{"id":"e2f2f2f2-70b3-4937-887f-6c57f7e2e371","arxiv_id":"2602.07883","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"An LLM agent that can call a reconfiguration tool to update its sub-goals, toolbox, strategy, and context outperforms static-config agents across FRAMES, xbench, GAIA, and SWE-bench Lite.","lead":"This paper proposes ToolSelf, an LLM agent design that lets the agent call a \"reconfigure\" tool to update its own sub-goals, tools, strategy, and context during a task. On four agent benchmarks, it reports consistent gains over static-configuration agents and several specialized baselines.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"CAT gains not isolated from training signal: no trained no-reconfiguration control","rationale":"The reader's weakest assumption focused on the agent's self-report quality and the heuristic context-cleanup, but our review identifies a more fundamental issue: the empirical comparison does not isolate the effect of the action space from the effect of training. The paper's strongest claims are about the trained system (Table 2 and Sec. 3.3), yet all non-TOOLSELF baselines are essentially zero-shot or use their own specialized training (e.g., WebSailor-7B), so a large portion of the gain could come from the RFT/KTO training signal, not from self-reconfiguration. This is particularly salient because the reconfiguration engine is instructed to 'adopt the proposed sub-goal whenever possible', making the system's planning behavior largely dependent on the inference agent's suggestions, which are themselves improved by training. The proposed control experiment—training a vanilla agent with the same CAT pipeline—is a direct, feasible test of whether the unified action space is the causal driver. If the control performs comparably, the central claim would need to be substantially weakened, potentially changing the verdict to REJECT; if the control performs poorly, the claim stands. Until this experiment is run, the paper's central causal claim is not fully supported, and the reader's CONDITIONAL verdict should be updated to include this specific experimental condition.","tokens_in":26195,"tokens_out":6015,"duration_ms":67520,"concrete_test":"Train a Vanilla Agent (standard ReAct, no reconfiguration tool) with the identical CAT pipeline: same base model (Qwen3-8B), same RFT data from the 300 FRAMES/MathVista tasks, same KTO sampling (4 trajectories per task, trajectory-level binary labels), and same LoRA hyperparameters. Use the same tool pool, context-management heuristics, and teacher trajectories (adapted for a static configuration). Evaluate on the same GAIA, FRAMES, and xbench test splits. If Vanilla+RFT+KTO reaches or approaches TOOLSELF+RFT+KTO's average (e.g., within 5 points), the central claim that the unified action space drives the improvement is undermined. If it remains far below (e.g., <30 points), the unified action space is validated as a key causal factor.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that unifying reconfiguration into the agent's action space causes the large gains (e.g., +24.1 points in Table 2). However, the comparison is confounded by training: TOOLSELF+RFT+KTO is trained on teacher trajectories and KTO reinforcement, while most baselines (Vanilla Agent, ReSum, OAgents, OWL, Co-Sight) are used zero-shot. The ablation in Sec. 3.4 only controls for compute (Vanilla + majority voting) but never for training signal. A vanilla ReAct agent trained on the same RFT data and trajectory-level KTO could plausibly achieve a large fraction of the observed gain, since it would benefit from the same high-quality teacher demonstrations and binary reward shaping. The paper's own prompt for the reconfiguration engine (Sec. A.6.3) instructs it to 'adopt the proposed sub-goal whenever possible', so the reconfiguration engine is largely a pass-through for the agent's proposal, meaning the 'unified action space' contribution may be mostly about giving the agent a way to emit structured sub-goals. Without a trained no-reconfiguration control, the causal role of the action space itself is not established.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":26456,"tokens_out":6647,"duration_ms":66253,"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":[{"comment":"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.","section":"Sec. 3.2 / Table 1"},{"comment":"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.","section":"Sec. 3.3, Table 2 / Sec. 3.4, Table 5"},{"comment":"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.","section":"All tables / Sec. A.5.7"},{"comment":"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.","section":"Abstract / Sec. 3.3 / Conclusion"},{"comment":"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.","section":"Sec. A.5.1 / Sec. A.6.3 / Sec. 2"}],"minor_comments":[{"comment":"The introduction refers to 'Prior work (Appendix 4)' but the related work appears in Section 4, not an appendix.","section":"Sec. 1"},{"comment":"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.","section":"Table 5 vs Table 1"},{"comment":"The KTO loss uses r(x), z0, v(·), and λD, λU without defining them in the main text. Please add definitions or a reference.","section":"Eq. (11)"},{"comment":"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.","section":"Sec. A.5.7"},{"comment":"The Level 3 row reads '7.693.85' without a spacing separator between 7.69 and 3.85. This is a formatting typo.","section":"Table 3"},{"comment":"No code, data, or model release is mentioned. Given the detail of the training setup, a reproducibility statement would strengthen the paper.