Pith. sign in

REVIEW 3 major objections 7 minor 26 references

A self-judging multi-agent loop makes agentic vision models use image tools only when the crop actually helps.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-31 14:05 UTC pith:FTU2ALUE

load-bearing objection Solid systems fix for unfaithful tool use: dual use of self-judgement as context and helpful-ratio reward, with real ablations—judge validation is the soft spot, not a collapse. the 3 major comments →

arxiv 2607.28225 v1 pith:FTU2ALUE submitted 2026-07-30 cs.CV

FaithEyes: Towards Faithful Tool Use via Multi-Agent Process-Image Verification

Haoqing Wang , Xingrun Xing , Wei Xia , Ziheng Li , Yehui Tang This is my paper
classification cs.CV
keywords agentic VLMstool faithfulnessprocess imagesmulti-agent self-judginghelpful-tool ratioGRPOvisual perceptionreward hacking
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Agentic vision-language models can call tools such as cropping or code-based image edits while they reason, yet many of those calls produce process images that miss the queried target. The model still answers correctly by leaning on prior knowledge or the original image, so the tools become decorative and waste computation. FaithEyes attributes this to two gaps: rewards that credit any tool call when the answer is right, and tool feedback that never says whether the returned image is useful. The method has the same model act as a subagent that judges each process image, injects that verdict into the next reasoning step, drops unhelpful images, and scales the tool reward by the helpful-tool ratio. Trained with supervised fine-tuning then reinforcement learning on adapted open data, it matches or beats prior 7B agentic models on perception and reasoning benchmarks while sharply raising the share of trajectories whose process images truly contain the needed evidence.

Core claim

Unfaithful tool use in agentic VLMs stems from undifferentiated tool rewards and usefulness-agnostic observations. FaithEyes closes both gaps at once: the model itself, under a subagent prompt, emits a binary helpfulness verdict plus rationale for every process image; that verdict is folded into the tool observation (keeping helpful images, discarding unhelpful ones) and simultaneously defines a tool reward equal to the fraction of executable, helpful calls, so decorative calls earn nothing.

What carries the argument

Multi-agent process-image verification: the same policy plays main agent (solver) and subagent (judge). The subagent maps question plus process image to a JSON verdict {is_helpful, reasons}; the main agent conditions on that verdict, and the helpful-tool ratio r_tool = 1 − (n_fail + n_unhelpful)/n_tool scales the tool reward independently of answer correctness.

Load-bearing premise

Binary helpfulness labels taken from trajectory structure in the open training data, plus the same 7B model acting as its own judge, are assumed accurate enough both to steer reasoning and to define the tool reward.

What would settle it

On correct-answer trajectories that contain process images, measure the fraction whose images actually show the queried target (using an independent strong judge or human labels); if FaithEyes does not raise that fraction substantially over baselines while keeping accuracy, the central claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Agentic VLMs can stop receiving credit for decorative crops and will be pushed to re-crop or fall back when a process image misses the target.
  • Inference cost drops because unhelpful process images are discarded from context rather than kept as visual tokens.
  • Train-test consistency is preserved without an external judge model at deployment, because the policy itself supplies the verdict.
  • Perception benchmarks that hinge on small distant objects gain the most, while dense numerical reasoning tasks gain less.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same self-judging loop could be applied to non-visual tools (search, code execution) wherever intermediate outputs currently lack an explicit usefulness signal.
  • If judge accuracy saturates, a modest external judge used only at training time might further lift faithfulness without breaking the no-external-model inference constraint.
  • Text-biased attention patterns noted in the paper suggest that process-level faithfulness is necessary but not yet sufficient for true answer-level reliance on retrieved evidence.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 7 minor

