{"id":"26b1b6ce-3cd5-47a4-939f-17a483f3d38c","arxiv_id":"2611.02839","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"An evidence-grounded recursive self-improvement loop, with blinded failure inspection, causal attribution, and verified repairs, brought a 31B customer-service model from below-incumbent quality to parity while cutting its inspected Bad Case rate to 1.97%.","lead":"This paper describes a production system in which an in-house AI customer-service model keeps learning from its own mistakes, with failures confirmed by independent reviewers and repaired before each retraining round. Over five rounds it matched the satisfaction of the commercial frontier model it replaced, cut median reply latency by 27.7%, and cut serving cost per order by 70.5%.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"The headline Bad Case decline and the training signal are co-produced by the same uncalibrated external inspection pipeline; without human-calibrated re-adjudication, the central improvement claim may partly reflect judge bias or drift.","rationale":"The reader's conditional verdict already rests on this weakest assumption, and my reading of the paper converges on the same point. What would have to be true for the central claim is that the externally driven inspection and attribution pipeline is accurate enough, and stable enough, that the measured Bad Case decline reflects genuine model improvement. The paper works hard to support this: evidence blinding, independent review, fail-closed terminal states, frozen simulation tests, and a separate online satisfaction measure. Those controls reduce obvious leakage and single-judge gaming, but they do not establish accuracy or invariance of the judge itself. The most distinctive numerical result, the drop from 6.42% to 1.97%, is produced by the same kind of model that also generates the training labels, and the independent human metric is consistent with only a modest parity claim. So the load-bearing weakness is not an internal inconsistency or a missing proof; it is an unverified empirical precondition. The proposed human re-adjudication is the direct check: if human raters reproduce the decline on blinded traces from the same frozen test, the concern is settled; if not, the central claim weakens substantially. Since the reader's verdict is already CONDITIONAL and explicitly flags this assumption, I recommend no change to the verdict.","tokens_in":41963,"tokens_out":6319,"duration_ms":64916,"concrete_test":"Re-adjudicate a fixed stratified sample of simulation traces from V0, V1, V4, and V5 with human expert raters using a frozen rubric, with model-version metadata removed, and compare human Bad Case rates and human attribution labels against the pipeline outputs. A sample of roughly 300 traces (about half the 559-case test) with each version's traces blinded and interleaved would be sufficient to detect a decline on the scale claimed: if the human-rated Bad Case rate still falls from the V0 level to the V5 level, the shared-judge concern is resolved; if the decline shrinks or disappears, the headline improvement is at least partly an artifact of judge bias or drift.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim depends on the inspection pipeline being accurate and stable, because that pipeline supplies both the training labels and the headline metric. Confirmed failures and diagnoses (Sections 5 and 6) feed OPSD, and the offline inspection Bad Case rate in Section 7.2 is computed by the same procedure on simulation traces. The paper acknowledges this exposure: attribution is a model-assisted judgment whose errors propagate into training, and attribution accuracy is not reported (Section 9.4), and calibration of the confirmed-bad rate against multilingual human annotation remains to be reported. The screener is also recalibrated on terminal states (Eq. 23), so the judge can drift across iterations. Case/denominator separation between training and evaluation prevents leakage but does not make the judge accurate or invariant. Because CARE-31B is post-trained from incumbent GPT-5.4 traces and judged by an external frontier model, correlated blind spots between the judge and the GPT-family behavior being imitated could lower judged Bad Cases without lowering true error rates. The independent human satisfaction metric reaches only non-significant parity in the final periods (p=0.85 and p=0.47), so it does not by itself confirm the 6.42% to 1.97% decline. The concern is therefore not that the paper is silent about the risk, but that the central quantitative claim has not been shown to survive an independent, calibrated judgment.