{"id":"9bf5dec1-f20b-45bc-a7da-24f6cd53674d","arxiv_id":"2611.23476","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"GUARD uses regret-guided ranking and a non-regression gate to let frozen-model agents continually improve their text harness without external graders or forgetting.","lead":"GUARD is a system that helps AI agents improve their own prompts and rules without forgetting what they already learned. It could make autonomous agents more reliable and efficient in real-world use.","discovery_kind":"new_method","skeptic_critique":{"audit":{"quote":"These results characterize the self-review objective on gated task buffers; held-out accuracy is evaluated empirically in \\S\\ref{sec:results}.","status":"reviewer_error","question":null,"consequence":null,"objection_id":"O1","published_as":"withdrawn"},"model":"deepseek/deepseek-v4-flash","headline":"Withdrawn: Pith's reviewer misread the paper.","rationale":"","tokens_in":70683,"tokens_out":770,"duration_ms":25250,"concrete_test":"","verdict_should_be":null,"load_bearing_attack":"The paper's words: “These results characterize the self-review objective on gated task buffers; held-out accuracy is evaluated empirically in \\S\\ref{sec:results}.”","agreement_with_reader":null},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"The paper introduces GUARD, a framework for recursive self-improvement of agent harnesses (prompts, skills, memory) around a frozen LLM, without external graders. It uses pre-execution checklists, paired regret ranking (Analyze), GRPO-inspired candidate pooling (Propose), and a Soft-HURT gate that promotes edits only when gains outweigh regressions. Theoretical guarantees (non-regression on self-review objective, targeted repair vs passive discovery, strict regression dominance) are provided. Experiments on PAST-Bench, EvoHarnessBench, Drift Gauntlet, and professional suites show GUARD improves or matches baselines, achieves 19/19 non-regressive promotions, zero post-block forgetting, and cuts token usage.","tokens_in":70907,"tokens_out":288,"duration_ms":16223,"significance":"If the results hold, GUARD addresses a key problem in autonomous agent improvement: how to continually update a persistent harness without external supervision while avoiding catastrophic forgetting and local overfitting. The combination of checklist-based self-review, paired regret ranking, group-relative candidate selection, and regression-aware gating is novel. The empirical evaluation is extensive across multiple benchmarks with careful baselines. The theoretical results provide formal guarantees on the self-review objective. The work is significant for the field of autonomous agents and continual learning.","major_comments":[],"minor_comments":[],"open_questions":null,"recommendation":"accept"},"author_rebuttal":null,"desk_editor":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek/deepseek-v4-flash","headline":"The paper argues that GUARD, a regret-guided framework, stabilizes recursive self-improvement of agent harnesses without external graders.","keywords":["recursive self-improvement","agent harness","regret guidance","continual learning","catastrophic forgetting","self-review","group-relative scoring","non-regressive gating"],"falsifier":"If a held-out evaluation shows that a promoted harness (accepted by Soft-HURT based on self-review scores) decreases accuracy on the same task cohort compared to the incumbent, that would falsify the claim that the gate guarantees non-regression on held-out performance. The paper provides calibration evidence (Table 6) showing 19/19 execution-gated promotions have Δ_test ≥ 0, but a single counterexample would challenge the assumption that self-review scores track hidden verifiers well enough.","tokens_in":72316,"feed_emoji":"🛡️","tokens_out":1812,"duration_ms":20985,"temperature":0.2,"pith_summary":"The paper introduces GUARD, a closed-loop framework for reference-free continual evolution of agent harnesses (prompts, skills, tool rules, memory) around a frozen language model. The central problem is that self-editing without external graders often backfires because scalar scores conflate what an edit newly broke with what remains unsolved, causing regressions, forgetting, and memory bloat. GUARD addresses this by decomposing each task into pre-execution self-checks, ranking criteria by directional