Pith. sign in

REVIEW 5 major objections 6 minor 66 references

A single LLM agent that can rewrite its own sub-goals, tools, and context mid-task beats static agents by up to 24.1 points on long-horizon benchmarks.

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

2026-08-03 03:25 UTC pith:7C3AAT2V

load-bearing objection A genuinely new agent self-reconfiguration architecture, but the key ablation is missing a trained control, so the headline gain is not yet pinned to the claimed mechanism. the 5 major comments →

arxiv 2602.07883 v3 pith:7C3AAT2V submitted 2026-02-08 cs.AI

ToolSelf: Unifying Task Execution and Self-Reconfiguration via Tool-Driven Emergent Adaptation

Jingqi Zhou , Sheng Wang , Dezhao Deng , Junwen Lu , Junwei Su , Qintong Li , Jiahui Gao , Hao Wu
show 4 more authors
Jiyue Jiang Lingpeng Kong Dunhong Jin Chuan Wu
This is my paper
classification cs.AI
keywords LLM agentsself-reconfigurationtool-driven adaptationlong-horizon tasksrejection sampling fine-tuningKTO reinforcement learningagent configurationruntime adaptation
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.

ToolSelf argues that the rigidity of LLM agents comes from configurations fixed before execution. It proposes treating configuration updates—new sub-goals, strategies, toolboxes, and context—as a callable tool, so the agent decides when and how to reconfigure itself in the same action space it uses for the task. The paper shows this single generalist matches or surpasses specialized multi-agent workflows across research, general-assistant, and coding benchmarks, and that a two-stage training recipe (imitation then trajectory-level reinforcement) adds substantial further gains. If right, hand-designed orchestration and context-management heuristics become unnecessary for long-horizon agentic tasks.

Core claim

The central claim is that self-reconfiguration can be internalized as a tool action rather than an external mechanism. ToolSelf equips the inference agent with a reconfiguration tool that, when invoked, archives a stage summary and a request (proposed next sub-goal, rationale, and requirements) into a reconfiguration engine µ, which outputs a fresh configuration C_{i+1} = (sub-goal, strategy, toolbox, context). Because the inference agent and the engine are two modes of the same base model, the loop is end-to-end trainable. The paper reports consistent gains: on a 14B model, ToolSelf reaches 56.0% on FRAMES versus 38.0% for a vanilla ReAct agent, 16.1% on SWE-bench Lite versus 14.2% for a sp

What carries the argument

The reconfiguration tool Treconfig and the reconfiguration engine µ: configurations are formalized as C_i = (sub-goal q_i, strategy σ_i, toolbox T_i, context K_i), and the engine rewrites them from the agent's own execution summary H_i and request r_i. The second piece is Configuration-Aware Two-stage Training (CAT): rejection-sampling fine-tuning to cold-start both adapters, then KTO reinforcement learning with trajectory-level credit assignment to propagate final success or failure back to every reconfiguration decision. The key move is turning “when to adapt” and “how to adapt” into decisions inside the policy's action space, so they can be optimized rather than heuristically patched.

Load-bearing premise

The load-bearing premise is that the inference agent can reliably judge its own progress and write a stage summary and proposed next sub-goal that the reconfiguration engine can trust; if that self-report is wrong, the error propagates into every subsequent configuration.

What would settle it

Corrupt the agent's proposed next sub-goal by replacing it with a random sub-goal drawn from a different task in the same benchmark, keeping the rest of the loop intact. If accuracy on a held-out suite does not drop materially, the claimed benefit of intent-driven adaptation is not carrying the result.

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

If this is right

  • A single agent can replace task-specific multi-agent pipelines on long-horizon tasks without losing generalization, because it can emulate specialized toolsets and strategies on demand.
  • Context stays bounded even when the agent explores far longer: the paper reports roughly four times more execution steps with nearly flat maximum input tokens, so self-reconfiguration doubles as a context-management primitive.
  • The training recipe is portable: RFT provides the cold start, KTO refines cross-stage meta-decisions, and the same two-stage scheme should transfer to new base models and domains.
  • Agents learn task-aware reconfiguration cadence—fewer reconfigurations on easy tasks, more on hard ones, plus task-appropriate tool selection—so the adaptation behavior itself becomes an emergent output of training.
  • Because the reconfiguration engine's output depends on the agent's own summary, the loop can become self-reinforcing: better execution yields better summaries, which yield better configurations.

