Pith. sign in

REVIEW 3 major objections 5 minor 52 references

Visual tool use can be learned as a compositional, adaptive capability rather than a fixed set of calling patterns, and the training recipe that achieves it reaches state-of-the-art open-model results across eight visual reasoning benchmark

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 ↗

VC-Tooler trains Qwen3-VL-8B on hierarchically synthesized tool trajectories via SFT then GRPO with a judge-based tool reward, reaching open-source SOTA on V* (95.8) and VTC-Bench (35.3) and transferring to 11 unseen tools at inference.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection A well-built training pipeline with a clever relabeling trick, but the adaptive-tool-use claim is supported by a weak proxy: unseen tools are scored on call validity, not on whether the model actually uses their returned images. the 3 major comments →

arxiv 2608.02217 v1 pith:K6XZZLPB submitted 2026-08-03 cs.CV

VC-Tooler: Learning Compositional and Adaptive Visual Tool Use

Yizheng Wu , Jiashen Hua , Bing Deng , Jieping Ye This is my paper
classification cs.CV
keywords visual tool useagentic multimodal reasoningcompositional tool useadaptive tool usetrajectory synthesisreinforcement learningtool rewardschema generalization
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.

The reading

This paper argues that visual tool use—zooming, rotating, querying, editing images—should be learned as a general capability, not a memorized set of invocation patterns. The authors synthesize a large trajectory bank covering single-tool grounding, multi-step composition, and thousands of alternative tool interfaces, then train a vision-language model with supervised fine-tuning followed by reinforcement learning that explicitly rewards using tool-returned information. The resulting model, VC-Tooler, reaches state-of-the-art results among open models on V* (95.8%), VTC-Bench (35.3%), and six other benchmarks, and invokes eleven previously unseen tools with schema-level success rates between 73.9% and 100%. If the claim holds, the recipe—synthetic trajectories plus a tool-feedback reward—offers a scalable path to agentic visual reasoning without hand-engineering tool-specific supervision.

Core claim

The paper's central claim is that visual tool use is a compositional and adaptive skill: a system should ground each tool call in the current visual state, compose multiple tools in sequence, and re-ground its decisions in the observations those tools return, even when the available tools were never seen during training. VC-Tooler instantiates this by building a hierarchical trajectory bank—plan-then-execute synthesis produces single-tool and multi-tool trajectories, and a reinstantiation step relabels the same visual transition under more than a thousand mock tool schemas—then training in two stages: a cold-start SFT that establishes grounding, composition, and adaptation, followed by reinf

What carries the argument

The central mechanism is the diverse-tool-context reinstantiation step: given an observed visual transition (input image, reasoning, output image), a strong VLM is asked to either match the operation to an existing tool or invent a new schema and arguments that realize the same transition. The observation is kept intact, so the model sees many different interfaces for the same underlying visual operation. This is combined with a plan-then-execute synthesizer for single- and multi-tool trajectories, and a two-stage training loop whose RL stage uses a five-dimension tool reward (plan coherence, call efficiency, feedback responsiveness, observation fidelity, visual grounding) to sharpen adaptiv

Load-bearing premise

The load-bearing premise is that relabeling an already-observed visual transition with a new tool schema—keeping the observation intact—teaches a capability that transfers to genuinely novel tools when those tools are actually executed at inference; if that transfer fails, the unseen-tool results reduce to schema-fitting rather than adaptation.

What would settle it

Run VC-Tooler on a held-out set of reinstantiated schemas never seen in training, and execute each call against a real tool environment with actual image transformations. If the schema-level success rate drops to chance or the model fails to call the tool at all when the observation is genuinely new, the adaptation claim collapses. The paper's S10 already shows a partial warning sign: E2 (wrong argument values) is unmoved after RL (682→717), meaning the reward's Visual Grounding dimension is not yet fixing misgrounded calls.

