{"id":"d81fbfa2-74cb-4fe9-a9dc-d2bc92e933e2","arxiv_id":"2609.17523","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"ScienceBuddy demonstrates that coupling harness evolution with model reinforcement learning in a nested loop can improve scientific task performance over multiple cycles.","lead":"ScienceBuddy is an interactive scientific workspace that continually improves its AI agent through researcher collaboration. It introduces a recursive-in-recursive self-improvement paradigm where the agent's instructions and skills are refined separately from the model weights, enabling sustained improvement across cycles.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"The three-cycle experiment lacks baselines that ablate the recursive coupling; observed improvement could be driven by harness adaptation alone or model learning alone, not by their nested interaction.","rationale":"The reader’s weakest assumption was the use of simulated user feedback, which is a separate limitation about external validity. My concern is about internal validity: the experiment does not isolate the effect of the nested recursion. Both are legitimate concerns, but mine targets the core claim that the coupling sustains improvement. Since the reader’s verdict was already CONDITIONAL and flagged the need for more realistic validation, adding this methodological point does not change the verdict—it strengthens the conditions under which the claim would be accepted. I recommend UNCHANGED because the concern does not fundamentally contradict the reader’s assessment but rather adds specificity about what additional evidence is needed.","tokens_in":22020,"tokens_out":3063,"duration_ms":32990,"concrete_test":"Run a matched ablation: starting from the same initial (θ₀, H₀), execute three cycles of harness-only adaptation (inner recursion, 10 steps per cycle) without any model RL, and separately three cycles of model-only RL (20 updates per cycle) with a fixed harness. Evaluate both baselines on the same held-out test set used in §4.2. Compare their final accuracy to the full system’s 73.3%. If either baseline matches or exceeds that figure, the recursive-in-recursive coupling is not the cause of the improvement.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central novelty of ScienceBuddy is the recursive-in-recursive coupling of harness evolution and model RL. The headline three-cycle experiment (§4.2) runs both components together: each cycle does 10 harness-evolution steps followed by 20 RL updates. But the paper never compares this full system to a baseline that runs only harness adaptation across multiple cycles (without model updates) or only model RL across multiple cycles (with a fixed harness). The standalone harness-only (§4.3) and model-only (§4.4) experiments each show improvement, but they are conducted on different task sets and evaluation protocols (first-response accuracy vs. pass@4) and are not repeated over multiple cycles. Thus we cannot determine whether the coupling is necessary or whether a simpler repeated application of either component alone would yield comparable or better gains. If, for example, three cycles of harness-only adaptation achieve the same 73.3% test accuracy, the recursive-in-recursive claim would lose force. The paper acknowledges that the reflector remains fixed (§5, p.17: “improved task performance does not imply that the improvement mechanism itself has become stronger”), which limits the strength of self-improvement, but does not address this more fundamental ablation gap. This is load-bearing because the entire paradigm is defined by the nested recursion; without an ablation control, the evidence that the coupling sustains improvement across cycles is inconclusive.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"The paper introduces ScienceBuddy, an interactive scientific research workspace that implements recursive-in-recursive self-improvement: an inner loop refines the agent's harness (instructions, skills) through evaluated edits while the model is fixed, and an outer loop performs reinforcement learning on the model using rubric rewards under the improved harness. The system is released as a product. Case studies include real researcher interactions, a three-cycle coupled experiment (Qwen3.5-4B, 895 tasks) showing test accuracy improvement from 42.2% to 73.3%, a harness-only experiment (fixed model) showing 31.1% to 51.1% validation accuracy, and a model-only experiment (fixed harness) showing pass@4 coverage from 48.3% to 67.8%.","tokens_in":22398,"tokens_out":1871,"duration_ms":19676,"significance":"If the central claim holds, the recursive-in-recursive paradigm could provide a principled mechanism for continual improvement of scientific AI agents through routine researcher collaboration. The coupling of procedural editing with model-level RL is a natural direction, and the release of a functional workspace with 224 tools across 22 modules could enable community experiments. The paper also demonstrates measurable gains on established benchmarks (LAB-Bench, Biomni-Eval1) with a small 4B model, which is encouraging for cost-sensitive deployment. However, the significance is currently