Pith. sign in

REVIEW 3 major objections 6 minor 114 references

ToolAnchor claims that verified counterfactual contexts at critical decision points let post-trained agents adopt new tools while preserving old skills.

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 · deepseek-v4-flash

2026-08-02 06:32 UTC pith:W75EGZGZ

load-bearing objection ToolAnchor has a genuinely new mechanism and a coherent story, but the headline numbers rest on an unvalidated LLM judge, so treat them as provisional until the evaluation is tightened. the 3 major comments →

arxiv 2607.14145 v1 pith:W75EGZGZ submitted 2026-07-14 cs.AI cs.CVcs.LG

ToolAnchor: Anchoring Counterfactual Context to Boost Agentic Tool-use Capability

Weiting Liu , Jieyi Bi , Wanqi Zhou , Jianfeng Feng , Yining Ma , Ai Han , Wenlian Lu This is my paper
classification cs.AI cs.CVcs.LG
keywords toolset expansionbehavioral inertiacounterfactual anchor contextagentic post-trainingagentic reinforcement learningtool-use agentsvisual searchLLM-as-a-judge
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.

This paper tackles a practical problem: when an AI agent has already been trained to use one set of tools, simply adding new tools rarely works, and retraining from scratch is too expensive. The authors argue the obstacle is behavioral inertia—the agent keeps falling back on familiar tools and reasoning patterns even when better options are available. They show that a teacher can propose a counterfactual "anchor" at a single critical decision point, replacing the agent's local thought and action with a better alternative while keeping the earlier history unchanged, and that the student often completes the task after resuming from that revised context. ToolAnchor scales this by having multiple teachers hypothesize anchors, verifying each one through student rollouts with both task-success and anchor-usefulness checks, then internalizing only verified anchors via supervised fine-tuning followed by reinforcement learning. On general-assistant, textual-search, and visual-search benchmarks, the method improves new-tool use while preserving or slightly improving prior task performance.

Core claim

The central claim is that failures in toolset expansion are often local and redirectable: within a failed trajectory there is a decision round where a different thought–action pair, under identical prior history, would reroute the rollout toward success. The paper calls that revised context a counterfactual anchor context and shows that resuming the student from teacher-hypothesized anchors recovers failed tasks more often than continuing from the original context. ToolAnchor operationalizes this by collecting failed trajectories, having several teacher models independently hypothesize anchor rounds and alternative decisions using only pre-anchor history, verifying each candidate by letting

What carries the argument

The counterfactual anchor context: a revised thought–action pair plus the resulting observation at one decision round, inserted in place of the original local decision under unchanged preceding history. It acts as a contextual reset point from which the student resumes generation. The second load-bearing mechanism is dual-level verification: a task-level correctness judge and an anchor-level usefulness judge filter teacher hypotheses, so the training data stays close to the student's own rollout distribution rather than imitating full teacher trajectories.

Load-bearing premise

The entire pipeline—which anchors are kept and how success is scored—rests on LLM-as-a-judge labels for both task correctness and anchor usefulness, with no reported human-agreement or sensitivity analysis; if those labels are biased or noisy, the verified set and the reported gains are not trustworthy.

What would settle it

A concrete check: take a sample of verified and rejected anchors, have human annotators label task success and anchor usefulness, and measure agreement with the LLM judge; if agreement is near chance, or if training on human-verified anchors fails to reproduce the reported visual-search gain, the central claim is undermined. Alternatively, rerun the pipeline with the anchor-level judge removed and compare verified-SFT versus unverified-SFT performance: if the gap disappears, verification is not the operative mechanism.

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

If this is right

  • Agents post-trained on fixed toolsets can be extended with new tools through targeted local supervision instead of full retraining, substantially reducing compute and data requirements.
  • Teacher-generated guidance can be treated as hypotheses rather than demonstrations; student-side verification filters noisy or unhelpful supervision, which matters for scaling data collection.
  • Internalizing verified anchors can improve not only new-tool adoption but also existing tool use and reasoning, as suggested by gains on general-assistant and textual-search benchmarks.
  • The anchor-round selection strategy matters: criterion-based selection outperforms first-error, last-error, and random interventions, implying that precise diagnosis is as important as the proposed fix.
  • Because anchor hypothesis is not restricted to new-tool rounds, the framework may generalize to suboptimal use of existing tools and flawed reasoning steps, not just toolset expansion.

Where Pith is reading between the lines

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

  • The same verified-anchor loop could be run online as tools are added over time, pointing toward a continual agentic-learning setting where the tool catalog grows without repeated full RL runs; the paper does not explore this dynamic case.
  • A testable extension is to replace or augment the LLM judge with execution-based or human-verified labels; if the performance gap between verified and unverified SFT persists, the verification mechanism is robust, and if it disappears, the judge is the real driver.
  • If the causal story is right, anchors that pass verification should cluster at identifiable causal pivots, and post-training should measurably change the student's behavior at those types of rounds; the paper reports anchor distributions but not per-round behavioral shifts.
  • The mechanism is not modality-specific, so similar gains might appear when adding non-visual tools such as calculators or database interfaces, but that remains an open empirical question.

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 / 6 minor

Summary. The paper addresses the toolset-expansion problem for post-trained tool-augmented LLM agents. It argues that such agents exhibit behavioral inertia, and that failed trajectories can be recovered by injecting a counterfactual anchor context at a critical round. ToolAnchor automates this in three stages: (i) teacher models hypothesize an anchor round and an alternative thought--action pair from the history preceding it; (ii) the student policy rolls out from the candidate context, and the context is retained only if a task-level judge and an anchor-level judge both approve (Eq. 1); (iii) the retained contexts are used for anchor-only agentic SFT followed by GRPO. On GAIA, BrowseComp, and VDR-Bench, the resulting agent reports 74.5%, 45.0%, and 28.8%, improving over the base model and most non-VDR-native baselines, with ablations suggesting that verification and SFT are both needed.