regret (separating new regressions from old gaps), sampling parallel candidate edits scored by group-relative fitness, and promoting only net-positive updates via a Soft-HURT gate that guarantees non-regression on its gated buffer. The authors claim that across benchmarks on memory, skill evolution, continual task streams, and domain reasoning, GUARD consistently improves performance without catastrophic forgetting or memory pollution while cutting deployment token usage.","feed_headline":"Regret-guided framework stabilizes recursive self-improvement of AI agents","feed_subtitle":"GUARD claims to prevent forgetting and regressions by separating edit-induced harms from unresolved gaps via checklist self-review and a…","key_machinery":"The central object is the paired directional regret decomposition for each criterion c: Priority_t(c) = Δ_hurt,t(c) + Δ_gap,t(c), where Δ_hurt,t(c) = max(0, obey(c; W_{t-2}) - obey(c; W_{t-1})) captures edit-induced regressions and Δ_gap,t(c) = 1 - obey(c; W_{t-1}) captures unresolved gaps. This ranking separates newly broken criteria from static gaps, pins mastered criteria (Priority=0), and feeds into the Soft-HURT gate which promotes a candidate only when net gain exceeds induced hurt. The group-relative candidate pooling (inspired by GRPO) samples M parallel harnesses, scores them on the same tasks, and selects by standardized fitness without a learned critic.","core_discovery":"The paper claims that by decomposing each task into pre-execution atomic binary checks (criteria) and guiding harness updates via paired directional regret—separating what an edit newly broke (Δ_hurt) from what remains unsolved (Δ_gap)—an agent can continually improve its persistent harness in a reference-free setting without interfering with or catastrophically forgetting previously learned skills. The key mechanism is a closed loop: Analyze (paired regret ranking pins mastered criteria and prioritizes regressions), Propose (parallel candidate harnesses scored by group-relative fitness, with contrastive prompt evolution), and Gate (Soft-HURT promotion that accepts an edit only when gains outweigh induced regressions, with [iff] precondition scoping). The authors report that this yields 19/19 non-regressive execution-gated promotions across EvoHarnessBench, Drift Gauntlet, and professional suites, zero post-block forgetting on all four Drift Gauntlet streams, and 54.5% fewer tokens per task on hosted EvoHarnessBench while improving accuracy.","headline_check":{"rule":"report_a_claim_v1","sent_back":false,"unattributed":{}},"pith_inferences":["The paper's reliance on a self-reviewer's checklist-based grading as a proxy for the unobserved true task evaluator V* is load-bearing; if the self-reviewer systematically misses or misjudges criteria, the gate's non-regression guarantee on the self-review objective may not translate to held-out accuracy. The authors provide calibration evidence (Table 6) but acknowledge this gap.","The approach could be extended to multi-agent settings where each agent maintains its own harness and shares regret bundles, potentially enabling collaborative skill discovery without centralized grading.","The contrastive prompt evolution (contrasting best and worst candidates) is a novel mechanism for iteratively improving proposal prompts; it would be interesting to test whether this technique generalizes to other domains like code generation or robotic policy search.","The paper's theoretical results (Theorem 1) guarantee non-regression on the self-review objective, not on held-out accuracy; a practical extension would be to bound the gap between self-review scores and true task performance under distribution shift."],"forward_implications":["If GUARD's claims hold, autonomous agents could continually improve their behavior from experience without weight updates or external supervision, reducing the need for human prompt engineering and manual harness tuning.","The approach could enable agents to adapt to shifting tools, schemas, and user workflows over long deployments without catastrophic forgetting, addressing a key bottleneck in practical agent systems.","The 54.5% reduction in tokens per task on EvoHarnessBench suggests that compact, scoped rules (via [iff] preconditions) can replace verbose, unscoped instructions, lowering inference cost while improving accuracy.","The