Where Pith is reading between the lines

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

  • If the approach is as general as claimed, the same tool-as-configuration trick could extend to other agent properties—memory policies, reward functions, or code-generation prompts—turning any meta-decision into a trainable action.
  • The paper's hard 80%-token heuristic context cleanup sits in tension with the “autonomous triggering” narrative; a purely learned trigger might remove even that threshold and improve behavior on tasks with abrupt context spikes.
  • A stress test worth running: feed the agent deliberately misleading stage summaries and measure how far performance drops; since the engine is told to adopt the proposed sub-goal “whenever possible,” the method's ceiling is set by the agent's self-assessment accuracy.
  • Because the two modules share a base model with separate LoRA adapters, the paradigm suggests a practical design space where task execution and self-management specialize within one network—cheap to ablate and easy to extend.

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

5 major / 6 minor

Summary. The paper introduces ToolSelf, a paradigm that lets an LLM agent treat its own configuration (sub-goal, execution strategy, toolbox, context) as a callable tool, so that task execution and self-reconfiguration share one action space. A second component, the reconfiguration engine, converts the agent's request into a new stage configuration. The authors propose a two-stage training scheme, CAT, combining rejection-sampling fine-tuning (RFT) with trajectory-level KTO reinforcement learning. Evaluations on FRAMES, xbench, GAIA(WS), GAIA, and SWE-bench Lite with Qwen3-8B/14B report gains over several single-agent, single-agent-extension, and multi-agent baselines, with the trained model reaching 40.6% average accuracy on Qwen3-8B versus 16.5% for the vanilla agent. The paper also provides ablations on the triggering mechanism, configuration components, and token efficiency.

Significance. If the causal claims were established, ToolSelf would be a useful step toward agents that can adapt their own tools, goals, and context at runtime without separate orchestration modules. The unification of reconfiguration into a standard tool-call interface is simple and potentially general. The paper is also commendably detailed in its appendices: full prompts, training hyperparameters, tool definitions, and two worked case studies are provided. The trajectory-level credit assignment is a sensible way to train a meta-capability whose effect is only visible through later execution. However, the current evidence does not establish the central claim: the headline 'consistently outperforms all baselines' is contradicted by the paper's own Table 1, and the CAT gains are confounded with the presence of any RFT+KTO training signal. The contribution is plausible but needs additional controlled experiments and more cautious claims.

major comments (5)
  1. [Sec. 3.2 / Table 1] The claim that 'TOOLSELF consistently outperforms all baselines across benchmarks and model scales' is directly contradicted by Table 1. On Qwen3-8B, TOOLSELF scores 10.0 on xbench while Co-Sight scores 13.0; on Qwen3-14B, TOOLSELF ties Co-Sight at 16.0 on xbench. The sentence 'surpassing all single-agent extensions and multi-agent workflows' is therefore not accurate. This is a load-bearing inconsistency in the paper's central empirical claim. Please restrict the claim to the benchmarks and scales where it is supported, or provide additional evidence.
  2. [Sec. 3.3, Table 2 / Sec. 3.4, Table 5] The CAT training gain is confounded with training signal. TOOLSELF+RFT+KTO is trained on teacher trajectories and KTO labels, while Vanilla Agent and all other baselines in Table 2 are used zero-shot. The average gain of 24.1 points therefore conflates the unified action space with the benefit of any RFT+KTO training. The compute control in Table 5 (Vanilla + majority voting) does not control for training signal. A static ReAct agent trained on the same RFT data and the same trajectory-level KTO scheme is necessary to attribute the gain to self-reconfiguration. Without such a control, the causal claim in Sec. 3.3 is not established.
  3. [All tables / Sec. A.5.7] All reported results are pass@1 with no standard errors, no seeds, and no reruns. Several margins are small (e.g., 10.0 vs 13.0 on xbench, 16.1 vs 14.6 on SWE-Lite), and single-run evaluation cannot distinguish sampling noise from systematic gains. Report means and variances over multiple seeds, or explicitly justify determinism of the evaluation. This is necessary for the 'consistently outperforms' claim.
  4. [Abstract / Sec. 3.3 / Conclusion] The abstract states 'gains 28.8 points over the static-configuration baseline on average', but the body reports '24.1% average performance gain' (Sec. 3.3 and Conclusion), and Table 2's numbers give 40.6 - 16.5 = 24.1. Also, Table 2's 'Avg.' column averages over different benchmark subsets for different rows: ReSum, OAgents, and Co-Sight do not report GAIA (and sometimes not SWE-Lite), so the average comparisons are not apples-to-apples. Reconcile the abstract number and state explicitly which benchmarks enter each average.
  5. [Sec. A.5.1 / Sec. A.6.3 / Sec. 2] The claim of 'autonomous triggering' is weakened by external thresholds in the implementation. Sec. A.5.1 imposes an automatic context cleanup at 80% of token capacity (removing all but the last 10 iterations) and treats exceeding 30 reconfigurations as task failure; these are externally fixed heuristics, not agent decisions. In addition, the reconfiguration engine prompt (Sec. A.6.3) instructs the engine to 'adopt it whenever possible' for the agent's proposed sub-goal. This makes the engine largely a pass-through for the agent's self-report. While this is a design choice, the paper should discuss how it interacts with the claim that the reconfiguration engine contributes independently to performance, especially given the absence of a trained no-reconfiguration control.
minor comments (6)
  1. [Sec. 1] The introduction refers to 'Prior work (Appendix 4)' but the related work appears in Section 4, not an appendix.
  2. [Table 5 vs Table 1] Vanilla Agent's GAIA overall accuracy is 18.8% in Table 5 but 19.7% in Table 1 for the same Qwen3-8B setting. This discrepancy needs reconciliation or a clear statement of different evaluation subsets.
  3. [Eq. (11)] The KTO loss uses r(x), z0, v(·), and λD, λU without defining them in the main text. Please add definitions or a reference.
  4. [Sec. A.5.7] The manual verification of LLM-as-Judge is described as showing 'high agreement' but no quantitative agreement score is reported. Please provide the agreement rate and sample size.
  5. [Table 3] The Level 3 row reads '7.693.85' without a spacing separator between 7.69 and 3.85. This is a formatting typo.
  6. [General] No code, data, or model release is mentioned. Given the detail of the training setup, a reproducibility statement would strengthen the paper.