Significance. If the results are reproducible, the paper would make a useful empirical contribution: an efficient way to extend a text-only deep-research agent with visual tools without retraining from scratch, with the teacher treated as a hypothesis generator rather than a demonstration oracle. The open base model and full hyperparameters (Tables 5--7) facilitate independent verification. The main risk is that every stage -- training-data filtering and final evaluation -- rests on a single unvalidated LLM judge; until judge reliability and baseline comparability are established, the exact magnitudes (especially the 7.5% to 28.8% VDR-Bench gain) should be treated as provisional. With those gaps filled, the paper would be a solid empirical contribution.

major comments (3)
  1. [Sec. 4.3, Eq. (1); Sec. 5.1; Appendices B.3/B.4] The entire verification and evaluation chain relies on a single LLM judge, Qwen3-235B-A22B, with no validation against human judgments, a second judge, or known ground-truth labels. Eq. (1) retains a counterfactual context only if Jtask(H'_T)=1 and Janchor(c*,H'_T)=1, both produced by this judge, and Table 3 and Figure 3 are scored by the same judge using the prompts in Appendix B.4. The anchor-level prompt in B.3 is a new, custom task with no calibration data. If the judge is systematically lenient or prefers certain answer styles, the verified-vs-unverified contrast in Figure 3 and the headline VDR-Bench jump could reflect judge bias rather than counterfactual utility. The limitation section (Sec. 6) acknowledges limited teacher diversity and data, but not judge reliability. I request a human-annotated sample (e.g., 100 or more trajectories per benchmark) with inter-annotator agreement
  2. [Table 3; Sec. 5.2] The main-table comparisons mix the authors' runs with numbers taken from other papers (marked *), and the benchmarks are not all evaluated under identical conditions. In particular, BrowseComp is described as a 200-sample subset following [34], but starred baselines such as Tongyi-DeepResearch (43.4%) and OpenSeeker (29.5%) are likely quoted from their original reports and may use different subsets or judge settings. The claimed improvements over the base model are small on GAIA (+3.6pp) and BrowseComp (+1.6pp), while inference is stochastic (temperature 0.85, Table 7) and no seeds, error bars, or significance tests are reported. I request that all baselines be rerun under the same judge, the same subset, and the same inference protocol, with at least 3 seeds or bootstrap confidence intervals. Without this, the 'competitive performance' claim is not quantitatively supported.
  3. [Figure 3; Sec. 5.3] The ablation comparing 'Agentic RL from scratch', 'Anchor-only agentic SFT' (verified and unverified), and ToolAnchor does not specify whether the RL-from-scratch and SFT-only conditions are matched in compute, rollout budget, and optimization steps with the corresponding stages of ToolAnchor. If RL-from-scratch is given fewer rollouts or less training than the RL stage used inside ToolAnchor, the conclusion that 'RL alone exhibits limited exploration' may be an artifact of unequal budgets. Please state the exact rollout count, batch size, and number of optimization steps for each ablation arm, and, if not already matched, rerun with matched budgets.
minor comments (6)
  1. [Table 1] Tool-call frequencies are averaged over 100 instances with no variance or significance measures. The 'stark contrast' is illustrative; adding per-model total turn counts or a simple statistical test would strengthen the inertia interpretation.
  2. [Eq. (1) and Sec. 4.3] The notation H'_T is ambiguous when Pass@3 verification is used: if three student rollouts are sampled, clarify whether all successful rollouts enter Deff or only one, and how Janchor is applied to rollouts that were not sampled for the successful pass.
  3. [Figure 2] The three-stage diagram is dense and the text is small; the flow from 'hypothesis' to 'verification' to 'internalization' would be easier to follow with numbered stage labels matching Sec. 4.2--4.4.
  4. [Table 3] The star notation is applied inconsistently (e.g., Gemini-2.5-Pro GAIA and BrowseComp are unstarred while its VDR-Bench is starred). Please list explicitly which numbers are the authors' runs and which are quoted, with citations and subset sizes.
  5. [Sec. 5.4 and Figure 4] The anchor-selection study uses Gemini-3.1-Pro-Preview, while the main experiments use Gemini-2.5-Pro and GPT-5. State whether this is a separate, earlier study and whether the conclusion generalizes to the teachers used in the main pipeline.
  6. [Appendix C.2] The RL-from-scratch ablation in Figure 3 is not described in the hyperparameter tables; please add an explicit row or paragraph for that configuration.

Circularity Check

0 steps flagged

No significant circularity: ToolAnchor is an empirical pipeline evaluated against external benchmarks; the shared LLM-judge issue is a validation risk, not a derivation-equals-input loop.

full rationale

I walked the claimed derivation chain: failed trajectories are collected from student rollouts (Sec. 4.2), teacher models hypothesize anchor rounds and alternative thought-action pairs using only history before the anchor round, student rollouts verify candidate contexts, and Eq. (1) defines Deff via task-level success (Jtask) and anchor-level usefulness (Janchor). Eq. (2) then trains only on the verified anchor pair. None of these steps defines the final benchmark numbers in Table 3 in terms of the method's own inputs: the teacher pair is generated under the student's information constraint, the verification rollout is executed by the student policy, and the final metrics are measured on external benchmarks (GAIA, BrowseComp, VDR-Bench) with benchmark ground-truth answers supplied to the judge. There is no fitted parameter that is later renamed a prediction, and no uniqueness theorem or ansatz is imported from the authors' prior work to force the conclusion. The closest concern is methodological: Qwen3-235B-A22B is used both as the anchor-level judge in Eq. (1) and as the final benchmark judge in Sec. 5.1, with no human-agreement or sensitivity analysis. That is a real validation/reward-validity risk, but it is not circularity: the judge is a fixed external model, the tasks are externally defined, and the reported gains are empirical measurements rather than quantities derived by construction from the filtering criterion. If the judge is biased, the numbers may be inflated, but that would be an artifact of the evaluation, not a circular derivation. The paper is self-contained against external benchmarks, so the appropriate circularity score is 0.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 2 invented entities

The method has no fitted physical constants; the free parameters are hand-set training choices and verification thresholds. Core assumptions are the validity of LLM judges, teacher anchor quality, comparability of published baselines, and the behavioral-inertia mechanism.