limited by the reliance on simulated user feedback and the absence of an ablation that isolates the contribution of the recursive coupling itself.","major_comments":[{"comment":"The three-cycle experiment runs harness evolution and model RL together but never compares against a baseline that applies only harness adaptation across multiple cycles or only model RL across multiple cycles. The standalone experiments in §4.3 and §4.4 are conducted on different task sets and evaluation protocols (first-response accuracy vs. pass@4) and are not repeated over multiple cycles, so they cannot serve as multi-cycle ablations. The central novelty of the paper is the recursive-in-recursive coupling; without a controlled ablation, the evidence that the coupling sustains improvement beyond either component alone is inconclusive. The authors should either provide multi-cycle baselines for each component separately, or substantially weaken the claim to describe the full system as one plausible instantiation rather than a demonstrated necessity.","section":"§4.2, Fig. 8"},{"comment":"The harness evolution in both the coupled experiment (§4.2) and the harness-only experiment (§4.3) relies on simulated user feedback generated by a bounded user simulator (Qwen3.8-27B helper with a fixed reply set). The paper acknowledges this limitation but does not discuss how real researcher feedback might differ in distribution, noise, or informativeness, nor whether the bounded simulator biases the harness toward a narrow set of procedural patterns. A sensitivity analysis or an informal comparison with real interaction trajectories (e.g., from §4.1) would help assess how representative the simulated feedback is.","section":"S1.2 (simulated user feedback)"},{"comment":"The paper states that 'the reflector remains fixed, so improved task performance does not imply that the improvement mechanism itself has become stronger.' This is an honest limitation, but it undercuts the term 'self-improvement.' The nested recursion is the paper's headline contribution; the reader should be told more explicitly what kind of self-improvement is and is not being demonstrated. Future work or discussion should address whether the fixed reflector is a temporary engineering choice or a principled limitation of the paradigm.","section":"§5 (Related Work, p. 17)"}],"minor_comments":[{"comment":"The abstract uses 'continually improving scientific agents' but does not clarify that the improvement is in task performance, not in the improvement mechanism itself. Rephrasing to 'continually improving scientific task performance' would prevent over-interpretation.","section":"Abstract"},{"comment":"ConstructRubric is called but not fully defined in the main text; the function relies on the full collaboration trajectory, reconstructed instruction, and assets. A brief note on what happens when the rubric derivation fails or produces conflicting criteria would improve clarity.","section":"§2.2, Eq. (2)"},{"comment":"The caption refers to 'three RinR cycles' but the figure legend uses stages 1, 2, 3. Align terminology (e.g., cycle 1, 2, 3) for consistency.","section":"§4.2, Fig. 8 caption"},{"comment":"The information-boundary table is helpful. The row for 'Scientific verifier' says it 'Requires output' – this should read 'Required output' for parallel phrasing with other rows.","section":"§S1.2, Table S2"},{"comment":"The group-relative advantage definition (Eq. S1) uses std denominator with a small 𝛿. It would be helpful to state the value of 𝛿 used in the experiments.","section":"§S1.3, Eq. (S1)–(S3)"}],"recommendation":"major_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"This paper introduces a novel recursive-in-recursive framework for scientific agent self-improvement, coupling harness evolution with model RL. The system is well-described, the code is released, and the case studies show consistent improvement on held-out tasks (42.2% to 73.3% over three cycles). The standalone harness and model experiments each demonstrate that both components can improve individually.\n\nThe soft spot is load-bearing: the central novelty is the nested coupling, but the paper never compares the full system to repeated application of either component alone. The three-cycle experiment couples both, while the harness-only and model-only experiments use different tasks and metrics (first-response accuracy vs. pass@4) and are not repeated over multiple cycles. We cannot tell whether the coupling is necessary or whether simply running harness evolution or model RL alone over multiple cycles would match or exceed those gains. This weakness is explicit in the stress test and, on reading the paper, I agree it is not addressed.\n\nThe paper is transparent about other limitations (simulated user feedback, small 4B model, few cycles, fixed reflector), and those are honestly stated. But the key empirical claim that the nested recursion sustains improvement lacks the direct ablation control it needs. The idea is interesting and the infrastructure is real, but the evidence for the coupling is circumstantial.