Circularity Check

0 steps flagged

No significant circularity: empirical benchmark claims are externally evaluated; no derivation reduces to its inputs.

full rationale

ToolSelf makes no first-principles derivation that could reduce to its inputs. Eqs. (6)-(7) are only probability decompositions of stage trajectories and the full ToolSelf trajectory; they state that reconfiguration outputs depend on requests and history, but they are not used to derive the benchmark gains. The gains in Tables 1-2 are empirical measurements on held-out/test splits (FRAMES test, GAIA, GAIA(WS), xbench, SWE-bench Lite), and CAT training uses external binary success labels from official evaluation scripts and teacher trajectories from separate models (Qwen3-235B-A22B-Thinking, DeepSeek-V3.2), not from the evaluated ToolSelf model. The nearest thing to a reduction is the reconfiguration engine prompt (Sec. A.6.3) instructing the engine to "adopt it whenever possible" for the agent-proposed sub-goal; the paper itself describes µ as a "compiler" (Sec. 2.4), so this is explicit delegation rather than a hidden equivalence. The absence of a trained no-reconfiguration control is a genuine experimental confound for attributing gains to the unified action space, and the hard 80% context-cleanup heuristic (Sec. A.5.1) sits uneasily with the claim of fully autonomous triggering, but these are experimental-design and consistency concerns, not circularity. No self-citation chain is load-bearing; cited methods (RFT, KTO, LoRA) are external standard techniques. Therefore no circular step is established.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 0 invented entities

The primary novel artifact is the reconfiguration tool and engine, which is an interface design rather than a new physical or mathematical entity. It does not introduce a new force, particle, or conserved quantity, and no independent falsifiable handle outside the paper is claimed. The free parameters are the usual hyperparameters of an LLM-agent system, several of which are explicitly tuned via grid search (Sec. A.5.1).

free parameters (6)
  • max_reconfiguration_count = 30
    Set as a termination threshold for reconfigurations; beyond 30 is treated as failure (Sec. A.5.1).
  • context_cleanup_threshold = 80% of 32k tokens; keep last 10 iterations
    A fixed heuristic that triggers context compression, contradicting the claim of autonomous triggering (Sec. A.5.1).
  • max_iterations = 200
    Task horizon limit for deep-research and GAIA tasks (Sec. A.5.1).
  • inference_temperature/top_p = 0.6/0.95 for inference; 0.7/0.9 for KTO sampling
    Chosen to balance exploration and stability; influence all reported results (Sec. A.5.1, A.5.3).
  • LoRA rank and scaling = r=16, alpha=32
    Parameter-efficient fine-tuning choices; no sensitivity analysis (Sec. A.5.3).
  • KTO and RFT hyperparameters = lambda_D=1.0, lambda_U=1.0, beta=0.1; lr=8e-6 RFT, lr=3e-6 KTO
    Standard RL fine-tuning coefficients chosen without reported sensitivity analysis (Sec. A.5.3).
axioms (4)
  • domain assumption The inference agent can accurately summarize its own stage and propose a useful next sub-goal.
    The reconfiguration engine is told to adopt the proposed sub-goal whenever possible (Sec. A.6.3), so the whole system depends on this self-assessment capability.
  • domain assumption Trajectory-level binary credit assignment is sufficient to learn good reconfiguration decisions.
    KTO propagates task success/failure to every stage decision (Sec. 2.3), which may wash out credit or reinforce lucky trajectories.
  • domain assumption LLM-as-judge with Qwen2.5-72B-Instruct matches human correctness judgments.
    Used to evaluate FRAMES, xbench, and GAIA; the paper states manual verification on a subset showed high agreement but gives no quantitative metric (Sec. A.5.7).
  • domain assumption Benchmark test sets are representative of agent capability and the reported single-run pass@1 is stable.
    No repeated seeds or confidence intervals are reported; variance from LLM sampling could be substantial (Sec. 3.2).