Summary. FaithEyes targets unfaithful tool use in agentic VLMs, where process images are often decorative or off-target yet still receive full tool credit. The authors attribute this to undifferentiated tool rewards and usefulness-agnostic tool feedback. They propose a self-calling multi-agent design: a main agent solves the question with code-based image tools; the same model under a subagent prompt emits a binary helpfulness verdict and rationale for each process image. Helpful images are returned with the verdict; unhelpful ones are dropped and only the verdict is kept. The same helpfulness labels scale a tool reward by the helpful-tool ratio (Eq. 4), independent of answer correctness. After cold-start SFT on adapted Thyme trajectories (structural helpfulness labels plus generated rationales) and GRPO RL with accuracy, format, consistency, and tool rewards, a 7B model reports competitive or superior accuracy on V*, HR-Bench 4K/8K, MathVista, and MathVerse versus prior 7B agentic baselines from the same base, with substantially higher external-judge faithfulness ratios (Figure 3). Ablations (Table 3), λ_tool sweeps (Figure 4), keep-unhelpful controls, attention rollout (Table 4, Figure 2), and training dynamics (Figures 5–6) support complementarity of judgement injection and reward scaling and motivate decoupling tool reward from accuracy.

Significance. Unfaithful or decorative tool calls are a documented failure mode of agentic VLMs and undermine both efficiency and the claim that models are “thinking with images.” Dual use of a process-level helpfulness signal—as context that steers re-cropping or fallback, and as a ratio-based reward that blocks reward hacking—is a clean and transferable design. Keeping the judge inside the same model removes external-judge dependence at inference and enforces train–test consistency, which is practically valuable. The empirical package is stronger than many concurrent agentic-VLM papers: same-base 7B comparisons, complementary ablations, keep-unhelpful controls, attention probes, and an explicit analysis of why accuracy-gated tool rewards collapse tool use (Figure 6). If the faithfulness gains track genuine target visibility rather than judge collusion, the recipe is a useful process-level alignment pattern for multimodal tool agents.

major comments (3)
  1. [§4.1 Data preparation; §3.2 Subagent] §4.1 and §3.2: SFT helpfulness labels are assigned purely from trajectory structure (single-call ⇒ helpful; first of two-call ⇒ unhelpful, second ⇒ helpful), with Qwen3-VL-32B only writing rationales that match those fixed labels (Prompt D.2), not verifying them. The central faithfulness claim requires that ht mean “queried target is visible.” Without human or independent verification that structural labels and the subsequent 7B self-judge agree with true target visibility on the training distribution—including the clarity/crop-concentration preferences the authors themselves flag in §3.3—both cold-start and r_tool (Eq. 4) may optimize a proxy. Please report judge–human (or strong held-out judge) agreement on a labeled sample of SFT/RL process images, and discuss systematic failure modes (lazy near-full-frame crops counted helpful; small on-target but blurry crops counted unhelpful).
  2. [§3.2; Eq. (4); Table 3] §3.2 and Eq. (4): Main agent and subagent share θ. Because r_tool is defined by the subagent’s ht, a lenient or collusive self-judge can raise the tool reward without better crops. Figure 3’s external 235B metric and Table 3’s keep-unhelpful protocol show the final policy improves under an independent automatic judge, which is reassuring, but they do not bound training-time self-judge error or gaming. The external-judge row in Table 3 lifts faithfulness further while accuracy stays flat—suggesting headroom and residual self-judge bias. A short analysis of main/subagent disagreement rates over training, or a frozen-external-judge RL control, would make the non-collusion claim load-bearing rather than inferred.
  3. [Figure 3; Prompt D.3; §4.2] Figure 3 / Prompt D.3: Faithfulness is defined as “trajectory contains at least one helpful process image” under a top-tier automatic judge with a prompt distinct from training. That is better than self-evaluation alone, but the paper never reports human agreement with D.3, nor inter-judge agreement between D.2-style and D.3-style verdicts on the same images. Given that the abstract’s headline claim is “markedly improving tool faithfulness,” the metric needs a brief human calibration (even on a few hundred V*/HR-Bench trajectories) or an explicit limitation that faithfulness is automatic-judge-relative.
minor comments (7)
  1. [Table 2; §4.2] Table 2: FaithEyes trails CodeV on MathVision (29.9 vs 33.6). The text attributes this to limited room for visual tools; a one-sentence breakdown of tool-call rate and helpfulness on MathVision versus MathVista would make that claim concrete.
  2. [Eq. (4); §3.3 Reward design] Eq. (4): Define r_tool when n_tool = 0 explicitly (currently only the n_tool > 0 branch is written). State whether trajectories with no tool calls receive 0 or are left undefined in the advantage.
  3. [Table 1; §3.1] Table 1: Injecting external judgements into frozen DeepEyes/Thyme helps most benchmarks but hurts DeepEyes on HR-Bench. The motivation for joint training is clear; a short note on calibration mismatch would help readers who might try judgement-only prompting.
  4. [Figure 3] Figure 3 caption and bar annotations are hard to parse in the rendered text (garbled glyph runs). Ensure vector labels and “faithful / total” counts are legible in the camera-ready figure.
  5. [Prompt D.1; §4.1] Prompt D.1 tells the main agent a “critic subagent will help… You must carefully consider its judgment,” which is appropriate for FaithEyes but should be noted as part of the trained interface so comparisons to baselines without that instruction remain fair.
  6. [§2] Related work: briefly contrast with CodeV’s process-level judge-as-reward-only design earlier in §2 (you already do this later); a single sentence up front would sharpen the novelty claim.
  7. [Abstract; passim] Minor prose: “This may stem from two limitations” (abstract) and occasional article issues (“the judgement”); standardize “judgment/judgement” spelling.