\n\nFor a researcher interested in agent architectures or continual learning, this paper provides a clear description and a practical system. But they should read it with a skeptical eye on the central claim. I would recommend conditional acceptance with a request for an ablation study that compares the full recursive system to at least one of the component-only baselines over matched cycles.","headline":"Interesting architecture and solid engineering, but the central claim about nested recursion is undersupported by missing ablation controls.","tokens_in":23431,"tokens_out":611,"duration_ms":21352},"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek/deepseek-v4-flash","headline":"ScienceBuddy introduces recursive-in-recursive self-improvement that couples harness evolution with model reinforcement learning, enabling a scientific agent to improve its task accuracy from 42.2% to 73.3% over three cycles of researcher…","keywords":["scientific AI agent","recursive self-improvement","interactive learning","harness evolution","rubric reward","GRPO","continual learning","science workspace"],"falsifier":"Run an identical three-cycle experiment where real researchers provide feedback instead of the bounded simulator. If accuracy gains shrink significantly or regress, the method's dependence on simulated feedback would limit its real-world utility.","tokens_in":23620,"feed_emoji":"🧪","tokens_out":4591,"duration_ms":44731,"temperature":0.7,"pith_summary":"ScienceBuddy is an interactive scientific workspace where a language-model agent improves by alternating two nested loops. In the inner loop, the agent's instructions and skills (the harness) are revised while the model remains frozen. In the outer loop, the model is updated via reinforcement learning under the improved harness, using task-specific rubric scores as rewards. Across three cycles of simulated researcher interaction, the agent's single-attempt test accuracy on four scientific task families rose from 42.2% to 73.3%, with only 2.2% regression. The paper argues this coupling of procedural and policy learning can sustain improvement as the system continues to interact with researchers.","feed_headline":"Recursive-in-recursive learning boosts agent accuracy by 31 points","feed_subtitle":"A three-cycle experiment on four scientific task families shows sustained gains with minimal regression, suggesting open-ended improvement.","key_machinery":"Recursive-in-recursive self-improvement: an inner recursion that keeps the task model fixed while a separate auxiliary model diagnoses execution failures and proposes bounded edits to the harness (instruction entries, reusable skills, and context settings), accepting candidates only when they improve a paired development evaluation; and an outer recursion that augments the task environment, collects fresh on-policy rollouts under the selected harness, and trains the model via group-relative policy optimization (GRPO) with rubric-based trajectory rewards.","core_discovery":"The central claim is that a scientific AI agent can continuously improve both its working procedures and its underlying capabilities by coordinating harness evolution and model reinforcement learning. The harness is refined by a fixed auxiliary model that diagnoses failures and proposes bounded edits to instructions and skills; the updated harness then guides fresh rollouts whose rubric rewards drive GRPO-based model training. The case studies show that harness adaptation alone raises first-response accuracy by 20 percentage points at fixed model weights, and model learning alone increases problem coverage by 19.5 percentage points under a fixed harness. The coupled three-cycle experiment demonstrates overall accuracy gains of 31.1 percentage points, suggesting that the two improvement mechanisms complement each other and do not plateau after a single round.","pith_inferences":["If the method transfers to real researcher interactions (not just simulators), it could turn routine scientific chat into a continuous training signal, reducing the need for manual data curation.","The recursive-in-recursive scheme effectively uses the agent's history as a self-generated curriculum for both procedures and policy, implying a general architecture for open-ended agent learning.","The fixed auxiliary model that proposes harness edits is a bottleneck; replacing it with a learned editor could make the inner loop itself improvable, though the paper explicitly avoids this to isolate the coupling.","Because rubrics are derived from collaboration but fixed during RL, the quality of rubric construction directly bounds the reward signal; poorly designed criteria could misguide training or miss meaningful improvements."],"forward_implications":["Scientific AI agents can improve without full retraining by adapting how they present instructions and use tools, enabling cost-effective capability growth.","The methodology provides a template for building interactive agents in other domains that learn from user feedback, provided rubrics can be derived from collaboration.","The three-cycle experiment shows sustained improvement without saturation, suggesting potential for indefinite self-improvement as researcher interactions accumulate.","Problem-coverage gains (pass@4 from 48.3% to 67.8%) indicate