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

If this is right

  • If visual tool use is compositional and adaptive, agentic multimodal models can be trained for arbitrary tool pools from synthesized trajectories, reducing the need for executable environments for every tool.
  • The two-stage recipe (SFT plus tool-reward RL) transfers to unseen tools, suggesting that schema-conditioned tool invocation can be learned without per-tool training data.
  • The approach improves general-purpose visual reasoning benchmarks, not just tool-use benchmarks, implying that tool-use training is a viable route to stronger perception.
  • The RL tool reward, rather than simply encouraging tool calls, is what drives improvement; rewarding mere invocation can hurt performance.
  • The model learns task-dependent tool-use strategies—mostly zoom on simple tasks, longer chains on agentic tasks—rather than a single fixed interaction pattern.

Where Pith is reading between the lines

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

  • The reinstantiation assumption deserves scrutiny: because mock-schema trajectories never produce real environment feedback during training, the transfer to unseen tools may reflect schema-fitting to observed observations rather than true adaptation; a test that executes those schemas against real transformations would distinguish.
  • The paper's own error analysis (S10) shows E2 (wrong argument values) and E5 (wrong reasoning from valid outputs) barely move after RL; if visual grounding of arguments remains the bottleneck, the next improvement likely comes from richer grounding supervision rather than more RL shaping.
  • The 'relabel the transition, reuse the observation' trick may extend to other agentic domains, such as text-based tools or code execution, where interface diversity can be generated cheaply without new executors.
  • Since the tool reward is a learned judge, its reliability limits the RL signal; if the judge cannot detect visually misgrounded arguments (E2), the policy may optimize for plausible-looking calls instead of correct ones.
Share X Bluesky LinkedIn Reddit HN

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 / 5 minor

Summary. The paper presents VC-Tooler, a two-stage training pipeline (cold-start SFT followed by GRPO) that aims to teach a Qwen3-VL-8B backbone compositional and adaptive visual tool use. Training data is produced by a hierarchical trajectory synthesis pipeline: a plan-then-execute stage generates single- and multi-tool trajectories, and a reinstantiation stage relabels existing visual transitions with 1,000+ novel tool schemas. The paper claims state-of-the-art open-source results on general-purpose and agentic benchmarks (95.8% on V*, 35.3% on VTC-Bench), and zero-shot transfer to 11 unseen tools. The main evidence consists of Table 1, the ablations in Tables 2–3, the tool-call statistics in Section S9, and the error analysis in Section S10.

Significance. If the central claims hold, the hierarchical synthesis plus two-stage training recipe is a practical and reproducible contribution to training VLMs for tool-rich environments. The DirectSFT control in Table 2 is a clean attribution device: it shows the gains come from tool supervision rather than from SFT on the same data alone, and the ST/MT/DTC and reward-design ablations are well posed. However, the paper's distinctive claim—adaptation to genuinely novel tool interfaces—is currently supported only by schema-level success rates and by non-degradation under a mixed tool pool, not by evidence that the model correctly interprets real novel-tool outputs. The reward analysis also shows no improvement in the semantic grounding dimensions. These gaps are fixable with additional experiments, so I view this as a major revision rather than a rejection.