Circularity Check

1 steps flagged

No significant circularity: accuracy and reported faithfulness are external; self-judge reward is a transparent training design, not a by-construction prediction.

specific steps
  1. other [§3.2–3.3, Eq. (4); §4.1 data prep]
    "r_tool = 1 − (n_fail + n_unhelpful)/n_tool, if n_tool > 0 ... nunhelpful counts those judged as unhelpful by the subagent. ... the model itself serves as a subagent to judge the main agent’s tool-call helpfulness ... Their structure gives natural ground-truth judgement: the sole call in a single tool call trajectory is labeled as ht = True, and in a two tool call trajectory, the first call is labeled as ht = False and the second one is labeled as ht = True."

    Training-time faithfulness credit is defined by the same policy (under a subagent prompt) that is being optimized, and SFT ht labels are taken from demo structure rather than independent visibility annotation. This is a self-referential training signal, not a derivation that forces the external accuracy or 235B-judged faithfulness numbers by construction; those remain independently measured. Flagged only as mild other-type self-reference, not as a collapsed prediction.

full rationale

FaithEyes is an empirical methods paper, not a first-principles derivation. Its load-bearing claims are (i) higher accuracy on held-out perception/reasoning benchmarks and (ii) higher process-image faithfulness under an external Qwen3-VL-235B judge with a prompt (D.3) distinct from the training subagent prompt (D.2). Neither quantity is defined by, or algebraically forced by, the training-time helpful-tool ratio r_tool or the structural SFT labels. The dual use of subagent ht in observation and in Eq. 4 is an intentional self-consistency design (train-test parity without an external judge at inference), not a case where a claimed prediction reduces to its fitted input. SFT ‘ground-truth’ ht from Thyme trajectory structure (single-call ⇒ helpful; first of two-call ⇒ unhelpful) is a labeling heuristic for cold-start, not a uniqueness theorem or a renamed known law. Mild self-reference exists—the policy is optimized partly against a same-θ subagent—but Table 3’s keep-unhelpful protocol and the external-235B faithfulness metric show the reported gains are not equivalent to that internal signal by construction. No self-citation uniqueness chain, no fitted parameter re-presented as prediction, and no equation that equals its inputs. Score 1 only for the transparent self-referential reward loop, which is not load-bearing for the external claims.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 2 invented entities

Empirical ML methods paper. Load-bearing commitments are modeling and data-construction choices, not physical axioms. Free parameters are RL/SFT hyperparameters and reward weights. Domain assumptions include standard VLM tool-use setup, GRPO, and automatic judges for rewards and metrics. Invented entities are framework roles (main/subagent, helpful-tool ratio reward), not new physical objects.