free parameters (5)
  • Pass@3 verification budget = 3 rollouts
    Hand-set in Section 4.3; a candidate context is retained if any of 3 student rollouts solves the task. This threshold controls the size and quality of Deff and is not swept.
  • Agentic RL learning rate = 1e-6
    Table 5; no sweep reported. Affects whether post-SFT RL preserves old vs new behaviors.
  • Agentic SFT learning rate = 5e-7
    Table 6; hand-chosen, no sweep reported.
  • Training rollout budget = 3,000 rollouts
    Table 5; the RL-from-scratch baseline is given the same budget, but convergence is not shown.
  • Training data mixture = OpenSeeker-1.5k + VDR-1.5k
    1,500 samples per dataset chosen without scaling analysis; affects how much anchor diversity is available.
axioms (5)
  • domain assumption LLM-as-a-judge labels (Qwen3-235B) are correct for both task success and anchor usefulness.
    Section 5.1 and Appendix B.3/B.4; no human agreement, error analysis, or judge sensitivity is reported.
  • domain assumption Teacher models can generate effective anchors without leaking future information or ground truth.
    Section 4.2 separates anchor-round diagnosis from context generation, but leakage is only mitigated by prompt design, not verified.
  • domain assumption Behavioral inertia is caused by policy convergence, not by prompt/tool-description artifacts.
    Section 3.3 infers this from tool-call frequencies; alternative explanations such as discoverability of tool descriptions are not controlled.
  • ad hoc to paper Pass@3 is a suitable verification criterion for counterfactual context quality.
    Section 4.3; threshold of 3 rollouts is chosen by hand and no sensitivity analysis is given.
  • domain assumption Baseline results marked '*' are comparable to the authors' own runs.
    Table 3 mixes cited numbers with new runs on a 200-sample BrowseComp subset and VDR test-mini; equivalence of evaluation protocols is assumed.
invented entities (2)
  • Counterfactual anchor context (c*) no independent evidence
    purpose: A revised history+thought+action+observation at a critical decision round, used as a training context for agentic SFT.
    The construct is validated only by in-paper benchmark numbers and judge labels; no external falsifiable handle is provided outside this paper.
  • Behavioral inertia (as a mechanistic explanation) no independent evidence
    purpose: Explains why post-trained agents underuse newly introduced tools.
    Table 1 measures tool-call frequencies, but the mechanistic claim that policy convergence causes the inertia is not independently established outside this paper.

pith-pipeline@v1.3.0-alltime-deepseek · 27110 in / 15823 out tokens · 163362 ms · 2026-08-02T06:32:01.066515+00:00 · methodology

0 comments
read the original abstract

Tool-augmented large language model agents excel at long-horizon tasks, yet they are typically post-trained on fixed toolsets. When tasks demand new tools, these agents struggle to incorporate them effectively, and retraining from scratch is often impractical. We identify the core obstacle in such toolset expansion problem as behavioral inertia: the tendency of agents to fall back on familiar tools and established reasoning patterns despite having access to new ones. We demonstrate that injecting counterfactual anchor contexts at critical decision points can break this inertia, recovering failed trajectories by eliciting suppressed agent capabilities. To scale this insight, we propose ToolAnchor, a framework that uses teacher models to hypothesize these counterfactual contexts, verifies them via student rollouts, and internalizes the successful interventions through agentic post-training. Extensive evaluations across general AI assistant (GAIA), textual search (BrowseComp), and visual search (VDR-Bench) tasks demonstrate that ToolAnchor consistently exhibits competitive performance under expanded toolsets. Our work bridges the gap between static post-training and dynamic adaptation, charting a new path for scalable agentic reinforcement learning.

Figures

Figures reproduced from arXiv: 2607.14145 by Ai Han, Jianfeng Feng, Jieyi Bi, Wanqi Zhou, Weiting Liu, Wenlian Lu, Yining Ma.

Figure 1
Figure 1. Figure 1: A case study of counterfactual anchor context generation [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the ToolAnchor framework 4.1 Problem formulation Building on the setup in Section 3.1, we study toolset expansion for a post-trained tool-use agent πθ. The agent has been trained with the original toolset Uold and is then exposed to the expanded toolset U = Uold ∪ Unew. The goal is to learn an updated policy πθ ′ that improves task success under U while preserving the agent’s existing capabilit… view at source ↗
Figure 3
Figure 3. Figure 3: Effects of post-training components 0 5 10 15 20 25 30 Anchor-conditioned rollout success rate (%) Last Error Round (Teacher Context) First Error Round (Teacher Context) Random Round (Teacher Context) Criteria Round (Student Context) Criteria Round (Teacher Context) Anchor selection strategy 8.0 16.0 12.0 10.5 20.5 9.5 21.0 17.0 15.5 23.5 VDR OpenSeeker +4.5 +2.5 [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Distribution of anchor rounds over 50 repeated teacher hypotheses for a failed trajectory. [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Frequency Statistics of Counterfactual Anchor Rounds in Gemini-2.5-Pro [PITH_FULL_IMAGE:figures/full_fig_p034_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

114 extracted references · 27 linked inside Pith

  1. [1]

    Agentic reasoning and tool integration for llms via reinforcement learning.arXiv preprint arXiv:2505.01441, 2025

    Joykirat Singh, Raghav Magazine, Yash Pandya, and Akshay Nambi. Agentic reasoning and tool integration for llms via reinforcement learning.arXiv preprint arXiv:2505.01441, 2025

    Pith/arXiv arXiv 2025

  2. [2]

    Verltool: Towards holistic agentic reinforcement learning with tool use.arXiv preprint arXiv:2509.01055, 2025

    Dongfu Jiang, Yi Lu, Zhuofeng Li, Zhiheng Lyu, Ping Nie, Haozhe Wang, Alex Su, Hui Chen, Kai Zou, Chao Du, et al. Verltool: Towards holistic agentic reinforcement learning with tool use.arXiv preprint arXiv:2509.01055, 2025

    arXiv 2025

  3. [3]

    Reinforcement learning foundations for deep research systems: A survey.arXiv preprint arXiv:2509.06733, 2025

    Wenjun Li, Zhi Chen, Jingru Lin, Hannan Cao, Wei Han, Sheng Liang, Zhi Zhang, Kuicai Dong, Dexun Li, Chen Zhang, et al. Reinforcement learning foundations for deep research systems: A survey.arXiv preprint arXiv:2509.06733, 2025

    arXiv 2025

  4. [4]

    Exploring reasoning reward model for agents

    Kaixuan Fan, Kaituo Feng, Manyuan Zhang, Tianshuo Peng, Zhixun Li, Yilei Jiang, Shuang Chen, Peng Pei, Xunliang Cai, and Xiangyu Yue. Exploring reasoning reward model for agents. arXiv preprint arXiv:2601.22154, 2026

    Pith/arXiv arXiv 2026

  5. [5]

    Reasoning and tool-use compete in agentic rl: From quantifying interference to disentangled tuning.arXiv preprint arXiv:2602.00994, 2026