major comments (3)
  1. [§4.1, Eq. (4); Table S6; Fig. 5] The central 'adaptive visual tool use' claim is not established by the current evidence. In reinstantiation, the post-action observation o_k is kept intact from the original execution, and the 1,000+ mock interfaces are never executed during training. The model therefore learns to associate a new schema string with a pre-supplied observation. At inference, unseen tools are actually executed and return real observations, yet the only reported metric for them (Table S6) is schema-level execution success—how often a call is schema-valid and executes without error—not whether the model correctly interprets the real tool output and uses it to improve the final answer. Fig. 5's Mixed configuration measures overall accuracy, but this only shows that adding unseen tools does not hurt; it does not show that novel outputs are consumed. An answer-level or observation-utilization evaluation on tasks
  2. [§S10; §S5, Prompt 4] The tool reward's five binary dimensions are claimed to shape accurate tool use, with Visual Grounding targeting E2 (wrong argument values) and Observation Fidelity targeting E5 (incorrect reasoning from a valid tool output). Yet the paper's own error analysis shows E2 682→717 and E5 770→811 after RL, i.e., no improvement on exactly the dimensions that determine semantic correctness. The claim that RL encourages 'accurate … visual tool use' is therefore overstated. More importantly, this pattern raises the question of whether the Qwen3.5-Plus critic provides a reliable training signal for the failure modes that matter most for adaptation. Please report the reward model's agreement with human labels on these dimensions, or otherwise show that the Visual Grounding and Observation Fidelity scores correlate with downstream accuracy, before attributing the RL gains to grounded adaptation.
  3. [§5.3, Table 1] The 'state-of-the-art among open-source models' claim rests on comparisons in which each baseline is evaluated under its native tool configuration, while VC-Tooler uses a tool pool that includes code execution and an online image-edit model (Qwen-Image). This conflates model capability with the tool environment: a method with access to stronger tools may improve even if its tool-use policy is no better. The ablations in Table 2 are a good start, but a matched tool-pool comparison (same tools, same tool-call budget, same evaluation harness for all methods) is needed to support the SOTA claim as stated. If this is infeasible, the claim should be qualified as 'SOTA among open-source models under each method's native tool environment.'
minor comments (5)
  1. [Table 1] The column header is misaligned ('HRBench HRBench CharXiv MME-RW VTC-Bench TIR-Bench4K 8K DQ RQ'), making it difficult to map numbers to benchmarks. Please reformat.
  2. [§5.2 and §S2] The main text lists the seen training tools as zoom, rotate, enhance, code, mark, edit, while §S2 and the planning prompt include multimodal_search as a base training tool. Please clarify whether search is in the training pool.
  3. [§S6] The statement 'We will release the full dataset composition, source breakdown, and filtering scripts upon acceptance' should be replaced with a concrete data-availability statement in the final version.
  4. [Fig. 7] The figure lacks a legend explaining the hatched segments and the benchmark-to-bar mapping. The caption says 'per benchmark' but the x-axis is not labeled.
  5. [§S9, Table S6] The success rate for the 'others' tool is reported (97.1%) but the tool is not described in Table S4. Please define what 'others' covers.

Circularity Check

0 steps flagged

No significant circularity; the central claims rest on external benchmarks and the flagged passages are metric-validity limitations rather than definitional reductions.

full rationale

The claimed derivation chain is not circular. The central results in Table 1 are measured on external benchmarks (V*, HRBench, CharXiv, MME-RealWorld, VTC-Bench, TIR-Bench) with fixed evaluation protocols, so the SOTA claim does not reduce to the paper's own training objectives. The diverse-context reinstantiation step (§4.1, Eq. 4) does keep the observation o_k from the original execution and therefore limits what the mock-interface trajectories can teach about real novel-tool outputs; and Table S6's 'success' is explicitly schema-level execution success, not correctness of interpreting returned observations. These are legitimate empirical limitations of the adaptation claim, but they are not definitional: the 11 unseen test schemas are not part of the training pool, and schema-valid calls are not guaranteed by construction. Similarly, §S10's finding that RL reduces E4 ('tool result not used') is aligned with the tool reward's design (Prompt 4), so it is a mechanism check rather than independent evidence; the paper's main evidence for the reward's value is the external benchmark gain in Table 3. The single self-citation (Pan et al. 2026) is in related work and is not load-bearing. Sec. 7 acknowledges bounded tool spaces and short interaction horizons. Overall, the suspicious passages concern metric validity and generalization, not circular reduction.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 0 invented entities

The paper's central claim rests on a chain of curation and modeling choices rather than mathematical derivation. The dominant free inputs are: hand-chosen reward weights and criteria, aggressive relevance/difficulty filtering thresholds (70–90% removal), and a judge stack (Qwen3-VL-235B planner/executor, Qwen3.5-Plus reward critic, Qwen-Max correctness judge) that both sources the training supervision and evaluates tool-use quality, all from the same model family as the base model. No new theoretical entities are postulated: the 1,000+ mock tool schemas and 11 unseen tools are software interfaces, not physical postulates. The most consequential unproven assumption is that schema relabeling with inherited observations (Eq. 4) transfers to real execution of unseen tools at inference.