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The central idea is interesting and the appendix is unusually thorough, but the paper overclaims relative to its own Table 1, and the main training gain is not isolated from the effect of RFT+KTO alone. Both issues are fixable with additional experiments or appropriately weakened claims, so I do not recommend rejection. However, the current version would not support an accept."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a read if you work on LLM agents for long-horizon tasks. The central idea—treating configuration updates as a callable tool in the same action space as task execution—is a real conceptual step, and the two-stage training (RFT cold-start, then trajectory-level KTO) is a sensible and non-obvious addition. The architecture is described concretely enough that someone could re-implement it, and the ablations in the appendix (removing sub-goals, context management, toolbox, strategy) give useful signal about which components carry the weight. The case studies are also genuinely informative: they show the agent decomposing a hard GAIA question into stages, pruning its toolbox at the right moment, and recovering from a failed strategy. Credit where it's due: this is a coherent, well-motivated piece of engineering work, not a repackaging of existing pieces.\n\nThe soft spots are mostly about evidence. The abstract says a 28.8-point gain, the body says 24.1; that's a factual mismatch that needs fixing. More importantly, the central claim that the unified action space is what drives the improvement is not cleanly tested. The trained TOOLSELF model is compared against zero-shot static baselines and against compute-matched majority voting, but never against a vanilla ReAct agent trained on the same RFT data and KTO reward. That control is essential: without it, a large fraction of the gain could come from the training signal itself rather than from the reconfiguration tool. The stress-test concern is valid. The appendix prompt also tells the reconfiguration engine to 'adopt the proposed sub-goal whenever possible,' which makes the engine a near pass-through; that further weakens the claim that the unification itself is the engine of improvement. And the paper relies on single-run pass@1 results with no error bars or seed variance, so we don't know how robust the margins are. The heuristic context-cleanup at 80% token capacity is a minor contradiction to 'autonomous triggering,' but I'd call that a qualification, not a fatal flaw.\n\nIf I were the editor, I'd send this to peer review—the architecture is worth discussing and the authors clearly know what they're doing. But I'd ask them to train a no-reconfiguration control, fix the abstract/body inconsistency, and report variance. The math is not the issue; there isn't any load-bearing derivation. The paper is a solid, reproducible-in-principle contribution that currently overclaims its main result. A serious referee would catch the missing control, and the paper would be stronger for it.","headline":"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.","tokens_in":786,"tokens_out":1116,"would_cite":true,"duration_ms":25835,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"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.","keywords":["LLM agents","self-reconfiguration","tool-driven adaptation","long-horizon tasks","rejection sampling fine-tuning","KTO reinforcement learning","agent configuration","runtime adaptation"],"falsifier":"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.","tokens_in":26078,"feed_emoji":"🤖","tokens_out":5399,"duration_ms":49699,"temperature":0.7,"pith_summary":"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.","feed_headline":"One agent rewrites its own plan and beats static agents by 24 points","feed_subtitle":"Turning configuration updates into a tool call lets one generalist match specialized workflows on long-horizon tasks.","key_machinery":"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.","core_discovery":"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","pith_inferences":["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."],"forward_implications":["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."],"fun_headline_variants":["Agent self-reconfigures via tool calls, beating static configs by 28.8 pts","Execution and self-reconfiguration unified into one policy's action space","Tool-driven adaptation lets agent update its own sub-goals and toolbox","No more static agent: reconfiguration becomes a tool action","Emergent adaptivity: agent rewrites its plan without manual guidance"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Agent self-reconfigures via tool calls, beating static configs by 28.8 pts","Execution and self-reconfiguration unified into one policy's action space","Tool-driven adaptation lets agent update its own sub-goals and toolbox","No more static agent: reconfiguration becomes a tool action","Emergent adaptivity: agent rewrites its plan without manual guidance"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000421,"raw_usage":{"total_tokens":2024,"prompt_tokens":792,"completion_tokens":1232,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":536,"completion_tokens_details":{"reasoning_tokens":1138}},"tokens_in":536,"tokens_out":1232,"duration_ms":12202,"temperature":1.0,"reasoning_tokens":1138,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T03:25:40.034550+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[],"review_version":1}