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"The paper presents CARE, a grounded recursive self-improvement framework for production customer-service agents. Phase I (cold start) builds an initial in-house policy from governed production traces of a frontier incumbent, using a data engine with evidence-anchored admission, capacity-constrained hierarchical allocation (Eq. 8), and state-aligned expression refinement. Phase II is a recurring loop: the deployed in-house model serves part of production traffic; a 24/7 evidence-blinded online inspection (Eq. 18) confirms failures; attribution (Section 5.2) separates model causes from component causes; component failures are repaired outside the model; model-attributable failures are truncated at the first erroneous decision and paired with verified minimal repairs (Section 6.1); and on-policy self-distillation (OPSD) trains the model on its own sampled decisions against a frozen copy of itself that sees diagnosis and repair as privileged information (Eqs. 27-31). Candidates must pass an executable, oracle-blind simulation test compiled from the latest Skill/SOP (Section 7.1), an offline inspection of the resulting traces (Section 7.2), an online comparison with the incumbent (Section 7.3), and human approval.","tokens_in":42309,"tokens_out":18849,"duration_ms":140143,"significance":"If the improvement claim survives independent calibration, this is one of the more credible demonstrations of bounded, human-on-the-loop RSI in production. The paper's strengths are the specificity of its grounding mechanisms: reviewed and evidence-blinded failure confirmation; executed business state as the arbiter of actions; a promotion gate requiring parity on a fixed evaluation; case/denominator separation between training failures and evaluation; and a limitation section (Section 9.4) that candidly discloses the uncalibrated judge, unmeasured simulator fidelity, harness/model confounding, and non-finalized ablation batches. The formal conditions (Eq. 18 evidence blinding, Eq. 37 oracle blindness, Definition 2, and the OPSD loss family with its ablations) are stated precisely enough to be checked and reused. The loop demonstrably converts the model's own production failures into stable updates on a small data increment, and the cost and latency advantages are measured on concurrent traffic.","major_comments":[{"comment":"The headline result, the decline of the inspection Bad Case rate from 6.42% to 1.97% (below the incumbent's 4.60%), and the training labels of Section 6.1 are produced by the same evidence-blinded inspection pipeline. The authors acknowledge the exposure in Section 9.4: attribution accuracy is not reported, and calibration of the confirmed-bad rate against multilingual human annotation remains to be reported. The case/denominator separation between training and evaluation prevents label leakage but does not make the judge accurate or invariant. The judge is a frontier model, plausibly from the same family as the GPT-5.4 incumbent whose behavior is imitated in cold start, so correlated blind spots could lower judged Bad Cases without lowering true error rates; and the screener is recalibrated on the loop's own terminal states (Eq. 23), so the judge can drift across iterations even on a fixed case set. The independent human metric reaches only non-significant parity in the final two periods (p=0.85, p=0.47, Section 8.2), so it does not corroborate the 6.42-to-1.97 decline. Because this is the central quantitative claim, I ask for a concrete calibration: re-adjudicate a fixed sample of each version's traces on the 559-case test with multilingual human annotation and show that the judged decline tracks the human-judged decline; until then the gap to the incumbent is a model-assessed rate, not a calibrated error rate.","section":"Sections 5.1, 7.2, 8.2 (Table 6), 9.4, Eq. (23)","authors_acknowledge":"Its calibration against multilingual human annotation remains to be reported.","authors_acknowledge_p":0.9},{"comment":"Section 8.1 states that all versions and reference models are evaluated on the same frozen test of 559 cases, but Section 7.1 states that whenever the Skill/SOP changes, the scenario compiler automatically regenerates the simulation test from the latest version, and Section 7.1 warns that results on tests compiled from different Skill/SOP versions are not a controlled comparison. The paper does not state whether any Skill/SOP revision occurred between V0 (July 30) and V5 (October 15). If one did, either the 559-case test was recompiled, contradicting the same-frozen-test claim, or the cross-version comparison in Table 6 is not controlled. Please state explicitly whether the test remained identical across all six versions and how that is consistent with the automatic recompilation rule; if the test changed, restrict the claimed decline to versions that shared the same test and say which test each version used.","section":"Sections 7.1 and 8.1, Table 6","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The replacement criterion of Definition 1 is per-context: it requires Q(pi_theta,c) >= Q(pi_star,c) - delta and H(pi_theta,c) <= H(pi_star,c) for every c in C, with a free tolerance delta. Section 8.3 verifies