free parameters (4)
  • Reward weights in Eq. 5 = 0.8 accuracy / 0.2 tool / 0.2 format
    Hand-chosen weighting of the total reward; no sensitivity analysis is reported, and the weights determine how much the judge-based tool reward shapes the final policy.
  • Tool-reward criteria set = 5 binary dimensions averaged uniformly
    The five criteria and their pass/fail definitions (plan coherence, call efficiency, feedback responsiveness, observation fidelity, visual grounding) are hand-designed in §S5 and target the paper's own E1–E5 error taxonomy; the uniform averaging is another unexamined choice.
  • Difficulty and relevance filtering thresholds = solved ≥4/8 attempts → discard; 8 attempts; 70–90% raw-candidate removal
    Arbitrary cutoffs in §S1 define which samples enter the trajectory bank; since the bank is the training signal, these thresholds shape capability at every later stage.
  • RL data filtering = remove samples the base policy fails in 8 attempts; ~28K retained
    §S6 selects a 'moderate-difficulty' RL set, so the measured RL gains apply to a curated subset; the filter's effect on benchmark transfer is not analyzed.
axioms (6)
  • domain assumption ReAct-style interleaved reasoning-and-acting is the correct operationalization of visual tool use (§3).
    The whole framework (Eqs. 1–3) presupposes this paradigm; alternative formulations (programmatic tools only, no language reasoning) are not considered.
  • domain assumption Large-Qwen trajectories and judgments are valid supervision for small-Qwen tool use.
    Planner/executor (Qwen3-VL-235B variants) generate the SFT bank; Qwen3.5-Plus judges the tool reward; Qwen-Max judges correctness — all share the base model family.
  • domain assumption Answer-conditioned planning does not leak the answer into execution (§4.1).
    The planner sees y-hat and the final bank is filtered for answer consistency; the paper claims separation 'avoids answer leakage,' but executor plans are still selected toward ground truth.
  • ad hoc to paper Schema relabeling with inherited observations preserves the learning signal for adaptation (§4.1, Eq. 4).
    Mock tools keep o_k from the original execution; the transfer from this signal to real execution of unseen tools is the paper's core unverified premise.
  • domain assumption LLM-as-judge correctness checks (Qwen-Max) are reliable enough to filter trajectories and score RL answers.
    Used in §S1 (trajectory filtering), §S5 (tool reward), §S6 (RL data correctness); no agreement study against human labels is reported.
  • domain assumption Benchmark scores (V*, HRBench, CharXiv, MME-RW, VTC-Bench, TIR-Bench) measure the claimed capabilities and are not contaminated by the training data sources.
    VTC-Bench and TIR-Bench are recent third-party benchmarks, but training sources in Table S1/§S6 (e.g., ChartVerse, DRIM, VisualProbe) overlap thematically with the evaluation suite; no contamination analysis is given.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of VC-Tooler: Learning Compositional and Adaptive Visual Tool Use." pith.science (2026). https://pith.science/paper/K6XZZLPB

@misc{pith2026260802217,
  author       = {Pith},
  title        = {Pith review of: VC-Tooler: Learning Compositional and Adaptive Visual Tool Use},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/K6XZZLPB}},
  note         = {Machine review of arXiv:2608.02217}
}
Share X Bluesky LinkedIn Reddit HN
abstract