    Yu Li, Mingyang Yi, Xiuyu Li, Ju Fan, Fuxin Jiang, Binbin Chen, Peng Li, Jie Song, and Tieying Zhang. Reasoning and tool-use compete in agentic rl: From quantifying interference to disentangled tuning.arXiv preprint arXiv:2602.00994, 2026

    Pith/arXiv arXiv 2026

  6. [6]

    Proximity-based multi-turn optimization: Practical credit assignment for llm agent training

    Yangyi Fang, Jiaye Lin, Xiaoliang Fu, Cong Qin, Haolin Shi, and Chang Liu. Proximity-based multi-turn optimization: Practical credit assignment for llm agent training. InProceedings of the 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026), pages 285–307, 2026

    2026

  7. [7]

    Webwalker: Benchmarking llms in web traversal

    Jialong Wu, Wenbiao Yin, Yong Jiang, Zhenglin Wang, Zekun Xi, Runnan Fang, Linhai Zhang, Yulan He, Deyu Zhou, Pengjun Xie, et al. Webwalker: Benchmarking llms in web traversal. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 10290–10305, 2025

    2025

  8. [8]

    Browsecomp: A simple yet challenging benchmark for browsing agents.arXiv preprint arXiv:2504.12516, 2025

    Jason Wei, Zhiqing Sun, Spencer Papay, Scott McKinney, Jeffrey Han, Isa Fulford, Hyung Won Chung, Alex Tachard Passos, William Fedus, and Amelia Glaese. Browsecomp: A simple yet challenging benchmark for browsing agents.arXiv preprint arXiv:2504.12516, 2025

    Pith/arXiv arXiv 2025

  9. [9]

    Browsecomp-zh: Benchmarking web browsing ability of large language models in chinese.arXiv preprint arXiv:2504.19314, 2025

    Peilin Zhou, Bruce Leon, Xiang Ying, Can Zhang, Yifan Shao, Qichen Ye, Dading Chong, Zhiling Jin, Chenxuan Xie, Meng Cao, et al. Browsecomp-zh: Benchmarking web browsing ability of large language models in chinese.arXiv preprint arXiv:2504.19314, 2025

    Pith/arXiv arXiv 2025

  10. [10]

    Gaia: a benchmark for general ai assistants

    Grégoire Mialon, Clémentine Fourrier, Thomas Wolf, Yann LeCun, and Thomas Scialom. Gaia: a benchmark for general ai assistants. InThe Twelfth International Conference on Learning Representations, 2023

    2023

  11. [11]

    Magic-vqa: Multimodal and grounded inference with commonsense knowledge for visual question answering

    Shuo Yang, Caren Han, Siwen Luo, and Eduard Hovy. Magic-vqa: Multimodal and grounded inference with commonsense knowledge for visual question answering. InFindings of the Association for Computational Linguistics: ACL 2025, pages 16967–16986, 2025

    2025

  12. [12]

    Vision-deepresearch: Incentivizing deepresearch capability in multimodal large language models.arXiv preprint arXiv:2601.22060, 2026

    Wenxuan Huang, Yu Zeng, Qiuchen Wang, Zhen Fang, Shaosheng Cao, Zheng Chu, Qingyu Yin, Shuang Chen, Zhenfei Yin, Lin Chen, et al. Vision-deepresearch: Incentivizing deepresearch capability in multimodal large language models.arXiv preprint arXiv:2601.22060, 2026

    arXiv 2026

  13. [13]

    Visbrowse-bench: Benchmarking visual-native search for multimodal browsing agents.arXiv preprint arXiv:2603.16289, 2026

    Zhengbo Zhang, Jinbo Su, Zhaowen Zhou, Changtao Miao, Yuhan Hong, Qimeng Wu, Yumeng Liu, Feier Wu, Yihe Tian, Yuhao Liang, et al. Visbrowse-bench: Benchmarking visual-native search for multimodal browsing agents.arXiv preprint arXiv:2603.16289, 2026

    arXiv 2026

  14. [14]

    Webweaver: Structuring web-scale evidence with dynamic outlines for open-ended deep research.arXiv preprint arXiv:2509.13312, 2025

    Zijian Li, Xin Guan, Bo Zhang, Shen Huang, Houquan Zhou, Shaopeng Lai, Ming Yan, Yong Jiang, Pengjun Xie, Fei Huang, et al. Webweaver: Structuring web-scale evidence with dynamic outlines for open-ended deep research.arXiv preprint arXiv:2509.13312, 2025

    arXiv 2025

  15. [15]

    Deepresearch bench: A comprehensive benchmark for deep research agents.arXiv preprint arXiv:2506.11763, 2025

    Mingxuan Du, Benfeng Xu, Chiwei Zhu, Xiaorui Wang, and Zhendong Mao. Deepresearch bench: A comprehensive benchmark for deep research agents.arXiv preprint arXiv:2506.11763, 2025. 11

    Pith/arXiv arXiv 2025

  16. [16]

    τ-bench: A benchmark for tool-agent-user interaction in real-world domains.arXiv preprint arXiv:2406.12045, 2024

    Shunyu Yao, Noah Shinn, Pedram Razavi, and Karthik Narasimhan. τ-bench: A benchmark for tool-agent-user interaction in real-world domains.arXiv preprint arXiv:2406.12045, 2024

    Pith/arXiv arXiv 2024

  17. [17]

    τ 2-Bench: Eval- uating conversational agents in a dual-control environment.arXiv preprint arXiv:2506.07982, 2025

    Victor Barres, Honghua Dong, Soham Ray, Xujie Si, and Karthik Narasimhan. τ 2-Bench: Eval- uating conversational agents in a dual-control environment.arXiv preprint arXiv:2506.07982, 2025

    Pith/arXiv arXiv 2025

  18. [18]

    React: Synergizing reasoning and acting in language models

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. InThe eleventh international conference on learning representations, 2022

    2022

  19. [19]

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models.arXiv preprint arXiv:2402.03300, 2024

    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 mathematical reasoning in open language models.arXiv preprint arXiv:2402.03300, 2024

    Pith/arXiv arXiv 2024

  20. [20]

    Dapo: An open-source llm reinforcement learning system at scale.arXiv preprint arXiv:2503.14476, 2025