the criterion with aggregate quantities: simulation accuracy on the 559-case test (94.10% vs 96.80%, a 2.7-point deficit), overall online satisfaction parity, and the inspection Bad Case rate mapped to the hard-violation term H even though that rate includes soft quality issues. As written, Eq. (3) is not instantiated: delta is never chosen, Q and H are not measured per context, and the aggregate quality proxy shows a deficit, so the claim that V5 satisfies the replacement criteria requires either an explicit delta with per-context measurements or a documented relaxation of Definition 1.","section":"Section 3.2 (Definition 1), Section 8.3, Eq. (3)","authors_acknowledge":null,"authors_acknowledge_p":null}],"minor_comments":[{"comment":"The statement that the combined score never decreases (V0 to V5 in Table 6) is a direct consequence of the promotion gate in Eq. (46), which admits a candidate only if S(theta') >= S(theta_k); labeling this as a property of the gate rather than an empirical finding would avoid a misleading reading.","section":"Section 8.2, Eq. (46)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The symbol B_t is overloaded: in Eq. (20) it denotes the admission capacity, while in Eq. (22) it denotes the set of confirmed Bad Cases; please rename one of the two.","section":"Section 5.1, Eqs. (20) and (22)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The abstract's headline figure (inspection Bad Case rate from 6.42% after cold start to 1.97%) is measured on the frozen 559-case simulation test, not on production conversations; Section 8.1 clarifies this, but the abstract should state it explicitly so the decline is not read as a production-traffic error rate.","section":"Abstract and Section 8.1","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The Bad Case rates are proportions over 559 cases (6.42% is about 36 cases, 1.97% about 11), and the offline metrics are reported without confidence intervals, as acknowledged in Section 9.4; the V4-to-V5 drop of 2.14 points is roughly 23 to 11 cases and is plausibly within sampling noise, so reporting Wilson intervals or counts, as Table 10 does, would help readers judge which steps of the decline are meaningful.","section":"Section 8.2, Table 6","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The vertical scale of the online-satisfaction panel is omitted, so the reader cannot verify the reported point differences or interval widths; please restore a labeled scale.","section":"Figure 7(a)","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"major_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"The one thing to know: this is the most complete production write-up of a bounded, human-on-the-loop RSI loop I have seen, and the headline offline gain rests on an inspection judge whose calibration is reported nowhere. The authors know this — Section 9.4 says attribution accuracy is not reported and the confirmed-bad rate's calibration against multilingual human annotation remains to be done.\n\nWhat is actually new: not the learning core, which is OPSD with privileged information from cited prior work, but the full production architecture around it — evidence-blinded 24/7 inspection with a publication gate, causal attribution that routes non-model failures to component repair, first-error truncation with verified minimal counterfactual repairs, the teacher-only privileged boundary, and a promotion gate on a frozen SOP-compiled simulation test. Five real iterations in a seven-region deployment is a serious study, and the ablations are informative: on a few hundred repair tuples, SFT collapses and DPO collapses into immediate handoff, while OPSD holds at 94.1% simulation accuracy.\n\nWhere it is soft, in proportion. The stress-test concern is real and unresolved: training failures and the headline Bad Case rate are co-produced by the same external-model inspection procedure. Case/denominator separation prevents leakage but does not make the judge accurate or invariant, and the screener is recalibrated on its own terminal states (Eq. 23). The online satisfaction result reaches only non-significant parity in the final periods (p = 0.85, p = 0.47), so it does not independently confirm the 6.42% to 1.97% decline — though it is a genuine, if weak, check that the model did not degrade service while cutting cost 70.5%. Offline metrics have no confidence intervals; 11 confirmed bad cases out of 559 carries a wide binomial interval, so the gap over the incumbent's 4.60% is plausible but not as sharp as the headline implies. The ablation counts come from separately finalized review batches, which the authors disclose, so those results rank configurations rather than attribute causes. None of this is hidden; the paper's limitation section is unusually honest.