Agentic multimodal reasoning extends passive image understanding by allowing VLMs to actively acquire and refine visual evidence through visual tool interactions. Effective visual tool use requires three capabilities: grounding tool calls in visual context, composing tools across multiple steps, and adapting reasoning to tool-returned observations. However, existing approaches largely focus on grounding within fixed tool spaces and rigid invocation patterns, leaving composition and adaptation insufficiently addressed. We present VC-Tooler, which learns visual tool use as a compositional and adaptive capability. To this end, we first build a trajectory bank through a hierarchical synthesis pipeline covering three capability levels: single-tool grounding, multi-tool composition, and diverse tool contexts and interfaces. We then train the model in two stages: a supervised cold start that establishes these capabilities, followed by reinforcement learning that encourages accurate, efficient, and context-aware visual tool use. VC-Tooler achieves state-of-the-art performance among open-source models on both general-purpose and agentic benchmarks, including $95.8\%$ on V* and $35.3\%$ on VTC-Bench, and shows promising transfer under richer tool settings at inference time. Project page: https://w1zheng.github.io/VC-Tooler

Figures

Figures reproduced from arXiv: 2608.02217 by Bing Deng, Jiashen Hua, Jieping Ye, Yizheng Wu.

Figure 1
Figure 1. Figure 1: Existing methods often ground a single familiar tool [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of VC-Tooler. A two-stage framework with hierarchical trajectory supervision for cold start and tool-reward RL for policy refinement [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Trajectory synthesis via plan-then-execute and diverse tool-context reinstantiating. [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Composition and diversity of the synthesized SFT trajectory bank. (a) Breakdown of the curated trajectories into single-tool, multi-tool, and diverse-tool-context data, together with their sample and pattern distributions. (b) Word cloud of tool and operation patterns covered in the trajectory bank. into execution. Second, executing tools under a planned se￾quence provides a controlled interaction context,… view at source ↗
Figure 5
Figure 5. Figure 5: Generalization across tool availability. Per￾benchmark accuracy of all methods under four tool con￾figurations (Original, No tools, Seen, Mixed). Tool pool in evaluation. We evaluate VC-Tooler with a di￾verse tool pool consisting of seen tools and unseen tools. The seen tools are the six tools available during train￾ing (zoom, rotate, enhance, code, mark, edit), whereas the unseen tools are eleven addition… view at source ↗
Figure 6
Figure 6. Figure 6: Case studies of compositional and adaptive visual tool use of VC-Tooler. VC-Tooler performs strongly across both general￾purpose and agentic benchmarks. On general-purpose benchmarks, VC-Tooler-RL attains 95.8 on V*, 83.8 on HRBench-8K, and 69.5 on MME-RealWorld, outperform￾ing prior open-source agentic baselines and remaining com￾petitive with strong proprietary models. It also improves over prior open-so… view at source ↗
Figure 7
Figure 7. Figure 7: Tool-call distribution before and after RL. Each bar shows the proportion of invoked tools, with solid colors denoting tools seen during training and hatched segments denoting novel tools. The number on top of each bar indicates the average number of tool calls per benchmark. 5.4 Ablation Study Effect of SFT Data Composition To understand how dif￾ferent types of supervised data contribute to VC-Tooler, we … 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

52 extracted references · 24 linked inside Pith

  1. [1]

    frontal lobe

    The unity and diversity of executive functions and their contributions to complex “frontal lobe” tasks: A latent variable analysis , author=. Cognitive psychology , volume=. 2000 , publisher=

    2000

  2. [2]

    arXiv preprint arXiv:2506.23918 , year=

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

    Pith/arXiv arXiv

  3. [3]

    arXiv preprint arXiv:2601.12538 , year=

    Agentic reasoning for large language models , author=. arXiv preprint arXiv:2601.12538 , year=

    Pith/arXiv 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=

    Pith/arXiv arXiv

  5. [5]

    thinking with images

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

    Pith/arXiv arXiv

  6. [6]

    arXiv preprint arXiv:2512.17306 , year=

    Deep But Reliable: Advancing Multi-turn Reasoning for Thinking with Images , author=. arXiv preprint arXiv:2512.17306 , year=

    arXiv

  7. [7]

    arXiv preprint arXiv:2508.11630 , year=

    Thyme: Think Beyond Images , author=. arXiv preprint arXiv:2508.11630 , year=

    Pith/arXiv arXiv

  8. [8]

    arXiv preprint arXiv:2512.02361 , year=

    VACoT: Rethinking Visual Data Augmentation with VLMs , author=. arXiv preprint arXiv:2512.02361 , year=

    arXiv

  9. [9]

    arXiv preprint arXiv:2511.05271 , year=

    Deepeyesv2: Toward agentic multimodal model , author=. arXiv preprint arXiv:2511.05271 , year=