    Qiying Yu, Zheng Zhang, Ruofei Zhu, Yufeng Yuan, Xiaochen Zuo, Yu Yue, Weinan Dai, Tiantian Fan, Gaohong Liu, Lingjun Liu, et al. Dapo: An open-source llm reinforcement learning system at scale.arXiv preprint arXiv:2503.14476, 2025

    Pith/arXiv arXiv 2025

  21. [21]

    The landscape of agentic reinforcement learning for llms: A survey.arXiv preprint arXiv:2509.02547, 2025

    Guibin Zhang, Hejia Geng, Xiaohang Yu, Zhenfei Yin, Zaibin Zhang, Zelin Tan, Heng Zhou, Zhongzhi Li, Xiangyuan Xue, Yijiang Li, et al. The landscape of agentic reinforcement learning for llms: A survey.arXiv preprint arXiv:2509.02547, 2025

    Pith/arXiv arXiv 2025

  22. [22]

    Search-r1: Training llms to reason and leverage search engines with reinforcement learning.arXiv preprint arXiv:2503.09516, 2025

    Bowen Jin, Hansi Zeng, Zhenrui Yue, Jinsung Yoon, Sercan Arik, Dong Wang, Hamed Za- mani, and Jiawei Han. Search-r1: Training llms to reason and leverage search engines with reinforcement learning.arXiv preprint arXiv:2503.09516, 2025

    Pith/arXiv arXiv 2025

  23. [23]

    Search-o1: Agentic search-enhanced large reasoning models

    Xiaoxi Li, Guanting Dong, Jiajie Jin, Yuyao Zhang, Yujia Zhou, Yutao Zhu, Peitian Zhang, and Zhicheng Dou. Search-o1: Agentic search-enhanced large reasoning models. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 5420–5438, 2025

    2025

  24. [24]

    Retool: Reinforcement learning for strategic tool use in llms.arXiv preprint arXiv:2504.11536, 2025

    Jiazhan Feng, Shijue Huang, Xingwei Qu, Ge Zhang, Yujia Qin, Baoquan Zhong, Chengquan Jiang, Jinxin Chi, and Wanjun Zhong. Retool: Reinforcement learning for strategic tool use in llms.arXiv preprint arXiv:2504.11536, 2025

    Pith/arXiv arXiv 2025

  25. [25]

    Simpletir: End-to-end reinforcement learning for multi-turn tool-integrated reasoning.arXiv preprint arXiv:2509.02479, 2025

    Zhenghai Xue, Longtao Zheng, Qian Liu, Yingru Li, Xiaosen Zheng, Zejun Ma, and Bo An. Simpletir: End-to-end reinforcement learning for multi-turn tool-integrated reasoning.arXiv preprint arXiv:2509.02479, 2025

    Pith/arXiv arXiv 2025

  26. [26]

    Tongyi deepresearch technical report.arXiv preprint arXiv:2510.24701, 2025

    Tongyi DeepResearch Team, Baixuan Li, Bo Zhang, Dingchu Zhang, Fei Huang, Guangyu Li, Guoxin Chen, Huifeng Yin, Jialong Wu, Jingren Zhou, et al. Tongyi deepresearch technical report.arXiv preprint arXiv:2510.24701, 2025

    Pith/arXiv arXiv 2025

  27. [27]

    Autotool: Dynamic tool selection and integration for agentic reasoning.arXiv preprint arXiv:2512.13278, 2025

    Jiaru Zou, Ling Yang, Yunzhe Qi, Sirui Chen, Mengting Ai, Ke Shen, Jingrui He, and Mengdi Wang. Autotool: Dynamic tool selection and integration for agentic reasoning.arXiv preprint arXiv:2512.13278, 2025

    Pith/arXiv arXiv 2025

  28. [28]

    Generalizable end-to-end tool-use rl with synthetic codegym.arXiv preprint arXiv:2509.17325, 2025

    Weihua Du, Hailei Gong, Zhan Ling, Kang Liu, Lingfeng Shen, Xuesong Yao, Yufei Xu, Dingyuan Shi, Yiming Yang, and Jiecao Chen. Generalizable end-to-end tool-use rl with synthetic codegym.arXiv preprint arXiv:2509.17325, 2025

    arXiv 2025

  29. [29]

    Webthinker: Empowering large reasoning models with deep research capability.arXiv preprint arXiv:2504.21776, 2025

    Xiaoxi Li, Jiajie Jin, Guanting Dong, Hongjin Qian, Yongkang Wu, Ji-Rong Wen, Yutao Zhu, and Zhicheng Dou. Webthinker: Empowering large reasoning models with deep research capability.arXiv preprint arXiv:2504.21776, 2025

    Pith/arXiv arXiv 2025

  30. [30]

    Webdancer: Towards autonomous information seeking agency.arXiv preprint arXiv:2505.22648, 2025

    Jialong Wu, Baixuan Li, Runnan Fang, Wenbiao Yin, Liwen Zhang, Zhengwei Tao, Dingchu Zhang, Zekun Xi, Gang Fu, Yong Jiang, et al. Webdancer: Towards autonomous information seeking agency.arXiv preprint arXiv:2505.22648, 2025

    Pith/arXiv arXiv 2025

  31. [31]

    Websailor-v2: Bridging the chasm to proprietary agents via synthetic data and scalable reinforcement learning.arXiv preprint arXiv:2509.13305, 2025

    Kuan Li, Zhongwang Zhang, Huifeng Yin, Rui Ye, Yida Zhao, Liwen Zhang, Litu Ou, Dingchu Zhang, Xixi Wu, Jialong Wu, et al. Websailor-v2: Bridging the chasm to proprietary agents via synthetic data and scalable reinforcement learning.arXiv preprint arXiv:2509.13305, 2025. 12

    arXiv 2025

  32. [32]

    Webresearcher: Unleashing unbounded reasoning capability in long-horizon agents.arXiv preprint arXiv:2509.13309, 2025

    Zile Qiao, Guoxin Chen, Xuanzhong Chen, Donglei Yu, Wenbiao Yin, Xinyu Wang, Zhen Zhang, Baixuan Li, Huifeng Yin, Kuan Li, et al. Webresearcher: Unleashing unbounded reasoning capability in long-horizon agents.arXiv preprint arXiv:2509.13309, 2025

    arXiv 2025

  33. [33]