\n\nWho it is for: applied RSI researchers and anyone building production agent-improvement loops. It deserves a serious referee. The architecture is worth engaging with, and the honest limitation statement makes it a useful template for what an industrial RSI report should disclose. The main thing I would ask of a revision is a human-calibrated re-adjudication of the judge on a sample, plus intervals on the offline metrics.","headline":"A serious production RSI architecture with an honest limitation section; the headline Bad Case gain and the training signal share one uncalibrated judge, which is the load-bearing question for referees.","tokens_in":43026,"tokens_out":5282,"duration_ms":54875},"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek/deepseek-v4-flash","headline":"A production customer-service model learned reliably from its own verified failures, cutting its bad-case rate from 6.42% to 1.97%, below a frontier incumbent's 4.60%, over five self-improvement rounds.","keywords":["recursive self-improvement","on-policy self-distillation","privileged information","customer-service agents","failure attribution","production deployment","grounded feedback","LLM agents"],"falsifier":"Take a frozen sample of failures from all five iterations, have independent human experts produce a gold set of verdicts and attributions, and compare: if the external judge disagrees with the gold set on a large share of Bad Case labels, or if attribution accuracy on model-attributable cases is low, the loop would be reinforcing the judge's opinions rather than real improvement. A cheaper first check is to re-score V0's 6.42% bad-case corpus with the final iteration's judge and attribution settings to see how much of the measured decline to 1.97% survives a changed evaluator.","tokens_in":45396,"feed_emoji":"🤖","tokens_out":2391,"duration_ms":84773,"temperature":0.7,"pith_summary":"The paper claims that a production customer-service agent can keep improving itself after deployment by turning its own confirmed failures into training data, provided every failure is certified by evidence that is independent of the model being improved. CARE runs this loop on a 31B in-house model: a 24/7 inspection pipeline flags failures from the model's own traffic, attribution decides whether the model or a component is at fault, and only model-attributable failures are cut at the first wrong decision, paired with a verified repair, and fed back through on-policy self-distillation. Over five iterations the inspection bad-case rate falls from 6.42% to 1.97%, below the frontier incumbent's 4.60%, and online user satisfaction reaches parity while reply latency and serving cost drop sharply. If this holds, it means recursive self-improvement can be made safe enough for a live business by grounding it in external review rather than in the model's own judgment.","feed_headline":"In-house 31B model beats incumbent after five self-improvement rounds","feed_subtitle":"Its bad-case rate fell to 1.97%, less than half the incumbent's 4.60%, by learning from its own verified failures.","key_machinery":"The load-bearing mechanism is attribution-guided on-policy self-distillation (OPSD) at the first erroneous decision. For each confirmed, model-attributable failure, the trajectory is truncated at the exact request where the model first went wrong; the student samples a fresh decision from that input, while a frozen copy of the same model reads the same prefixes plus privileged information $I = (d, a^-, a^+)$ (diagnosis, erroneous output, verified minimal repair), and the loss is $\\ell_t = \\max(u^K_t, 0)$ with $u^K_t = \\sum_{v \\in S^K_t} q_t(v) \\log(q_t(v)/p_{\\theta,t}(v))$, a forward KL partial sum over the teacher's top-$K=32$ tokens, floored at zero per token and backpropagated directly. The teacher's only advantage is the privileged hindsight, which never reaches the student, so the student is corrected on what it would actually do now at the failing decision rather than on a fixed target text. Around this sit two supporting mechanisms: the evidence-blinded inspection pipeline (screening, deep analysis with a claim ledger, independent review, and a fail-closed publication gate, all driven by an external frontier model on frozen evidence without source metadata) and the executable simulation test compiled from the latest Skill/SOP and frozen, whose traces are inspected offline to produce the headline bad-case metric.","core_discovery":"On its own terms, the paper establishes that bounded recursive self-improvement works in production customer service when the loop is grounded by three independent checks: executed business state, reviewed evidence, and causal attribution. After a one-time cold start on governed traces of the incumbent GPT-5.4, the in-house CARE-31B model serves real traffic alongside the incumbent, and its own failures drive each update. A frozen, evidence-blinded inspection pipeline confirms failures, attribution routes non-model causes to component repair, and model-attributable failures become verified repair tuples at the first erroneous decision. Training