pith-pipeline@v1.3.0-alltime-deepseek · 3939 in / 4223 out tokens · 119475 ms · 2026-08-03T03:25:40.034550+00:00 · methodology

0 comments
read the original abstract

LLM-powered agentic systems excel at complex long-horizon tasks, but remain constrained by static configurations fixed before execution. Such rigidity forces a trade-off between domain-specific performance and cross-task generalization: strong priors and compact tool spaces aid specialization but weaken transfer, while task-agnostic workflows and broad action spaces expand coverage but dilute guidance. Existing pre-execution optimization, planner-worker orchestration, and configuration patching fall short of resolving this tension, as they decouple adaptation from execution, causing information loss, fragmented optimization, and ambiguous credit assignment. We propose ToolSelf, a tool-driven runtime self-reconfiguration paradigm that abstracts configuration updates as a standardized tool interface and unifies execution and adaptation within one policy's action space. The execution agent can dynamically update sub-goals, strategies, toolboxes, context, and context-management modes based on task progress and feedback. We further introduce Configuration-Aware Two-stage Training (CAT), which combines rejection sampling fine-tuning with trajectory-level KTO reinforcement learning to internalize self-reconfiguration. Across diverse benchmarks, zero-shot ToolSelf rivals task-specialized agents; after CAT training, ToolSelf gains 28.8 points over the static-configuration baseline on average, illuminating a path toward emergent adaptivity that obviates manually injected guidance.

Figures

Figures reproduced from arXiv: 2602.07883 by Chuan Wu, Dezhao Deng, Dunhong Jin, Hao Wu, Jiahui Gao, Jingqi Zhou, Jiyue Jiang, Junwei Su, Junwen Lu, Lingpeng Kong, Qintong Li, Sheng Wang.

Figure 1
Figure 1. Figure 1: Illustration of TOOLSELF. (a) Multi-Agent Workflows rely on manual priors with poor generalization; (b) Single-Agent Extensions apply fragmented patches via external heuristic mechanisms; (c) TOOLSELF unifies task execution and self-reconfiguration into a single action space, achieving intrinsic and learnable adaptation. (d) The system operates through an inference-reconfiguration loop. In each stage i, th… 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

66 extracted references · 23 linked inside Pith

  1. [1]

    Governing ai agents.arXiv preprint arXiv:2501.07913, 2025

    Noam Kolt. Governing ai agents.arXiv preprint arXiv:2501.07913, 2025

    Pith/arXiv arXiv 2025

  2. [2]

    React: Synergizing reasoning and acting in language models.arXiv preprint arXiv:2210.03629, 2022

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models.arXiv preprint arXiv:2210.03629, 2022

    Pith/arXiv arXiv 2022

  3. [3]

    Agentic reasoning: A streamlined framework for enhancing llm reasoning with agentic tools.arXiv preprint arXiv:2502.04644, 2025

    Junde Wu, Jiayuan Zhu, Yuyuan Liu, Min Xu, and Yueming Jin. Agentic reasoning: A streamlined framework for enhancing llm reasoning with agentic tools.arXiv preprint arXiv:2502.04644, 2025

    Pith/arXiv arXiv 2025

  4. [4]

    Swe-agent: Agent-computer interfaces enable automated software engineering.Advances in Neural Information Processing Systems, 37:50528–50652, 2024

    John Yang, Carlos E Jimenez, Alexander Wettig, Kilian Lieret, Shunyu Yao, Karthik Narasimhan, and Ofir Press. Swe-agent: Agent-computer interfaces enable automated software engineering.Advances in Neural Information Processing Systems, 37:50528–50652, 2024

    2024

  5. [5]

    Appbench: Planning of multiple apis from various apps for complex user instruction.arXiv preprint arXiv:2410.19743, 2024

    Hongru Wang, Rui Wang, Boyang Xue, Heming Xia, Jingtao Cao, Zeming Liu, Jeff Z Pan, and Kam-Fai Wong. Appbench: Planning of multiple apis from various apps for complex user instruction.arXiv preprint arXiv:2410.19743, 2024

    Pith/arXiv arXiv 2024

  6. [6]

    Ui-tars: Pioneering automated gui interaction with native agents.arXiv preprint arXiv:2501.12326, 2025