that model learning can broaden the set of solvable tasks, not just sharpen accuracy on known ones.","The coupling of harness and model updates means that improvements in one component create new opportunities for the other, producing a compounding effect."],"supporting_citations":[{"why":"provides the scientific tool catalog, execution environments, and data lake used by ScienceBuddy.","marker":"[9]"},{"why":"supplies the LAB-Bench and Biomni-Eval1 benchmark tasks that define the four scientific task families.","marker":"[10]"},{"why":"establishes the GRPO objective used in the outer recursion for model reinforcement learning.","marker":"[14]"},{"why":"introduces rubric-based rewards that serve as the trajectory-level reward signal in the outer loop.","marker":"[7]"},{"why":"develops the Meta-Harness concept for end-to-end harness optimization, which the inner recursion builds on.","marker":"[11]"},{"why":"provides the Reflexion paradigm for persistent procedural adaptation through verbal reflection.","marker":"[2]"},{"why":"proposes model-harness co-evolution (HELIX), a direct precedent for the joint adaptation studied here.","marker":"[5]"}],"fun_headline_variants":["Recursive self-improvement lifts scientific agent accuracy 31 points","Harness evolution and model RL combine for 31-point gain","ScienceBuddy: two-loop learning boosts accuracy 31 points","Inner recursion refines harness, outer recursion trains model","Open-ended agent improvement through coupled recursion cycles"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The harness-evolution case study relies on simulated user feedback generated by a bounded user simulator, not on real researcher interactions; if simulated feedback differs substantially from real behavior, the harness improvements may not transfer to actual scientific collaboration.","fun_headline_variants_meta":{"raw":{"variants":["Recursive self-improvement lifts scientific agent accuracy 31 points","Harness evolution and model RL combine for 31-point gain","ScienceBuddy: two-loop learning boosts accuracy 31 points","Inner recursion refines harness, outer recursion trains model","Open-ended agent improvement through coupled recursion cycles"]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":0.000256,"raw_usage":{"cost":0.000255808,"is_byok":false,"cost_details":{"upstream_inference_cost":0.000255808,"upstream_inference_prompt_cost":9.1728e-05,"upstream_inference_completions_cost":0.00016408},"total_tokens":1446,"prompt_tokens":860,"completion_tokens":586,"prompt_tokens_details":{"audio_tokens":0,"video_tokens":0,"cached_tokens":256,"cache_write_tokens":0},"completion_tokens_details":{"audio_tokens":0,"image_tokens":0,"reasoning_tokens":519}},"tokens_in":860,"tokens_out":586,"cost_source":"openrouter_usage","duration_ms":6871,"image_pages":[],"temperature":1.0,"reasoning_tokens":519,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-09-17T20:10:18.053468+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-medium","input_kind":null,"input_chars":null,"input_total_chars":null,"passes_missing":null,"published_standing":null,"published_confidence":null,"objection_audit":null,"run_record":null,"basis":{"kind":"not_recorded","source":null,"chars_given":null,"chars_total":null},"receipt":null,"falsifier":"Run an identical three-cycle experiment where real researchers provide feedback instead of the bounded simulator. If accuracy gains shrink significantly or regress, the method's dependence on simulated feedback would limit its real-world utility.","supporting_citations":[{"cited_title":"Biomni: A General-Purpose Biomedical AI Agent","cited_arxiv_id":null,"evidence_quote":"provides the scientific tool catalog, execution environments, and data lake used by ScienceBuddy."},{"cited_title":"LAB-Bench: Measuring Capabilities of Language Models for Biology Research","cited_arxiv_id":"2407.10362","evidence_quote":"supplies the LAB-Bench and Biomni-Eval1 benchmark tasks that define the four scientific task families."},{"cited_title":"DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models","cited_arxiv_id":"2402.03300","evidence_quote":"establishes the GRPO objective used in the outer recursion for model reinforcement learning."},{"cited_title":"Rubrics as Rewards: Reinforcement Learning Beyond Verifiable Domains","cited_arxiv_id":"2507.17746","evidence_quote":"introduces rubric-based rewards that serve as the trajectory-level reward signal in the outer loop."},{"cited_title":"Meta-Harness: End-to-End Optimization of Model Harnesses","cited_arxiv_id":"2603.28052","evidence_quote":"develops the Meta-Harness concept for end-to-end harness optimization, which the inner recursion builds on."},{"cited_title":"GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning","cited_arxiv_id":"2507.19457","evidence_quote":"provides the Reflexion paradigm for persistent procedural adaptation through verbal reflection."},{"cited_title":"HELIX: Model-Harness Co-evolution for Recursive Self-Improvement","cited_arxiv_id":"2608.13951","evidence_quote":"proposes model-harness co-evolution (HELIX), a direct precedent for the joint adaptation studied here."}],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}