    Scaling agents via continual pre-training.arXiv preprint arXiv:2509.13310, 2025

    Liangcai Su, Zhen Zhang, Guangyu Li, Zhuo Chen, Chenxi Wang, Maojia Song, Xinyu Wang, Kuan Li, Jialong Wu, Xuanzhong Chen, et al. Scaling agents via continual pre-training.arXiv preprint arXiv:2509.13310, 2025

    arXiv 2025

  34. [34]

    Openseeker: Democratizing frontier search agents by fully open-sourcing training data.arXiv preprint arXiv:2603.15594, 2026

    Yuwen Du, Rui Ye, Shuo Tang, Xinyu Zhu, Yijun Lu, Yuzhu Cai, and Siheng Chen. Openseeker: Democratizing frontier search agents by fully open-sourcing training data.arXiv preprint arXiv:2603.15594, 2026

    arXiv 2026

  35. [35]

    Vision-deepresearch benchmark: Rethinking visual and textual search for multimodal large language models.arXiv preprint arXiv:2602.02185, 2026

    Yu Zeng, Wenxuan Huang, Zhen Fang, Shuang Chen, Yufan Shen, Yishuo Cai, Xiaoman Wang, Zhenfei Yin, Lin Chen, Zehui Chen, et al. Vision-deepresearch benchmark: Rethinking visual and textual search for multimodal large language models.arXiv preprint arXiv:2602.02185, 2026

    arXiv 2026

  36. [36]

    Multi-modal agent tuning: Building a vlm-driven agent for efficient tool usage.arXiv preprint arXiv:2412.15606, 2024

    Zhi Gao, Bofei Zhang, Pengxiang Li, Xiaojian Ma, Tao Yuan, Yue Fan, Yuwei Wu, Yunde Jia, Song-Chun Zhu, and Qing Li. Multi-modal agent tuning: Building a vlm-driven agent for efficient tool usage.arXiv preprint arXiv:2412.15606, 2024

    Pith/arXiv arXiv 2024

  37. [37]

    Learning to deliberate: Meta-policy collaboration for agentic llms with multi-agent reinforcement learning

    Wei Yang and Jesse Thomason. Learning to deliberate: Meta-policy collaboration for agentic llms with multi-agent reinforcement learning. InProceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 29820–29828, 2026

    2026

  38. [38]

    Heterogeneous agent collaborative reinforcement learning.arXiv preprint arXiv:2603.02604, 2026

    Zhixia Zhang, Zixuan Huang, Gongxun Li, Huaiyang Wang, Chengyi Yuan, Xin Xia, De- qing Wang, Fuzhen Zhuang, Shuai Ma, Ning Ding, et al. Heterogeneous agent collaborative reinforcement learning.arXiv preprint arXiv:2603.02604, 2026

    Pith/arXiv arXiv 2026

  39. [39]

    Agentic reinforced policy optimization

    Guanting Dong, Hangyu Mao, Kai Ma, Licheng Bao, Yifei Chen, Zhongyuan Wang, Zhongxia Chen, Jiazhen Du, Huiyang Wang, Fuzheng Zhang, et al. Agentic reinforced policy optimization. arXiv preprint arXiv:2507.19849, 2025

    Pith/arXiv arXiv 2025

  40. [40]

    Agentic entropy-balanced policy optimization.arXiv preprint arXiv:2510.14545, 2025

    Guanting Dong, Licheng Bao, Zhongyuan Wang, Kangzhi Zhao, Xiaoxi Li, Jiajie Jin, Jing- han Yang, Hangyu Mao, Fuzheng Zhang, Kun Gai, et al. Agentic entropy-balanced policy optimization.arXiv preprint arXiv:2510.14545, 2025

    arXiv 2025

  41. [41]

    Distilling the knowledge in a neural network

    Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015

    Pith/arXiv arXiv 2015

  42. [42]

    A reduction of imitation learning and structured prediction to no-regret online learning

    Stéphane Ross, Geoffrey Gordon, and Drew Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. InProceedings of the fourteenth interna- tional conference on artificial intelligence and statistics, pages 627–635. JMLR Workshop and Conference Proceedings, 2011

    2011

  43. [43]

    Agentic context engineering: Evolving contexts for self-improving language models.arXiv preprint arXiv:2510.04618, 2025

    Qizheng Zhang, Changran Hu, Shubhangi Upasani, Boyuan Ma, Fenglu Hong, Vamsidhar Ka- manuru, Jay Rainton, Chen Wu, Mengmeng Ji, Hanchen Li, et al. Agentic context engineering: Evolving contexts for self-improving language models.arXiv preprint arXiv:2510.04618, 2025

    Pith/arXiv arXiv 2025

  44. [44]

    Cl-bench: A benchmark for context learning

    Shihan Dou, Ming Zhang, Zhangyue Yin, Chenhao Huang, Yujiong Shen, Junzhe Wang, Jiayi Chen, Yuchen Ni, Junjie Ye, Cheng Zhang, et al. Cl-bench: A benchmark for context learning. arXiv preprint arXiv:2602.03587, 2026

    arXiv 2026

  45. [45]

    Openai gpt-5 system card.arXiv preprint arXiv:2601.03267, 2025

    Aaditya Singh, Adam Fry, Adam Perelman, Adam Tart, Adi Ganesh, Ahmed El-Kishky, Aidan McLaughlin, Aiden Low, AJ Ostrow, Akhila Ananthram, et al. Openai gpt-5 system card.arXiv preprint arXiv:2601.03267, 2025

    Pith/arXiv arXiv 2025

  46. [46]

    Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities.arXiv preprint arXiv:2507.06261, 2025

    Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, et al. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities.arXiv preprint arXiv:2507.06261, 2025. 13

    Pith/arXiv arXiv 2025

  47. [47]

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

    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, 2025

    Pith/arXiv arXiv 2025

  48. [48]

    Through the lens of contrast: Self-improving visual reasoning in vlms.arXiv preprint arXiv:2603.02556, 2026

    Zhiyu Pan, Yizheng Wu, Jiashen Hua, Junyi Feng, Shaotian Yan, Bing Deng, Zhiguo Cao, and Jieping Ye. Through the lens of contrast: Self-improving visual reasoning in vlms.arXiv preprint arXiv:2603.02556, 2026

    arXiv 2026

  49. [49]

    slime: An llm post-training framework for rl scaling

    Zilin Zhu, Chengxing Xie, Xin Lv, and slime Contributors. slime: An llm post-training framework for rl scaling. https://github.com/THUDM/slime, 2025. GitHub repository. Corresponding author: Xin Lv

    2025

  50. [50]

    name": "parse_file

    Ziwen Han, Meher Mankikar, Julian Michael, and Zifan Wang. Search-time data contamination. arXiv preprint arXiv:2508.13180, 2025. 14 A Detailed case study We first present a failed inference trajectory from the Tongyi DeepResearch Agent in Appendix A.1, with the corresponding agent prompt provided in Appendix B.1. We then show the teacher model’s counterf...