    Yujia Qin, Yining Ye, Junjie Fang, Haoming Wang, Shihao Liang, Shizuo Tian, Junda Zhang, Jiahao Li, Yunxin Li, Shijue Huang, et al. Ui-tars: Pioneering automated gui interaction with native agents.arXiv preprint arXiv:2501.12326, 2025

    Pith/arXiv arXiv 2025

  7. [7]

    Autoagents: A framework for automatic agent generation.arXiv preprint arXiv:2309.17288, 2023

    Guangyao Chen, Siwei Dong, Yu Shu, Ge Zhang, Jaward Sesay, B ¨orje F Karlsson, Jie Fu, and Yemin Shi. Autoagents: A framework for automatic agent generation.arXiv preprint arXiv:2309.17288, 2023

    Pith/arXiv arXiv 2023

  8. [8]

    Octotools: An agentic framework with extensible tools for complex reasoning.arXiv preprint arXiv:2502.11271, 2025

    Pan Lu, Bowen Chen, Sheng Liu, Rahul Thapa, Joseph Boen, and James Zou. Octotools: An agentic framework with extensible tools for complex reasoning.arXiv preprint arXiv:2502.11271, 2025

    Pith/arXiv arXiv 2025

  9. [9]

    Mem1: Learning to synergize memory and reasoning for efficient long-horizon agents

    Zijian Zhou, Ao Qu, Zhaoxuan Wu, Sunghwan Kim, Alok Prakash, Daniela Rus, Jinhua Zhao, Bryan Kian Hsiang Low, and Paul Pu Liang. Mem1: Learning to synergize memory and reasoning for efficient long-horizon agents. arXiv preprint arXiv:2506.15841, 2025

    Pith/arXiv arXiv 2025

  10. [10]

    Large language model based multi-agents: A survey of progress and challenges.arXiv preprint arXiv:2402.01680, 2024

    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, 2024

    Pith/arXiv arXiv 2024

  11. [11]

    Llm agents making agent tools

    Georg W¨olflein, Dyke Ferber, Daniel Truhn, Ognjen Arandjelovic, and Jakob Nikolas Kather. Llm agents making agent tools. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 26092–26130, 2025

    2025

  12. [12]

    Owl: Optimized workforce learning for general multi-agent assistance in real-world task automation.arXiv preprint arXiv:2505.23885, 2025

    Mengkang Hu, Yuhang Zhou, Wendong Fan, Yuzhou Nie, Bowei Xia, Tao Sun, Ziyu Ye, Zhaoxuan Jin, Yingru Li, Qiguang Chen, et al. Owl: Optimized workforce learning for general multi-agent assistance in real-world task automation.arXiv preprint arXiv:2505.23885, 2025

    Pith/arXiv arXiv 2025

  13. [13]

    Magentic-one: A generalist multi-agent system for solving complex tasks.arXiv preprint arXiv:2411.04468, 2024

    Adam Fourney, Gagan Bansal, Hussein Mozannar, Cheng Tan, Eduardo Salinas, Friederike Niedtner, Grace Proebsting, Griffin Bassman, Jack Gerrits, Jacob Alber, et al. Magentic-one: A generalist multi-agent system for solving complex tasks.arXiv preprint arXiv:2411.04468, 2024

    Pith/arXiv arXiv 2024

  14. [14]

    Adaplanner: Adaptive planning from feedback with language models.Advances in neural information processing systems, 36:58202–58245, 2023

    Haotian Sun, Yuchen Zhuang, Lingkai Kong, Bo Dai, and Chao Zhang. Adaplanner: Adaptive planning from feedback with language models.Advances in neural information processing systems, 36:58202–58245, 2023

    2023

  15. [15]

    Oagents: An empirical study of building effective agents.arXiv preprint arXiv:2506.15741, 2025

    He Zhu, Tianrui Qin, King Zhu, Heyuan Huang, Yeyi Guan, Jinxiang Xia, Yi Yao, Hanhao Li, Ningning Wang, Pai Liu, et al. Oagents: An empirical study of building effective agents.arXiv preprint arXiv:2506.15741, 2025

    Pith/arXiv arXiv 2025

  16. [16]

    Resum: Unlocking long-horizon search intelligence via context summarization.arXiv preprint arXiv:2509.13313, 2025

    Xixi Wu, Kuan Li, Yida Zhao, Liwen Zhang, Litu Ou, Huifeng Yin, Zhongwang Zhang, Yong Jiang, Pengjun Xie, Fei Huang, et al. Resum: Unlocking long-horizon search intelligence via context summarization.arXiv preprint arXiv:2509.13313, 2025

    arXiv 2025

  17. [17]

    Language agent tree search unifies reasoning acting and planning in language models.arXiv preprint arXiv:2310.04406, 2023