uses on-policy self-distillation with the same model as teacher: the student samples its own decision where it failed, and a frozen copy of itself, given the diagnosis and repair as privileged information, defines the token-level target. After five iterations the confirmed-bad rate drops from 6.42% to 1.97% (less than half the incumbent's 4.60%), simulation accuracy rises from 93.0% to 94.1%, online satisfaction moves from 4.6 points below the incumbent to 1.7 points above, median reply latency falls 27.7%, and replacing the incumbent cuts serving cost per order by 70.5%. The ablations support the paper's central methodological claim: on a small increment of repairs, supervised fine-tuning collapses during training and preference optimization collapses into immediate handoff (15.1% simulated pass rate), while self-distillation trained on the same 280 repair tuples stays at 94.1%.","pith_inferences":["The authors state that attribution is a model-assisted judgment whose accuracy is not reported, so an unmeasured share of the headline decline could reflect the external judge's behavior; a natural check they do not run is to re-score V0's failure corpus with the final iteration's judge and attribution settings to see how much of the 6.42% to 1.97% drop survives a changed evaluator.","The same machinery transfers to other high-stakes service domains with an executable state model, such as banking or insurance claims, where a simulator can certify refunds, cancellations, and status transitions; the hard requirement is a state-transition function that is independent of the model's descriptions of its own actions.","Using the in-house model itself as the repair proposer, under the same independent verifier, would test whether the external frontier proposer is replaceable, which the paper names as its next step toward a more closed loop.","The paper argues, rather than measures, that mixing several SOP versions teaches policy-following; compiling the frozen simulation test separately for each SOP version, as the paper itself proposes, would turn that argument into a measurement."],"forward_implications":["A mid-sized in-house model can replace a frontier proprietary incumbent on quality, latency, and cost after five bounded self-improvement iterations, satisfying a formal replacement criterion of quality within tolerance, no more hard violations, and lower serving cost.","After the incumbent is retired, improvement does not stop, because the deployed model's own production failures supply both the training experience and the teacher, so fresh data no longer depends on the incumbent's traces.","Component failures (Skills, tools, harness, judge, infrastructure) are routed to component repair and kept out of the weights, so the model is not trained to work around broken parts of the system.","Small increments of repair data are usable only when the training target is the model's own distribution; fixed-text objectives (SFT, DPO) destabilize the model at this data scale, which is why self-distillation is the learning signal the loop depends on.","Business-policy revisions can be absorbed between model updates by mixing failures judged against the SOP in force when they occurred, so the model learns to follow whatever policy it is given rather than memorizing one version."],"supporting_citations":[{"why":"Supplies the RSI taxonomy (bounded versus open-ended, human-on-the-loop) that CARE uses to define its own design, scope, and safeguards.","marker":"[5]"},{"why":"Establishes model collapse from recursive training on self-generated data, the failure mode CARE's external grounding is designed to avoid.","marker":"[6]"},{"why":"Provides the evidence that LLM judges can be gamed, motivating the evidence-blinded, externally driven inspection pipeline.","marker":"[7]"},{"why":"Gives the reduction of imitation learning to no-regret on-policy learning, the basis for training on the student's own rollouts rather than fixed texts.","marker":"[16]"},{"why":"Defines generalized on-policy distillation of language models, the distillation family that OPSD extends to a self-teacher.","marker":"[17]"},{"why":"Introduces privileged information distillation for agents, the source of the teacher-sees-hindsight, student-does-not design.","marker":"[20]"},{"why":"Defines on-policy self-distillation with verified reasoning traces, the framework CARE adapts with failure-specific hindsight.","marker":"[21]"},{"why":"Motivates the executable, state-transitioning simulation test with outcome verification that the evaluation system implements.","marker":"[22]"},{"why":"Supplies DPO, the preference-optimization baseline whose collapse on small repair increments the ablation attributes to fixed-text targets.","marker":"[27]"}],"fun_headline_variants":["In-house 31B model cuts bad rate to 1.97% via self-improvement","CARE loop beats incumbent after five grounded self-improvement rounds","Self-distilled 31B agent beats