    Pith/arXiv arXiv 2025

  51. [51]

    However, there are subtle 16 differences such as the background color and the presence of text in the search result image

    [Wizard Tiny Style 2D Character Sprites - CraftPix.net](Web page URL: https://craftpix.net/product/wizard-tiny-style-2d-character-sprites/) Source: CraftPix The original query image and the search result image both feature a cartoonish character with green hair, yellow eyes, and a black mask holding a staff. However, there are subtle 16 differences such a...

  52. [52]

    Image URL: https://encrypted-tbn0.gstatic.com/images?q=tbn:ANd9GcTtGe3MHCcQf5- Z0NLydwHSLvYpwVGUOSCq8YmKUMZU_upmR7fF

    [Discover 12 enemies for games and skeleton characters in games ideas | skele- ton warriors cartoon, 2d game rock sprites, skeleton game character and more](Web page URL: https://www.pinterest.com/giorgoschar06/enemies-for-games/) Source: Pinterest The original query image and the search result image share several visual similarities including a character...

  53. [53]

    [Before The Halloween Update All Animals Evolution With Patrick Reaper Ghostly Pumpkin (EvoWorld.io) - YouTube](Web page URL: https://www.youtube.com/watch?v=ewpDuL2JsB0) Source: YouTube The original query image and the webpage image share several visual similarities including a character with green hair, yellow eyes, and a black mask but differ in contex...

  54. [54]

    [Skins | EvoWorld.io Wiki | Fandom](Web page URL: https://evoworldio.fandom.com/wiki/Skins) Source: Fandom The original query image and the search result image both feature cartoon characters wearing green hooded outfits but exhibit distinct differences in design and context. The character in the original query image has black facial features with yellow ...

  55. [55]

    [Kaden | The Rewrittenverse Wiki | Fandom](Web page URL: https://the- rewrittenverse.fandom.com/wiki/Kaden) Source: Fandom The original query image and the search result image both feature a character wearing a green hooded mask with glowing yellow eyes and a skull emblem on their chest indicating a 17 thematic similarity. However, there are notable diffe...

  56. [56]

    North-East Europe

    [North-East Europe Is Kinda Toxic | Evoworld - YouTube](Web page URL: https://www.youtube.com/watch?v=_kGnzn0mUVg) Source: YouTube The original query image and the search result image both feature a character with green hair, yellow eyes, and a black mask which suggests a strong visual connection. However, the original query image shows a single character...

  57. [57]

    Despite these differences, the central character appears visually consistent in design and style

    [Grinding to Rank #1 in West Europe 1 | EvoWorld Gameplay Live - YouTube](Web page URL: https://www.youtube.com/watch?v=JRxO_19M1gY) Source: YouTube The original query image and the search result image both feature a character with green hair, yellow eyes, and a black mask but the original query image shows a single character while the search result image...

  58. [58]

    Despite these differences, the core design elements including the pose, clothing style, and overall silhouette are remarkably similar suggesting a strong connection between the two

    [Green Witcher | Raid Heroes Wiki | Fandom](Web page URL: https://raid- heroes.fandom.com/wiki/Green_Witcher) Source: Fandom The original query image and the search result image both feature a cartoon character wearing a green hooded outfit with a mask but differ in several details such as eye color, additional accessories like a skull emblem, and backgro...

  59. [59]

    [GEE estimates of negative binomial models for number of first time... | Download Table](Web page URL: https://www.researchgate.net/figure/GEE-estimates-of-negative- binomial-models-for-number-of-first-time-entries-in-the-segment_tbl1_237290427) Source: ResearchGate The original query image and the search result image share striking similarities including...

  60. [60]

    Wizard Tiny Style 2D Character Sprites

    [Cloaked Plant Wizard | Prodigy Game Wiki | Fandom](Web page URL: https://prodigy- game.fandom.com/wiki/Cloaked_Plant_Wizard) Source: Fandom The original query image and the search result image share similar thematic elements such as a character wearing a green hooded outfit but differ significantly in design and context. The original query image features...

    2017

  61. [61]

    The green hood provides a Damage Resistance of 1 and can be worn with a gas mask,

    [Green hood (Fallout 4) - Fallout Wiki](https://fallout.fandom.com/wiki/Green_hood_(Fallout_ 4)) The green hood is a piece of headwear that can be obtained in Fallout 4. The green hood provides a Damage Resistance of 1 and can be worn with a gas mask, ... 20

  62. [62]

    Pet Level Required: 8

    [Green Hood | IMO: The World of Magic Wiki - Fan- dom](https://imotwom.fandom.com/wiki/Green_Hood) Green Hood A hooded fellow to help guide you. Pet Level Required: 8. Class Re- quired: Magician-Class · Ranger-Class · Warrior-Class All Critical: 1

  63. [63]

    [The Ultimater Guide to Green Hood! [3k Words!] | Hypixel Forums](https://hypixel.net/threads/the-ultimater-guide-to-green-hood-3k-words.1209267/) Date published: Jul 5, 2017 Cake Monster is a bad teammate for Hood, he likes to combo the hell out of every- one with insane DPS but has no way to work with Green Hood

    2017

  64. [64]

    [[TOMT] [GAME] very popular game where you play as a bald kid with ...](https://www.reddit.com/r/tipofmytongue/comments/16ff1qg/tomt_gam e_very_popular_game_where_you_play_as_a/) Date published: Sep 10, 2023 [TOMT] (anime character) blue hair green hood. 2. 4. Did ya see that ... [2007/2008][FLASH] Top Down game where the main character has a green

    2023

  65. [65]

    The Worst Character

    [Smash Heroes: Green Hood - YouTube](https://www.youtube.com/watch?v=4gPONz90lrs) Date published: Dec 23, 2016 Happy 100th video! In this, I will be showing you guys how to use Green Hood effectively and to the best of his ability ... The Worst Character

    2016

  66. [66]

    Discover videos related to Green Hood Guy Roblox on TikTok

    [Green Hood Guy Roblox - TikTok](https://www.tiktok.com/discover/green-hood-guy- roblox) Date published: Feb 2, 2026 5.8M posts. Discover videos related to Green Hood Guy Roblox on TikTok. See more videos about Roblox Green Guy, Who Is Green Guy with Green

    2026

  67. [67]

    games, Roblox hype culture, comment your favorite Roblox game, Roblox player community

    [Exploring the Green Hooded Avatars in Roblox - Tik- Tok](https://www.tiktok.com/@creepiecrawlies6/video/7540798118984502542) Date published: Aug 20, 2025 ... games, Roblox hype culture, comment your favorite Roblox game, Roblox player community ... Watch the hilarious video of a dog in a green hood

    arXiv 2025

  68. [68]

    *NEW* GREEN HOOD GAME- PLAY!

