXiv preprint arXiv:2511.19661 , year=

    arXiv

Show all 43 references
  1. [9]

    2009 , publisher=

    Causality , author=. 2009 , publisher=

    2009

  2. [10]

    arXiv preprint arXiv:2602.11858 , year=

    Zooming without Zooming: Region-to-Image Distillation for Fine-Grained Multimodal Perception , author=. arXiv preprint arXiv:2602.11858 , year=

  3. [11]

    arXiv preprint arXiv:2506.23918 , year=

    Thinking with Images for Multimodal Reasoning: Foundations, Methods, and Future Frontiers , author=. arXiv preprint arXiv:2506.23918 , year=

    arXiv

  4. [12]

    Proceedings of the Seventeenth Conference on Uncertainty in Artificial Intelligence (UAI) , pages=

    Direct and indirect effects , author=. Proceedings of the Seventeenth Conference on Uncertainty in Artificial Intelligence (UAI) , pages=

  5. [13]

    Proceedings of the 19th International Joint Conference on Artificial Intelligence (IJCAI) , pages=

    Identifiability of path-specific effects , author=. Proceedings of the 19th International Joint Conference on Artificial Intelligence (IJCAI) , pages=

  6. [14]

    Epidemiology , volume=

    Identifiability and exchangeability for direct and indirect effects , author=. Epidemiology , volume=

  7. [15]

    The Fourteenth International Conference on Learning Representations , year=

    Thyme: Think Beyond Images , author=. The Fourteenth International Conference on Learning Representations , year=

  8. [16]

    arXiv:2602.20739 , year=

    PyVision-RL: Forging Open Agentic Vision Models via RL , author=. arXiv:2602.20739 , year=

  9. [17]

    Wu, Penghao and Xie, Saining , booktitle=. V ^

  10. [18]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Divide, conquer and combine: A training-free framework for high-resolution image perception in multimodal large language models , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  11. [19]

    2025 , url=

    YiFan Zhang and Huanyu Zhang and Haochen Tian and Chaoyou Fu and Shuangqing Zhang and Junfei Wu and Feng Li and Kun Wang and Qingsong Wen and Zhang Zhang and Liang Wang and Rong Jin , booktitle=. 2025 , url=

    2025

  12. [20]

    arXiv preprint arXiv:2511.21631 , year=

    Qwen3-vl technical report , author=. arXiv preprint arXiv:2511.21631 , year=

    arXiv

  13. [21]

    Advances in Neural Information Processing Systems , year=

    Language models don't always say what they think: Unfaithful explanations in chain-of-thought prompting , author=. Advances in Neural Information Processing Systems , year=

  14. [22]

    arXiv preprint arXiv:2307.13702 , year=

    Measuring faithfulness in chain-of-thought reasoning , author=. arXiv preprint arXiv:2307.13702 , year=

    arXiv

  15. [23]

    International Conference on Machine Learning , pages=

    On the Relationship Between Explanation and Prediction: A Causal View , author=. International Conference on Machine Learning , pages=. 2023 , organization=

    2023

  16. [24]

    arXiv preprint arXiv:2402.14897 , year=

    Chain-of-thought unfaithfulness as disguised accuracy , author=. arXiv preprint arXiv:2402.14897 , year=

    arXiv

  17. [25]

    arXiv preprint arXiv:2305.20050 , year=

    Let's verify step by step , author=. arXiv preprint arXiv:2305.20050 , year=

    arXiv

  18. [26]

    arXiv preprint arXiv:2211.14275 , year=

    Solving math word problems with process- and outcome-based feedback , author=. arXiv preprint arXiv:2211.14275 , year=

    arXiv

  19. [27]

    ACL , year=

    Math-Shepherd: Verify and reinforce LLMs step-by-step without human annotations , author=. ACL , year=

  20. [28]

    arXiv preprint arXiv:2410.08146 , year=

    Rewarding progress: Scaling automated process verifiers for LLM reasoning , author=. arXiv preprint arXiv:2410.08146 , year=

    arXiv

  21. [29]

    arXiv preprint arXiv:2606.30251 , year=

    TACO: Tool-Augmented Credit Optimization for Agentic Tool Use , author=. arXiv preprint arXiv:2606.30251 , year=

    arXiv

  22. [30]

    arXiv preprint arXiv:2511.23031 , year=

    From Illusion to Intention: Visual Rationale Learning for Vision-Language Reasoning , author=. arXiv preprint arXiv:2511.23031 , year=

  23. [31]

    The Fourteenth International Conference on Learning Representations , year=

    Information gain-based policy optimization: A simple and effective approach for multi-turn search agents , author=. The Fourteenth International Conference on Learning Representations , year=

  24. [32]

    Aho and Jeffrey D

    Alfred V. Aho and Jeffrey D. Ullman , title =. 1972

    1972

  25. [33]

    Publications Manual , year = "1983", publisher =

    1983

  26. [34]

    Chandra and Dexter C

    Ashok K. Chandra and Dexter C. Kozen and Larry J. Stockmeyer , year = "1981", title =. doi:10.1145/322234.322243

    1981

  27. [35]

    Scalable training of

    Andrew, Galen and Gao, Jianfeng , booktitle=. Scalable training of

  28. [36]

    Dan Gusfield , title =. 1997

    1997

  29. [37]

    Tetreault , title =

    Mohammad Sadegh Rasooli and Joel R. Tetreault , title =. Computing Research Repository , volume =. 2015 , url =

    2015

  30. [38]

    A Framework for Learning Predictive Structures from Multiple Tasks and Unlabeled Data , Volume =

    Ando, Rie Kubota and Zhang, Tong , Issn =. A Framework for Learning Predictive Structures from Multiple Tasks and Unlabeled Data , Volume =. Journal of Machine Learning Research , Month = dec, Numpages =

  31. [39]

    arXiv preprint arXiv:2512.01827 , year=

    CauSight: Learning to Supersense for Visual Causal Discovery , author=. arXiv preprint arXiv:2512.01827 , year=

  32. [40]

    Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages=

    Cello: Causal evaluation of large vision-language models , author=. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages=

    2024

  33. [41]

    International Conference on Learning Representations , volume=

    Beyond Surface Structure: A Causal Assessment of LLMs' Comprehension Ability , author=. International Conference on Learning Representations , volume=

  34. [42]

    Findings of the Association for Computational Linguistics: ACL 2025 , pages=

    From imitation to introspection: Probing self-consciousness in language models , author=. Findings of the Association for Computational Linguistics: ACL 2025 , pages=

    2025

  35. [43]

    Findings of the Association for Computational Linguistics: EMNLP 2024 , pages=

    Quantifying and mitigating unimodal biases in multimodal large language models: A causal perspective , author=. Findings of the Association for Computational Linguistics: EMNLP 2024 , pages=

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.