incumbent in production customer service","Grounded self-improvement: 31B model halves incumbent's bad-case rate","CARE model learns from verified failures, beats incumbent in five rounds"],"cache_read_input_tokens":0,"weakest_assumption_plain":"Everything the model learns, and every error-rate figure in the paper, depends on an external frontier model's inspection and attribution judgments, and the paper assumes those judgments are accurate enough and sufficiently independent of the in-house model's own blind spots, an accuracy that the authors explicitly state is not reported.","fun_headline_variants_meta":{"raw":{"variants":["In-house 31B model cuts bad rate to 1.97% via self-improvement","CARE loop beats incumbent after five grounded self-improvement rounds","Self-distilled 31B agent beats incumbent in production customer service","Grounded self-improvement: 31B model halves incumbent's bad-case rate","CARE model learns from verified failures, beats incumbent in five rounds"]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":0.0001,"raw_usage":{"cost":0.000100191,"is_byok":false,"cost_details":{"upstream_inference_cost":0.000100191,"upstream_inference_prompt_cost":3.5217e-05,"upstream_inference_completions_cost":6.4974e-05},"total_tokens":1564,"prompt_tokens":1207,"completion_tokens":357,"prompt_tokens_details":{"audio_tokens":0,"video_tokens":0,"cached_tokens":1025,"cache_write_tokens":0},"completion_tokens_details":{"audio_tokens":0,"image_tokens":0,"reasoning_tokens":271}},"tokens_in":1207,"tokens_out":357,"cost_source":"openrouter_usage","duration_ms":4676,"image_pages":[],"temperature":1.0,"billed_attempts":1,"reasoning_tokens":271,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-11-03T07:09:46.085039+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":"Take a frozen sample of failures from all five iterations, have independent human experts produce a gold set of verdicts and attributions, and compare: if the external judge disagrees with the gold set on a large share of Bad Case labels, or if attribution accuracy on model-attributable cases is low, the loop would be reinforcing the judge's opinions rather than real improvement. A cheaper first check is to re-score V0's 6.42% bad-case corpus with the final iteration's judge and attribution settings to see how much of the measured decline to 1.97% survives a changed evaluator.","supporting_citations":[{"cited_title":"Recursive Self-Improvement in AI: From Bounded Self-Refinement to Autonomous Research Loops","cited_arxiv_id":"2607.07663","evidence_quote":"Supplies the RSI taxonomy (bounded versus open-ended, human-on-the-loop) that CARE uses to define its own design, scope, and safeguards."},{"cited_title":"AI models collapse when trained on recursively generated data.Nature, 631:755–759, 2024","cited_arxiv_id":null,"evidence_quote":"Establishes model collapse from recursive training on self-generated data, the failure mode CARE's external grounding is designed to avoid."},{"cited_title":"LLM-as-a-Judge Is Not an Oracle: Why Self-Improving Agents Need Deterministic Guardrails","cited_arxiv_id":"2609.02246","evidence_quote":"Provides the evidence that LLM judges can be gamed, motivating the evidence-blinded, externally driven inspection pipeline."},{"cited_title":"A reduction of imitation learning and structured prediction to no-regret online learning","cited_arxiv_id":null,"evidence_quote":"Gives the reduction of imitation learning to no-regret on-policy learning, the basis for training on the student's own rollouts rather than fixed texts."},{"cited_title":"On-Policy Distillation of Language Models: Learning from Self-Generated Mistakes","cited_arxiv_id":"2306.13649","evidence_quote":"Defines generalized on-policy distillation of language models, the distillation family that OPSD extends to a self-teacher."},{"cited_title":"Privileged Information Distillation for Language Models","cited_arxiv_id":"2602.04942","evidence_quote":"Introduces privileged information distillation for agents, the source of the teacher-sees-hindsight, student-does-not design."},{"cited_title":"Self-Distilled Reasoner: On-Policy Self-Distillation for Large Language Models","cited_arxiv_id":"2601.18734","evidence_quote":"Defines on-policy self-distillation with verified reasoning traces, the framework CARE adapts with failure-specific hindsight."},{"cited_title":"$\\tau$-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains","cited_arxiv_id":"2406.12045","evidence_quote":"Motivates the executable, state-transitioning simulation test with outcome verification that the evaluation system implements."},{"cited_title":"Direct Preference Optimization: Your Language Model is Secretly a Reward Model","cited_arxiv_id":"2305.18290","evidence_quote":"Supplies DPO, the preference-optimization baseline whose collapse on small repair increments the ablation attributes to fixed-text targets."}],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}