    [Smash Heroes] Minecraft - "*NEW* GREEN HOOD GAME- PLAY!"](https://www.youtube.com/watch?v=nhPS8UzR5_k) Date published: Aug 17, 2016 ... video description’s due to a false positive on YouTube’s end. As ... GREEN HOOD GAMEPLAY!" 3.2K views · 9 years ago. #Christian ...more

    2016

  69. [69]

    This piece of headwear appears as a hood normally connected to the assault gas

    [Green hood (Fallout 76) | Fallout Wiki - Fan- dom](https://fallout.fandom.com/wiki/Green_hood_(Fallout_76)) The green hood is a piece of headwear that can be obtained in Fallout 76. This piece of headwear appears as a hood normally connected to the assault gas

  70. [70]

    green hooded character game screenshot

    [Green Hood Character - Pinterest](https://www.pinterest.com/ideas/green-hood- character/907747242196/) 21 Discover Pinterest’s best ideas and inspiration for Green hood character. Get in- spired and try out new things. Last updated 4d. </tool_response> Step 5 <think> Search results are not particularly helpful. Let’s try search for "green hooded characte...

    arXiv

  71. [71]

    - TikTok](https://www.tiktok.com/discover/what-does-a-roblox-avatar-with-a-green-hood- green-skin-and-round-head-mean) Date published: Mar 2, 2026

    [What Does A Roblox Avatar with A Green Hood Green Skin ... - TikTok](https://www.tiktok.com/discover/what-does-a-roblox-avatar-with-a-green-hood- green-skin-and-round-head-mean) Date published: Mar 2, 2026 ... character styles, detecting green hooded avatars, gaming ... Squid Game Roblox Avatar 0 Robux Whit Green Tshirt and Tshirt Screenshot

    2026

  72. [72]

    [[TOMT] (anime character) blue hai r green hood - Reddit](https://www.reddit.com/r/tipofmytongue/comments/5l5zte/tomt_ textgreateranime_character_blue_hair_green_hood/) Date published: Dec 30, 2016 [TOMT] (anime character) blue hair green hood. Solved. I can’t ... screenshot that didn’t have the bunny ears, and thought the lace was

    2016

  73. [73]

    keep reading

    [How to get Green Bean outfit in Roblox & Green Avatar explained](https://www.dexerto.com/roblox/how-to-get-green-bean-outfit-in-roblox- green-avatar-explained-3240530/) Date published: Aug 20, 2025 Grow a garden green hood ... keep reading. Ro Ghoul Roblox character in the diner. RobloxRo Ghoul codes (April 2026) · A screenshot featuring

    2025

  74. [74]

    Green Hood in Roblox, What Is The Green Hood

    [Who Is Green Guy with Green Hoodie in Roblox - TikTok](https://www.tiktok.com/discover/who-is-green-guy-with-green-hoodie-in-roblox) Date published: 5 days ago ... Green Hood in Roblox, What Is The Green Hood ... Green Man Roblox stories, Roblox gameplay tips, Roblox character exploration, Roblox game

  75. [75]

    a close up of a cartoon character with a green hood

    [Berry avenue creature - Discover & Share GIFs - Tenor](https://tenor.com/view/berry- avenue-creature-gif-3355296347644183555) Date published: Oct 30, 2024 Berry Avenue Creature Meme. a close up of a cartoon character with a green hood ... a screenshot of a video game with the words the audacity

    2024

  76. [76]

    A 22 ...](https://www.instagram.com/p/DNN__sesyj4/) Plus a green hood

    [My loot of last week Plus a green hood, because of rain and cold. A 22 ...](https://www.instagram.com/p/DNN__sesyj4/) Plus a green hood ... May be a video game screenshot. 17. chimerical_dragonfly ... I create everything to complete a character and tell a story, from D&D decor to

  77. [77]

    Cara main Game PS2 di kom- puter kalian, cocok buat nostalgia game jadul kesukaan kalian nih

    [Roblox Green Avtsr Iwth Green Hood - TikTok](https://www.tiktok.com/discover/roblox- green-avtsr-iwth-green-hood) Date published: Jan 5, 2026 suggest character nga #fyp #roblox #character ... Cara main Game PS2 di kom- puter kalian, cocok buat nostalgia game jadul kesukaan kalian nih

    2026

  78. [78]

    [Gods will fall. Wishlist now! https://f.mtr.cool/hdtqxfalbl - Insta- gram](https://www.instagram.com/reel/DWCW0mxjG7J/) Date published: Mar 18, 2026 A stylized illustration featuring a character in a green hood with glowing eyes and horns, ... Photo by Pupsker on March 09, 2026

    2026

  79. [79]

    a stuffed doll with a green hood and a red zip- per

    [Makoto Naegi Danganronpa Sticker - Tenor](https://tenor.com/view/makoto- naegi-danganronpa-gif-4948538113774618876) Date published: Jan 19, 2024 Makoto Naegi Danganronpa Sticker. a stuffed doll with a green hood and a red zip- per ... a screenshot of a video game with a character named

    2024

  80. [80]

    David Cseh

    [HD wallpaper: Fantasy, Warrior, Forest, Green, Hood 4096x2304px ...](https://www.wallpaperflare.com/fantasy-warrior-forest-green-hood-wallpaper- might/download/4096x2304) Fantasy, Warrior, Forest, Green, Hood ... warrior holding sword wallpaper, video game screenshot ... black haired girl anime character illustration ... </tool_response> . . . . . . . . ...

    2017

Showing first 80 references.