    Andy Zhou, Kai Yan, Michal Shlapentokh-Rothman, Haohan Wang, and Yu-Xiong Wang. Language agent tree search unifies reasoning acting and planning in language models.arXiv preprint arXiv:2310.04406, 2023

    Pith/arXiv arXiv 2023

  18. [18]

    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

    arXiv 2025

  19. [19]

    Kto: Model alignment as prospect theoretic optimization.arXiv preprint arXiv:2402.01306, 2024

    Kawin Ethayarajh, Winnie Xu, Niklas Muennighoff, Dan Jurafsky, and Douwe Kiela. Kto: Model alignment as prospect theoretic optimization.arXiv preprint arXiv:2402.01306, 2024

    Pith/arXiv arXiv 2024

  20. [20]

    Lora: Low-rank adaptation of large language models.arXiv preprint arXiv:2106.09685, 2022

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models.arXiv preprint arXiv:2106.09685, 2022

    Pith/arXiv arXiv 2022

  21. [21]

    Fact, fetch, and reason: A unified evaluation of retrieval-augmented generation.CoRR, 2024

    Satyapriya Krishna, Kalpesh Krishna, Anhad Mohananey, Steven Schwarcz, Adam Stambler, Shyam Upadhyay, and Manaal Faruqui. Fact, fetch, and reason: A unified evaluation of retrieval-augmented generation.CoRR, 2024. 12

    2024

  22. [22]

    xbench: Tracking agents productivity scaling with profession-aligned real- world evaluations.arXiv preprint arXiv:2506.13651, 2025

    Kaiyuan Chen, Yixin Ren, Yang Liu, Xiaobo Hu, Haotong Tian, Tianbao Xie, Fangfu Liu, Haoye Zhang, Hongzhang Liu, Yuan Gong, et al. xbench: Tracking agents productivity scaling with profession-aligned real- world evaluations.arXiv preprint arXiv:2506.13651, 2025

    Pith/arXiv arXiv 2025

  23. [23]

    Gaia: a benchmark for general ai assistants

    Gr´egoire Mialon, Cl´ementine Fourrier, Thomas Wolf, Yann LeCun, and Thomas Scialom. Gaia: a benchmark for general ai assistants. InThe Twelfth International Conference on Learning Representations, 2023

    2023

  24. [24]

    Swe-bench: Can language models resolve real-world github issues?arXiv preprint arXiv:2310.06770, 2023

    Carlos E Jimenez, John Yang, Alexander Wettig, Shunyu Yao, Kexin Pei, Ofir Press, and Karthik Narasimhan. Swe-bench: Can language models resolve real-world github issues?arXiv preprint arXiv:2310.06770, 2023

    Pith/arXiv arXiv 2023

  25. [25]

    Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts.arXiv preprint arXiv:2310.02255, 2023

    Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts.arXiv preprint arXiv:2310.02255, 2023

    Pith/arXiv arXiv 2023

  26. [26]

    Co-sight: Enhancing llm-based agents via conflict-aware meta-verification and trustworthy reasoning with structured facts.arXiv preprint arXiv:2510.21557, 2025

    Hongwei Zhang, Ji Lu, Shiqing Jiang, Chenxiang Zhu, Li Xie, Chen Zhong, Haoran Chen, Yurui Zhu, Yongsheng Du, Yanqin Gao, et al. Co-sight: Enhancing llm-based agents via conflict-aware meta-verification and trustworthy reasoning with structured facts.arXiv preprint arXiv:2510.21557, 2025

    arXiv 2025

  27. [27]

    Swe-search: Enhancing software agents with monte carlo tree search and iterative refinement.arXiv preprint arXiv:2410.20285, 2024

    Antonis Antoniades, Albert ¨Orwall, Kexun Zhang, Yuxi Xie, Anirudh Goyal, and William Wang. Swe-search: Enhancing software agents with monte carlo tree search and iterative refinement.arXiv preprint arXiv:2410.20285, 2024

    Pith/arXiv arXiv 2024

  28. [28]

    The openhands software agent sdk: A composable and extensible foundation for production agents.arXiv preprint arXiv:2511.03690, 2025

    Xingyao Wang, Simon Rosenberg, Juan Michelini, Calvin Smith, Hoang Tran, Engel Nyst, Rohit Malhotra, Xuhui Zhou, Valerie Chen, Robert Brennan, et al. The openhands software agent sdk: A composable and extensible foundation for production agents.arXiv preprint arXiv:2511.03690, 2025

    Pith/arXiv arXiv 2025

  29. [29]

    Search-o1: Agentic search-enhanced large reasoning models.arXiv preprint arXiv:2501.05366, 2025

    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.arXiv preprint arXiv:2501.05366, 2025

    Pith/arXiv arXiv 2025

  30. [30]

    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

  31. [31]

    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

  32. [32]

    reconfig

    Jiahao Qiu, Xuan Qi, Tongcheng Zhang, Xinzhe Juan, Jiacheng Guo, Yifu Lu, Yimin Wang, Zixin Yao, Qihan Ren, Xun Jiang, et al. Alita: Generalist agent enabling scalable agentic reasoning with minimal predefinition and maximal self-evolution.arXiv preprint arXiv:2505.20286, 2025. 13 A Appendix A.1 Additional Experimental Results Table 5: Unified action spac...