free parameters (3)
  • λ_tool (and λ_fmt, λ_cons) = 0.2 (fmt=cons=0.2)
    Reward coefficients in r(τ); chosen with accuracy dominant. Ablated for λ_tool ∈ {0.1,0.2,0.4,0.8}; default 0.2 selected where faithfulness saturates before accuracy drops.
  • SFT/RL optimization hyperparameters = SFT 1e-5 3ep bs128; RL 1e-6 bs128 G=12; max 5 tool calls
    LR, epochs, batch size, rollouts, token/tool caps affect the reported policy; taken largely from Thyme/GRPO practice rather than derived.
  • Attention rollout depth and top-effective k = last 4 layers; k=round(exp(H))
    Last-four-layer rollout and k=round(exp(H)) are analysis choices that shape Table 4 attention claims, not the trained policy itself.
axioms (5)
  • domain assumption A process image is helpful iff the queried object/attribute is visible so target attributes can be identified; otherwise unhelpful (subagent rubric).
    Defines ht used both in context and in r_tool; §3.2 and Prompt D.2. Operational, not proved.
  • ad hoc to paper In adapted Thyme SFT demos, single-call trajectories are helpful and in two-call trajectories the first call is unhelpful and the second helpful.
    Structural pseudo-labels for cold-start judgement and interactive reasoning; §4.1. Not independently validated.
  • domain assumption GRPO with group-normalized advantages and KL to SFT reference is a valid RL update for this multimodal tool policy.
    Standard in recent LLM/VLM RL; §3.3 Eq. (2).
  • domain assumption LLM-as-judge (Qwen2.5-VL-72B) can score answer correctness and reasoning-answer consistency when rules fail.
    Used inside racc and rcons; §3.3. Common practice with known noise.
  • ad hoc to paper Dropping unhelpful process images and returning only the verdict does not remove information the main agent needs beyond the verdict text.
    Asymmetric observation design §3.2; supported by keep-unhelpful control but still a design premise.
invented entities (2)
  • FaithEyes main-agent / subagent self-calling loop no independent evidence
    purpose: Produce per-process-image helpfulness verdicts at train and test without an external model.
    Organizational construct over one VLM with two prompts; not a new physical entity. Independent evidence is empirical benchmark behavior only.
  • Helpful-tool ratio tool reward r_tool no independent evidence
    purpose: Credit only executable and subagent-helpful calls to suppress decorative tool use.
    Defined in Eq. (4) from n_fail, n_unhelpful, n_tool. Falsifiable only via downstream faithfulness/accuracy metrics.

pith-pipeline@v1.2.0-daily-grok45 · 27481 in / 3948 out tokens · 83113 ms · 2026-07-31T14:05:07.747960+00:00 · methodology

0 comments
read the original abstract

Agentic vision-language models (VLMs), which interleave textual reasoning with explicit tool calls such as cropping and code-based image manipulation, have emerged as a compelling paradigm for reliable and interpretable multimodal reasoning. However, recent studies have revealed that such models often use tools unfaithfully. Many process images are irrelevant to the question (e.g., the tool crops the wrong region or misses the queried target), yet the call still receives full credit and the model still answers correctly. Such decorative or misaligned tool calls waste computation and reveal that the model leans on prior knowledge or the original image rather than the evidence it retrieves. This may stem from two limitations of prevailing methods: the tool reward fails to distinguish useful from useless calls, and tool feedback carries no signal of usefulness. To this end, we introduce FaithEyes, a multi-agent self-judging framework. Concretely, we use a VLM to judge whether each process image helps answer the question. The judgement is injected into the reasoning context as part of the tool observation to help subsequent reasoning, and meanwhile is used to scale the tool reward by the helpful-tool ratio to suppress reward hacking. To keep judgement available at evaluation and thus ensure train-test consistency, we further design a multi-agent framework where the model itself serves as a subagent to judge the tool calls from main agent, eliminating any dependence on an external model at inference. Training via a two-stage SFT + RL pipeline on adapted open-source data, FaithEyes attains competitive or superior accuracy across visual perception and reasoning benchmarks, while markedly improving tool faithfulness. The homepage is at https://github.com/Mosi-AI/FaithEyes.