zero post-block forgetting on Drift Gauntlet implies that the Soft-HURT gate and signature-routed playbook sections can maintain performance across unannounced task distribution shifts, a common failure mode in continual learning."],"supporting_citations":[{"why":"Introduces unsupervised environment design (UED) and the concept of regret for curriculum generation, which GUARD adapts to criterion-level regret ranking.","marker":"[dennis2020emergent]"},{"why":"Introduces Group Relative Policy Optimization (GRPO), which inspires GUARD's group-relative candidate scoring without a value critic.","marker":"[shao2024deepseekmath]"},{"why":"Proposes training-free GRPO, which GUARD extends to non-parametric harness optimization without weight updates.","marker":"[cai2025trainingfree]"},{"why":"RRSI is a direct predecessor that performs recursive self-improvement of harnesses with a score-band gate; GUARD compares against it and claims to address its unbounded per-task drops.","marker":"[xia2026rrsi]"},{"why":"Introduces Drift Gauntlet, a benchmark for continual learning under hidden stream drift, which GUARD uses to evaluate forgetting and plasticity.","marker":"[chen2026continual]"},{"why":"Introduces EvoHarnessBench, a benchmark for continual evolution across tools and skills, which GUARD uses as a primary evaluation suite.","marker":"[evoharnessbench2026]"},{"why":"Introduces PAST-Bench, a benchmark for cross-session memory and write-path attribution, which GUARD uses to evaluate memory pollution and procedural task performance.","marker":"[xue2026pastbench]"},{"why":"Meta-Harness is a baseline that searches over harness code using raw trace archives; GUARD compares against it and claims better scalability and non-regression.","marker":"[lee2026metaharness]"},{"why":"GEPA is a baseline that evolves prompts via reflective Pareto archives; GUARD compares against it and claims better handling of regressions and forgetting.","marker":"[agrawal2026gepa]"}],"cache_read_input_tokens":0,"weakest_assumption_plain":"The self-reviewer's checklist-based grading (Layer-1 sandbox checks plus Layer-2 semantic grading) provides a reliable proxy for the unobserved true task evaluator V*, and the Soft-HURT gate's decisions based on this self-review score generalize to held-out test accuracy.","cache_creation_input_tokens":0},"created_at":"2026-11-19T03:15:35.571101+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-medium+deepseek/deepseek-v4-flash+p.ce10e2b95288","input_kind":"tex","input_chars":202705,"input_total_chars":202705,"passes_missing":["desk_editor"],"published_standing":"established_narrower","published_confidence":"moderate","objection_audit":{"version":"pith-objection-audit@1","counts":{"raised":1,"confirmed":0,"questions":0,"withdrawn":1,"moved_to_minor":0,"math_error_not_certified":0,"withdrawn_reviewer_error":1,"typesetting_minor_dropped":0,"withdrawn_already_addressed":0},"open_questions":[],"standing":{"basis":["first_pass_verdict"],"standing":"established_narrower","confidence":"moderate"},"standing_rule":"1. No standing is published (not_assessed) when the review read only the abstract, when the referee pass or the objection check did not run, or when the first pass gave no verdict. 2. The first pass's verdict sets the start: ACCEPT is established, CONDITIONAL is established_narrower, REJECT is not_supported. 3. Each objection that survived the check against the paper's text caps the standing: one that breaks the central claim caps it at not_supported, one that narrows it at established_narrower, one that weakens it at supported_with_gaps. The standing is the lowest of the start and the caps. 4. A first-pass REJECT stands only if a surviving objection breaks the central claim. Without one, the standing is the lowest cap of the surviving objections, or not_assessed when none survived. 5. The confidence is the lowest of the first pass's confidence and the confidence of each objection that set the standing, and is no higher than moderate while a question is open or part of the paper's text was not supplied. 