    Pith/arXiv arXiv 2025

  33. [33]

    Adaptive Toolbox Management.The agent dynamically adjusts its available tools based on current needs. Case 2 Stage 5 demonstrates significant toolbox pruning: once data collection is complete, the agent removes visit and search tools, retaining only code interpreter and execute bash for Haversine distance calculation. This focused toolset reduces action s...

  34. [34]

    Autonomous Strategy Adaptation.When initial strategies fail, agents can pivot without external intervention. In Case 2, when the ASEAN Wikipedia page lacks capital information (Stage 1-2), the agent autonomously decides to visit individual country pages (Stage 3)—a major strategic pivot recorded in the reconfiguration request and propagated through inter-...

  35. [35]

    doing” and “managing

    Self-Determined Task Completion.Agents decide when to terminate based on their own assessment. Both cases end with the agent invoking Tterm after verifying answer correctness, rather than relying on external completion signals or fixed iteration limits. These capabilities emerge from TOOLSELF’s unified action space design, where task execution tools and m...

    2006

  36. [36]

    visit (Web Access Tool).The tool definition prompt is shown in Sec. A.6.4. Implementation follows WebSailor [18], using the Jina Reader API service to access and parse webpage content, supporting single or batch URL access. This tool performs intelligent content extraction through an LLM, automatically locating and extracting relevant information and evid...

  37. [37]

    search.The tool definition prompt is shown in Sec. A.6.4. Implementation follows WebSailor [ 18], supporting batch query submission, automatic language detection, and concurrent request handling, returning formatted search results including title, link, summary, publication date, and source domain. We adopt differentiated search backends tailored to bench...

  38. [38]

    code interpreter (Python Code Interpreter).The tool definition prompt is shown in Sec. A.6.4. It executes Python code in an isolated Jupyter Kernel environment, with each agent instance maintaining an independent kernel state to support cross-call variable persistence and state retention. Pre-installed libraries include NumPy, Pandas, Matplotlib, SciPy, a...

  39. [39]

    execute bash.The tool definition prompt is shown in Sec. A.6.4. Implementation follows SWE Agent [ 4], executing Bash commands in an isolated Shell session, supporting file system operations (create, read, modify, delete files or directories), script execution, system queries (e.g., git version control operations, environment variable checks). It incorpor...

  40. [40]

    str replace editor.The tool definition prompt is shown in Sec. A.6.4. Implementation follows SWE Agent [ 4], an advanced file editing tool designed specifically for software engineering tasks. It supports five core operations: view for viewing file contents or specified line ranges, create for creating new files, str replace for content replacement based ...

  41. [41]

    file analyzer (File Analysis Tool).The tool definition prompt is shown in Sec. A.6.4. An intelligent tool designed specifically for multimodal file analysis, supporting two major file categories: text documents and images. For text documents (PDF, Word, Excel, PowerPoint, TXT, Markdown, etc.), it employs a document parser for intelligent chunking, using t...

  42. [42]

    <main_task> {MAIN_TASK_CONTENT} </main_task>

    **Main Task ‘<main_task>‘**: This is the ultimate goal to be completed. <main_task> {MAIN_TASK_CONTENT} </main_task>

  43. [43]

    <sub_goal> {SUB_GOAL_CONTENT} </sub_goal>

    **Current Sub_goal ‘<sub_goal>‘**: This is the specific task you need to focus on completing now. <sub_goal> {SUB_GOAL_CONTENT} </sub_goal>

  44. [44]

    <strategy> {EXECUTION_STRATEGY} </strategy>

    **Execution Strategy ‘<strategy>‘**: The execution methodology for the current sub_goal. <strategy> {EXECUTION_STRATEGY} </strategy>

  45. [45]

    <toolbox> {TOOLBOX_LIST} </toolbox>

    **Toolbox ‘<toolbox>‘**: This is the list of tools available for the current sub_goal. <toolbox> {TOOLBOX_LIST} </toolbox>

  46. [46]

    **Knowledge Information ‘<knowledge>‘**: This is the background information and constraints related to the task. <knowledge> {KNOWLEDGE_CONTENT} </knowledge> ## Tool Usage Guidelines - **Precision**: Select the most appropriate tools based on the current sub-goal - **Efficiency**: Prioritize tools that can directly solve the problem - **Completeness**: En...