Figures

Figures reproduced from arXiv: 2607.28225 by Haoqing Wang, Wei Xia, Xingrun Xing, Yehui Tang, Ziheng Li.

Figure 1
Figure 1. Figure 1: Illustration of our FaithEyes framework and interaction trajectory. The main agent calls tools to obtain the process images, and the subagent gives the judgement to help further reasoning. correct answers only because the model can shortcut through prior knowledge or the original image. Indeed, removing the process images barely changes its predictions (Liu et al., 2025; Yang et al., 2026). Consequently, t… view at source ↗
Figure 2
Figure 2. Figure 2: Attention maps of agentic VLMs on process images. We examine the spatial focus [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Tool faithfulness ratio on V∗ , HR-Bench 4K and HR-Bench 8K. The bottom of the bar shows the number of faithful tool trajectories and the total trajectories. More data and training details are provided in the appendix. 4.2 Main results In this work, we mainly examine visual perception (i.e, V∗ , HR-Bench 4K and HR-Bench 8K) and reasoning (i.e., MathVista, MathVerse and MathVision) benchmarks [PITH_FULL_IM… view at source ↗
Figure 4
Figure 4. Figure 4: Tool call number, accuracy and tool faithfulness ratio on different tool reward [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Training dynamics of accuracy reward, format reward, consistency reward, tool [PITH_FULL_IMAGE:figures/full_fig_p016_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Tool failure ratio and tool call number of acc-dependent and independent tool [PITH_FULL_IMAGE:figures/full_fig_p017_6.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

26 extracted references · 18 linked inside Pith

  1. [4]

    Refocus: Visual editing as a chain of thought for structured image understanding.arXiv preprint arXiv:2501.05452,

    Xingyu Fu, Minqian Liu, Zhengyuan Yang, John Corring, Yijuan Lu, Jianwei Yang, Dan Roth, Dinei Florencio, and Cha Zhang. Refocus: Visual editing as a chain of thought for structured image understanding.arXiv preprint arXiv:2501.05452,

    Pith/arXiv arXiv

  2. [5]

    Advancing vision-language models in front-end development via data synthesis.arXiv preprint arXiv:2503.01619,

    Tong Ge, Yashu Liu, Jieping Ye, Tianyi Li, and Chao Wang. Advancing vision-language models in front-end development via data synthesis.arXiv preprint arXiv:2503.01619,

    Pith/arXiv arXiv

  3. [6]

    Large language model based multi-agents: A survey of progress and challenges.arXiv preprint arXiv:2402.01680,

    Taicheng Guo, Xiuying Chen, Yaqi Wang, Ruidi Chang, Shichao Pei, Nitesh V Chawla, Olaf Wiest, and Xiangliang Zhang. Large language model based multi-agents: A survey of progress and challenges.arXiv preprint arXiv:2402.01680,

    Pith/arXiv arXiv

  4. [8]

    Gpt-4o system card.arXiv preprint arXiv:2410.21276,

    12 AI Model TF July, 2026 Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card.arXiv preprint arXiv:2410.21276,

    Pith/arXiv arXiv 2026

  5. [10]

    Llava-onevision: Easy visual task transfer

    Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326,

    Pith/arXiv arXiv

  6. [11]

    Encouraging divergent thinking in large language models through multi-agent debate.arXiv preprint arXiv:2305.19118,

    Tian Liang, Zhiwei He, Wenxiang Jiao, Xing Wang, Yan Wang, Rui Wang, Yujiu Yang, Shuming Shi, and Zhaopeng Tu. Encouraging divergent thinking in large language models through multi-agent debate.arXiv preprint arXiv:2305.19118,

    Pith/arXiv arXiv

  7. [12]

    Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101,

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101,

    Pith/arXiv arXiv

  8. [14]

    Chatdev: Communicative agents for software development.arXiv preprint arXiv:2307.07924,