    Pith/arXiv arXiv

  10. [10]

    arXiv preprint arXiv:2508.05748 , year=

    Webwatcher: Breaking new frontier of vision-language deep research agent , author=. arXiv preprint arXiv:2508.05748 , year=

    Pith/arXiv arXiv

  11. [11]

    arXiv preprint arXiv:2505.22525 , year=

    Thinking with generated images , author=. arXiv preprint arXiv:2505.22525 , year=

    Pith/arXiv arXiv

  12. [12]

    arXiv preprint arXiv:2507.16746 , year=

    Zebra-cot: A dataset for interleaved vision language reasoning , author=. arXiv preprint arXiv:2507.16746 , year=

    arXiv

  13. [13]

    arXiv preprint arXiv:2512.16918 , year=

    AdaTooler-V: Adaptive Tool-Use for Images and Videos , author=. arXiv preprint arXiv:2512.16918 , year=

    Pith/arXiv arXiv

  14. [14]

    arXiv preprint arXiv:2601.18631 , year=

    AdaReasoner: Dynamic Tool Orchestration for Iterative Visual Reasoning , author=. arXiv preprint arXiv:2601.18631 , year=

    arXiv

  15. [15]

    arXiv preprint arXiv:2603.15030 , year=

    VTC-Bench: Evaluating Agentic Multimodal Models via Compositional Visual Tool Chaining , author=. arXiv preprint arXiv:2603.15030 , year=

    arXiv

  16. [16]

    arXiv preprint arXiv:2511.01833 , year=

    TIR-Bench: A Comprehensive Benchmark for Agentic Thinking-with-Images Reasoning , author=. arXiv preprint arXiv:2511.01833 , year=

    arXiv

  17. [17]

    V*: Guided visual search as a core mechanism in multimodal llms , author=

  18. [18]

    arXiv preprint arXiv:2507.07998 , year=

    Pyvision: Agentic vision with dynamic tooling , author=. arXiv preprint arXiv:2507.07998 , year=

    Pith/arXiv arXiv

  19. [19]

    arXiv preprint arXiv:2512.18745 , year=

    InSight-o3: Empowering Multimodal Foundation Models with Generalized Visual Search , author=. arXiv preprint arXiv:2512.18745 , year=

    arXiv

  20. [20]

    Farrar, Straus and Giroux , year=

    Thinking, fast and slow , author=. Farrar, Straus and Giroux , year=

  21. [21]

    arXiv preprint arXiv:2507.05255 , year=

    Open vision reasoner: Transferring linguistic cognitive behavior for visual reasoning , author=. arXiv preprint arXiv:2507.05255 , year=

    arXiv

  22. [22]

    Llava-cot: Let vision language models reason step-by-step , author=

  23. [23]

    R1-onevision: Advancing generalized multimodal reasoning through cross-modal formalization , author=

  24. [24]

    Through the Lens of Contrast: Self-Improving Visual Reasoning in VLMs , author=

  25. [25]

    arXiv preprint arXiv:2412.16720 , year=

    Openai o1 system card , author=. arXiv preprint arXiv:2412.16720 , year=

    Pith/arXiv arXiv

  26. [26]

    arXiv preprint arXiv:2501.12948 , year=

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning , author=. arXiv preprint arXiv:2501.12948 , year=

    Pith/arXiv arXiv

  27. [27]

    Apigen: Automated pipeline for generating verifiable and diverse function-calling datasets , author=

  28. [28]

    arXiv preprint arXiv:2511.15718 , year=

    ToolMind Technical Report: A Large-Scale, Reasoning-Enhanced Tool-Use Dataset , author=. arXiv preprint arXiv:2511.15718 , year=

    arXiv

  29. [29]

    arXiv preprint arXiv:2504.11536 , year=

    Retool: Reinforcement learning for strategic tool use in llms , author=. arXiv preprint arXiv:2504.11536 , year=