  47. [47]

    DO NOT attempt to complete the entire main task

    **CRITICAL - Focus ONLY on Sub_goal**: Your ONLY objective is to complete the current sub_goal. DO NOT attempt to complete the entire main task

  48. [48]

    **Main Task Usage**: Use the main task information ONLY to understand the broader context and make informed decisions about the current sub_goal

  49. [49]

    **Tool Selection**: Choose appropriate tools from the toolbox to complete the sub_goal

  50. [50]

    Use the reconfiguration tool when needing to update agent configuration

    **Result Reporting**: Use the termination tool only when the MAIN TASK is completely finished. Use the reconfiguration tool when needing to update agent configuration

  51. [51]

    A conversation between User and Assistant. The user asks a question, and the assistant solves it by calling one or more of the following tools

    **Error Handling**: When encountering problems, use the reconfiguration tool to modify agent configuration rather than simply reporting errors. 24 A.6.2 ReAct Execution Format ReAct Format Specification The inference agent operates within a structured prompt format: USER_PROMPT_PREFIX: "A conversation between User and Assistant. The user asks a question, ...

  52. [52]

    Mentally plan the overall task execution to understand the big picture

  53. [53]

    Identify the immediate next step based on current progress

  54. [54]

    Generate appropriate configuration (sub_goal, strategy, toolbox, knowledge) ## Input Information You will receive the following parts:

  55. [55]

    <main_task> {MAIN_TASK_CONTENT} </main_task>

    **Main Task ‘<main_task>‘**: The final goal the Agent needs to accomplish. <main_task> {MAIN_TASK_CONTENT} </main_task>

  56. [56]

    <available_tools> {ALL_AVAILABLE_TOOLS} </available_tools>

    **Available Tools ‘<available_tools>‘**: Complete list of tools available to the Agent during the entire task lifecycle. <available_tools> {ALL_AVAILABLE_TOOLS} </available_tools>

  57. [57]

    Can be empty ("NONE")

    **Execution History ‘<execution_history>‘**: Summarized record of all previous steps. Can be empty ("NONE"). <execution_history> {EXECUTION_HISTORY} </execution_history>

  58. [58]

    Can be empty ("NONE")

    **Update Requirement ‘<update_requirement>‘**: Update suggestions from previous Agent. Can be empty ("NONE"). <update_requirement> {UPDATE_REQUIREMENT} </update_requirement> ## Configuration Generation Process ### Step 1: Determine Next Sub-Goal Carefully consider the proposed sub-goal in ‘<update_requirement>‘ and adopt it whenever possible. Only re-plan...

  59. [59]

    Senior Software Engineer

    **Define Persona**: Assign an expert persona (e.g., "Senior Software Engineer", "Data Analyst")

  60. [60]

    **Create a Step-by-Step Plan**: Simulate the thinking process to accomplish the next sub_goal

  61. [61]

    **Identify Decision Points**: Mark key decision points in the plan

  62. [62]

    **2.2 Toolbox Selection** Select tools for the **toolbox**: - **CRITICAL**: The toolbox must include at least three tools specifically needed for the NEXT sub_goal ONLY

    **Link to Tools**: Specify which concrete tool should be used in each step. **2.2 Toolbox Selection** Select tools for the **toolbox**: - **CRITICAL**: The toolbox must include at least three tools specifically needed for the NEXT sub_goal ONLY. - **VERY IMPORTANT**: Focus EXCLUSIVELY on tools required to complete the next sub_goal. - **IMPORTANT**: The t...

  63. [63]

    **Initial Step**: If ‘<execution_history>‘ is ‘NONE‘ or empty, ‘inter_agent_knowledge‘ must be an empty string ‘""‘

  64. [64]

    **Use ALL**: If history is concise and relevant, use the string ‘"ALL"‘

  65. [65]

    next_sub_goal

    **Summarize**: If history is long or contains exploratory steps, extract a concise summary. ## Output Format Your final output **must** be a strict JSON object: { "next_sub_goal": "Detailed description of the next step to execute", "execution_strategy": "As a [expert persona], I will follow these steps to accomplish the NEXT sub_goal: 1. [first step and tool]

  66. [66]

    , "toolbox

    [second step and tool] 3. [key decision points]...", "toolbox": ["ToolName1", "ToolName2", "ToolName3"], "inter_agent_knowledge": "One of three forms: summary, ’ALL’, or ’’" } 26 A.6.4 Tool Definitions The reconfiguration tool arguments correspond to the formal definitions in Sec. 2 as follows:new sub goal corresponds to qprop i+1, update reason to ρi, an...