    Chen Qian, Wei Liu, Hongzhang Liu, Nuo Chen, Yufan Dang, Jiahao Li, Cheng Yang, Weize Chen, Yusheng Su, Xin Cong, et al. Chatdev: Communicative agents for software development.arXiv preprint arXiv:2307.07924,

    Pith/arXiv arXiv

  9. [15]

    Deepseekmath: Pushing the limits of mathe- matical reasoning in open language models.arXiv preprint arXiv:2402.03300,

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. Deepseekmath: Pushing the limits of mathe- matical reasoning in open language models.arXiv preprint arXiv:2402.03300,

    Pith/arXiv arXiv

  10. [16]

    Thinking with images for multimodal reasoning: Foundations, methods, and future frontiers.arXiv preprint arXiv:2506.23918,

    Zhaochen Su, Peng Xia, Hangyu Guo, Zhenhua Liu, Yan Ma, Xiaoye Qu, Jiaqi Liu, Yanshu Li, Kaide Zeng, Zhengyuan Yang, et al. Thinking with images for multimodal reasoning: Foundations, methods, and future frontiers.arXiv preprint arXiv:2506.23918,

    Pith/arXiv arXiv

  11. [17]

    Kwai keye-vl technical report.arXiv preprint arXiv:2507.01949,

    Kwai Keye Team, Biao Yang, Bin Wen, Changyi Liu, Chenglong Chu, Chengru Song, Chongling Rao, Chuan Yi, Da Li, Dunju Zang, et al. Kwai keye-vl technical report.arXiv preprint arXiv:2507.01949,

    Pith/arXiv arXiv

  12. [18]

    Multi-agent collaboration mechanisms: A survey of llms.arXiv preprint arXiv:2501.06322,

    13 AI Model TF July, 2026 Khanh-Tung Tran, Dung Dao, Minh-Duong Nguyen, Quoc-Viet Pham, Barry O’Sullivan, and Hoang D Nguyen. Multi-agent collaboration mechanisms: A survey of llms.arXiv preprint arXiv:2501.06322,

    Pith/arXiv arXiv 2026

  13. [19]

    From illusion to intention: Visual rationale learning for vision-language reasoning.arXiv preprint arXiv:2511.23031, 2025a

    Changpeng Wang, Haozhe Wang, Xi Chen, Junhan Liu, Taofeng Xue, Chong Peng, Donglian Qi, Fangzhen Lin, and Yunfeng Yan. From illusion to intention: Visual rationale learning for vision-language reasoning.arXiv preprint arXiv:2511.23031, 2025a. Haozhe Wang, Alex Su, Weiming Ren, Fangzhen Lin, and Wenhu Chen. Pixel reasoner: Incentivizing pixel-space reasoni...

    arXiv

  14. [20]

    Divide, conquer and combine: A training-free framework for high- resolution image perception in multimodal large language models

    Wenbin Wang, Liang Ding, Minyan Zeng, Xiabin Zhou, Li Shen, Yong Luo, Wei Yu, and Dacheng Tao. Divide, conquer and combine: A training-free framework for high- resolution image perception in multimodal large language models. InProceedings of the AAAI Conference on Artificial Intelligence, pp. 7907–7915, 2025c. Yuntao Wang, Yanghe Pan, Zhou Su, Yi Deng, Qu...

    1906

  15. [21]

    Autogen: Enabling next-gen LLM applications via multi-agent conversations.arXiv preprint arXiv:2308.08155,

    Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Beibin Li, Erkang Zhu, Li Jiang, Xiaoyun Zhang, Shaokun Zhang, Jiale Liu, et al. Autogen: Enabling next-gen LLM applications via multi-agent conversations.arXiv preprint arXiv:2308.08155,

    Pith/arXiv arXiv

  16. [22]

    Visual planning: Let’s think only with images.arXiv preprint arXiv:2505.11409,

    Yi Xu, Chengzu Li, Han Zhou, Xingchen Wan, Caiqi Zhang, Anna Korhonen, and Ivan Vuli´ c. Visual planning: Let’s think only with images.arXiv preprint arXiv:2505.11409,

    arXiv

  17. [23]