    Pith/arXiv arXiv

  30. [30]

    arXiv preprint arXiv:2602.15763 , year=

    Glm-5: from vibe coding to agentic engineering , author=. arXiv preprint arXiv:2602.15763 , year=

    Pith/arXiv arXiv

  31. [31]

    Introducing Agentic Vision in Gemini 3 Flash , year =

  32. [32]

    arXiv preprint arXiv:2210.03629 , year=

    React: Synergizing reasoning and acting in language models , author=. arXiv preprint arXiv:2210.03629 , year=

    Pith/arXiv arXiv

  33. [33]

    arXiv preprint arXiv:2505.09388 , year=

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

    Pith/arXiv arXiv

  34. [34]

    arXiv preprint arXiv:2509.23661 , year=

    Llava-onevision-1.5: Fully open framework for democratized multimodal training , author=. arXiv preprint arXiv:2509.23661 , year=

    Pith/arXiv arXiv

  35. [35]

    arXiv preprint arXiv:2408.03326 , year=

    Llava-onevision: Easy visual task transfer , author=. arXiv preprint arXiv:2408.03326 , year=

    Pith/arXiv arXiv

  36. [36]

    arXiv preprint arXiv:2402.03300 , year=

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

    Pith/arXiv arXiv

  37. [37]

    Swift: a scalable lightweight infrastructure for fine-tuning , author=

  38. [38]

    Proceedings of the Twentieth European Conference on Computer Systems , pages=

    Hybridflow: A flexible and efficient rlhf framework , author=. Proceedings of the Twentieth European Conference on Computer Systems , pages=

  39. [39]

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

  40. [40]

    Charxiv: Charting gaps in realistic chart understanding in multimodal llms , author=

  41. [41]

    Mme-realworld: Could your multimodal llm challenge high-resolution real-world scenarios that are difficult for humans? , author=

  42. [42]

    Thinking with images , year =

  43. [43]

    , author=

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

    arXiv

  44. [44]

    arXiv preprint arXiv:2512.17312 , year=

    CodeDance: A Dynamic Tool-integrated MLLM for Executable Visual Reasoning , author=. arXiv preprint arXiv:2512.17312 , year=

    arXiv

  45. [45]

    arXiv preprint arXiv:2601.13606 , year=

    ChartVerse: Scaling Chart Reasoning via Reliable Programmatic Synthesis from Scratch , author=. arXiv preprint arXiv:2601.13606 , year=

    Pith/arXiv arXiv

  46. [46]

    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=

    Pith/arXiv arXiv

  47. [47]

    arXiv preprint arXiv:2504.10479 , year=

    Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models , author=. arXiv preprint arXiv:2504.10479 , year=

    Pith/arXiv arXiv

  48. [48]

    arXiv preprint arXiv:1711.05101 , year=

    Decoupled weight decay regularization , author=. arXiv preprint arXiv:1711.05101 , year=

    Pith/arXiv arXiv

  49. [49]

    arXiv preprint arXiv:2602.16742 , year=

    DeepVision-103K: A Visually Diverse, Broad-Coverage, and Verifiable Mathematical Dataset for Multimodal Reasoning , author=. arXiv preprint arXiv:2602.16742 , year=

    arXiv

  50. [50]

    arXiv preprint arXiv:2511.21395 , year=

    Monet: Reasoning in latent visual space beyond images and language , author=. arXiv preprint arXiv:2511.21395 , year=

    arXiv

  51. [51]

    arXiv preprint arXiv:2511.19773 , year=

    Scaling Agentic Reinforcement Learning for Tool-Integrated Reasoning in VLMs , author=. arXiv preprint arXiv:2511.19773 , year=

    arXiv

  52. [52]

    arXiv preprint arXiv:2508.02324 , year=

    Qwen-image technical report , author=. arXiv preprint arXiv:2508.02324 , year=

    Pith/arXiv arXiv

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.