6. If any objection was withdrawn because Pith's reviewer misread the paper, the confidence is no higher than moderate, whatever the first pass said: a review that misread the paper once may have misread it elsewhere.","input_kind":"tex","nothing_to_audit":false,"bar":{"enabled":false}},"run_record":{"models":["deepseek/deepseek-v4-flash"],"passes":{"reader":{"model":"deepseek/deepseek-v4-flash","effort":"medium","cost_usd":0.007181,"tokens_in":75344,"tokens_out":3318,"duration_ms":27603,"input_sha256":"eaf1390e9c14609baeadd2de0991bea21519c2af94e143eb3963db83b75ef5f6","prompt_sha256":"f9ead730038216984cfb56d7cb34332bdf691828296a1de4245f3d6667d4bf56","message_sha256":"70dbc1c5f1af7f6dbdd3133ad62d101fb722d0cef45e65c60563d76d3fd8e601","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"referee":{"model":"deepseek/deepseek-v4-flash","effort":"medium","cost_usd":0.006505,"tokens_in":70907,"tokens_out":288,"duration_ms":16223,"input_sha256":"eaf1390e9c14609baeadd2de0991bea21519c2af94e143eb3963db83b75ef5f6","prompt_sha256":"a852d99aea0ddfd900e1483f76e20777aae3afde9c016d5b8d6e5db01064c728","message_sha256":"4a405f382614ac26f3f894ef436c44a7477995fbd87e45681f123d4c9b7d7d09","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"skeptic":{"model":"deepseek/deepseek-v4-flash","effort":"medium","cost_usd":0.006572,"tokens_in":70683,"tokens_out":770,"duration_ms":25250,"input_sha256":"eaf1390e9c14609baeadd2de0991bea21519c2af94e143eb3963db83b75ef5f6","prompt_sha256":"8f88c0f61488b1e764c317998cddf35fffc33a1f20a3b1be54330d7843386fc7","message_sha256":"36e1d3ee25064da83e72acff3130c9da8363f7289371e9daac38b0a43eca1ae1","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"circularity":{"model":"deepseek/deepseek-v4-flash","effort":"medium","cost_usd":0.010545,"tokens_in":70659,"tokens_out":2742,"duration_ms":19007,"input_sha256":"eaf1390e9c14609baeadd2de0991bea21519c2af94e143eb3963db83b75ef5f6","prompt_sha256":"e9946e39d67b76cc60476379bc32abdc44665b5f0c438659796547bb1041a01f","message_sha256":"4de4168a5ea86c6f59c6a358649ef9d05831b841eef2e8786e08f4aa27d574cd","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"objection_audit":{"model":"deepseek/deepseek-v4-flash","effort":"medium","cost_usd":0.008032000000000001,"tokens_in":142626,"tokens_out":1289,"duration_ms":35091,"input_sha256":"eaf1390e9c14609baeadd2de0991bea21519c2af94e143eb3963db83b75ef5f6","prompt_sha256":"6c7a6883d5517cf9e4945d83947200a36dcd152b8b56c32baf7500d8b3841d0a","message_sha256":"160926c745e09e96ee5a7781e8236e41f138e50963f0819e1c89e66d3299a71f","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"pith_extraction":{"model":"deepseek/deepseek-v4-flash","effort":"medium","cost_usd":0.006911,"tokens_in":72316,"tokens_out":1812,"duration_ms":20985,"temperature":0.2,"input_sha256":"eaf1390e9c14609baeadd2de0991bea21519c2af94e143eb3963db83b75ef5f6","prompt_sha256":"e4f56bf568e7854d0a1d2efb6ec96f8697d5132737cdb31cb7fc64acbd923974","message_sha256":"09e4dc5b5b4ddf089f073a92c7304ec445b23b3c192019c079d075ca31878abb","cache_read_input_tokens":0,"cache_creation_input_tokens":0}},"version":"pith-run-record/v1","prompt_set_sha256":"ce10e2b952880e823f7b29deab33764e2644d21b6b7c6ddc90a261514435ee02","configured_pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-medium"},"basis":{"kind":"full_text","source":"recorded","chars_given":202705,"chars_total":202705},"receipt":{"review_sha256":"d9fe4be8576d89901791ec863a64e6d52b4c1c8c1a25940d1329bbe1c0811859","rule":"pith-review-sha256/v1","hashed_at":"2026-11-19T03:17:50.318508+00:00","fields":["id","arxiv_id","paper_version","created_at","pipeline_version","model_set","run_record","input_kind","input_chars","input_total_chars","passes_missing","verdict","confidence","novelty_score","correctness_risk","formal_verification","parameter_count","discovery_kind","one_line_summary","lead","skeptic_critique","referee_report","author_rebuttal","desk_editor","pith_extraction","objection_audit","published_standing","published_confidence"],"matches_current":true},"falsifier":"If a held-out evaluation shows that a promoted harness (accepted by Soft-HURT based on self-review scores) decreases accuracy on the same task cohort compared to the incumbent, that would falsify the claim that the gate guarantees non-regression on held-out performance. The paper provides calibration evidence (Table 6) showing 19/19 execution-gated promotions have Δ_test ≥ 0, but a single counterexample would challenge the assumption that self-review scores track hidden verifiers well enough.","supporting_citations":[],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":"tex"},"url":"https://pith.science/accuracy"},"review_version":1}