    Wenxi Yang, Yuzhong Zhao, Fang Wan, and Qixiang Ye

    URLhttps://openreview.net/forum?id=ivAHZuj8k1. Wenxi Yang, Yuzhong Zhao, Fang Wan, and Qixiang Ye. Thinking with images via self- calling agent.arXiv preprint arXiv:2512.08511,

    arXiv

  18. [24]

    thinking with images

    Yi-Fan Zhang, Xingyu Lu, Xiao Hu, Chaoyou Fu, Bin Wen, Tianke Zhang, Changyi Liu, Kaiyu Jiang, Kaibing Chen, Kaiyu Tang, et al. R1-reward: Training multimodal reward model through stable reinforcement learning.arXiv preprint arXiv:2505.02835, 2025a. Yi-Fan Zhang, Xingyu Lu, Shukang Yin, Chaoyou Fu, Wei Chen, Xiao Hu, Bin Wen, Kaiyu Jiang, Changyi Liu, Tia...

    Pith/arXiv arXiv 2026

  19. [25]

    All scores are averaged over the rollouts within the batch

    15 AI Model TF July, 2026 /uni00000018/uni00000013/uni00000014/uni00000013/uni00000013/uni00000014/uni00000018/uni00000013/uni00000015/uni00000013/uni00000013/uni00000015/uni00000018/uni00000013/uni00000016/uni00000013/uni00000013/uni00000016/uni00000018/uni00000013 /uni00000037/uni00000055/uni00000044/uni0000004c/uni00000051/uni0000004c/uni00000051/uni00...

    2026

  20. [26]

    {file_path}

    We plot six curves against the training step: the accuracy, format, and consistency rewards, the tool reward (i.e., the helpful- tool ratio rtool), the average number of tool calls per trajectory, and the average response 16 AI Model TF July, 2026 50 100 150 200 250 300 350 Training Step 0.0 2.5 5.0 7.5 10.0 12.5 15.0 17.5Code Failure Rate (%) FaithEyes (...

    2026

  21. [2017]

    Mathvista: Evaluating mathe- matical reasoning of foundation models in visual contexts.arXiv preprint arXiv:2310.02255,

    Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating mathe- matical reasoning of foundation models in visual contexts.arXiv preprint arXiv:2310.02255,

    Pith/arXiv arXiv

  22. [2020]

    Qwen3-vl technical report.arXiv preprint arXiv:2511.21631, 2025a

    Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, et al. Qwen3-vl technical report.arXiv preprint arXiv:2511.21631, 2025a. Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zha...

    Pith/arXiv arXiv

  23. [2023]

    Thinking with generated images.arXiv preprint arXiv:2505.22525,

    Ethan Chern, Zhulin Hu, Steffi Chern, Siqi Kou, Jiadi Su, Yan Ma, Zhijie Deng, and Pengfei Liu. Thinking with generated images.arXiv preprint arXiv:2505.22525,

    Pith/arXiv arXiv

  24. [2024]

    Controlthinker: Unveiling latent semantics for controllable image generation through visual reasoning.arXiv preprint arXiv:2506.03596,

    Feng Han, Yang Jiao, Shaoxiang Chen, Junhao Xu, Jingjing Chen, and Yu-Gang Jiang. Controlthinker: Unveiling latent semantics for controllable image generation through visual reasoning.arXiv preprint arXiv:2506.03596,

    arXiv

  25. [2025]

    A survey on the optimization of large language model-based agents.arXiv preprint arXiv:2503.12434,

    Shangheng Du, Jiabao Zhao, Jinxin Shi, Zhentao Xie, Xin Jiang, Yanhong Bai, and Liang He. A survey on the optimization of large language model-based agents.arXiv preprint arXiv:2503.12434,

    arXiv

  26. [2026]

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

    Xin Lai, Junyi Li, Wei Li, Tao Liu, Tianjian Li, and Hengshuang Zhao. Mini-o3: Scaling up reasoning patterns and interaction turns for visual search.arXiv preprint arXiv:2509.07969,

    Pith/arXiv arXiv