Abstention as an Action Can Kill Both the Reward Gradient and the KL Anchor: Collapse Law and Repair for Error-Penalized Reinforcement Learning

• URL: http://arxiv.org/abs/2608.00301v1
• Date: Fri, 31 Jul 2026 21:31:07 GMT
• Title: Abstention as an Action Can Kill Both the Reward Gradient and the KL Anchor: Collapse Law and Repair for Error-Penalized Reinforcement Learning
• Abstract summary: Error-penalized scoring rules ($+1$ for a correct answer, $$ for a wrong one, $0$ for abstaining) are increasingly prescribed against hallucination.<n>We prove that a KL-anchored learner can do the opposite.
• Score: 5.990509154718772
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Error-penalized scoring rules ($+1$ for a correct answer, $-λ$ for a wrong one, $0$ for abstaining) are increasingly prescribed against hallucination: a rational agent facing such a rule answers exactly when its correctness probability exceeds Chow's threshold $t^\ast=λ/(1+λ)$. We prove that a KL-anchored gradient learner can do the opposite. When abstention is a discrete action, the reward gradient and the anchor's restoring force are throttled by the same gate-saturation factor and die together: under explicit conditions (among them, blanket answering loses score in expectation and prompts share a bounded readout) the model drifts toward refusing everything, its mean training reward rising to zero like $1/t$ in training time $t$, so the curve reads as improvement while coverage collapses. The advantage estimator compounds the failure: in its sparse-answer regime, group normalization silently replaces every designed penalty with an effective penalty of one, moving the learned threshold from $t^\ast$ to $1/2$. The repair is structural: train a mandatory confidence report with a strictly proper score plus a correctness reward, and abstain only at deployment by thresholding the report. The always-emitted report has no gate to saturate, so no shared factor can kill its reward gradient and its anchor together, and its calibrated optimum is attracting. Simulations confirm every prediction, and experiments on language models at two scales confirm the mechanism live: the rule silences questions the models demonstrably still solve within ten optimizer steps, an ablation isolates the cause, and report-level training raises coverage, accuracy, and calibration together.

Related papers

• Constrained Online Learning with Noisy Constraint Values [55.29259818039367]
  Under common feasibility, our LEDGER algorithm attains $O(sqrt T)$ expected regret and $O(sqrtTlog(eT))$ expected budget violation.<n>Neither a Slater condition, independence between feedback channels, nor an absolute constraint-value bound is needed.
  arXiv  Detail & Related papers  (2026-09-07T01:38:41Z)
• Convex-Hull-Neighborhood Smooth Dual Generalization: Controlling Local Correction Propagation in Offline RL [7.803284875836737]
  bootstrapping can amplify estimation errors in offline reinforcement learning.<n>We propose Convex Hull Neighborhood Smooth Dual Generalization, which expresses the Bellman backup as an in-sample value target plus a CHN-local correction.<n> Experiments on Gym-MuJoCo and AntMaze show strong aggregate performance and stable value estimation.
  arXiv  Detail & Related papers  (2026-08-04T04:30:02Z)
• H$^2$SD: Hybrid Hindsight Self-Distillation [61.71131241473028]
  Reinforcement learning with verifiable rewards (RLVR) provides reliable outcome supervision for language model reasoning.<n>Existing self-distillation methods add a privileged teacher but typically assign it a fixed role.<n>We introduce Hybrid Hindsight Self-Distillation ($mathrmH2mathrmSD$), which jointly adapts teacher context and update strategy to trajectory correctness.
  arXiv  Detail & Related papers  (2026-07-21T10:47:27Z)
• Transferred QAOA Parameters Remember the Penalty Scale: A $λ$-Resonance Law for Constrained Quantum Optimization [0.0]
  We identify a new, independent axis that governs transfer whenever constraints are encoded as penalties.<n>We prove that at arbitrary fixed QAOA angles $(,)$ of depth $p$, the probability $F()$ on the feasible subspace is a finite real trigonometric mass.<n>The theorem is independent of how the angles were obtained and applies to any integer-penalty QUBO.
  arXiv  Detail & Related papers  (2026-07-10T19:23:41Z)
• A Pre-Registered Causal Partition of Self-Consistency Elicitation and Reward Design in RLVR [1.2958054117511815]
  Reinforcement learning from rewards improves reasoning even when the reward signal is spurious.<n> Practitioners commonly interpret naive = acc(TRUE) - acc(R) as the reward-design effect.<n>We prove this estimand is systematically biased.
  arXiv  Detail & Related papers  (2026-06-04T09:35:54Z)
• The Extrapolation Cliff in On-Policy Distillation of Near-Deterministic Structured Outputs [52.709361620508595]
  ListOPD brings a student to in-domain with an 8B-SFT baseline at one-fifth the parameters.<n>On Amazon Fashion, three pre-registered tests--a fine-grid cliff interval test, and a small-clip cross-prediction--fall within their locked prediction windows--with the small-clip value matching the closed-form prediction below grid resolution.
  arXiv  Detail & Related papers  (2026-05-09T06:48:00Z)
• MEMSAD: Gradient-Coupled Anomaly Detection for Memory Poisoning in Retrieval-Augmented Agents [0.0]
  We formalize memory poisoning attacks on retrieval-augmented agents as a Stackelberg game with a unified evaluation framework.<n>We show faithful evaluation increases measured attack success by $4times$ (ASR-R: 0.25 to 1.00$)<n>Our primary contribution is MEMSAD, a calibration-based defense grounded in a gradient coupling.
  arXiv  Detail & Related papers  (2026-05-05T08:15:41Z)
• Optimal Unconstrained Self-Distillation in Ridge Regression: Strict Improvements, Precise Asymptotics, and One-Shot Tuning [61.07540493350384]
  Self-distillation (SD) is the process of retraining a student on a mixture of ground-truth and the teacher's own predictions.<n>We show that for any prediction risk, the optimally mixed student improves upon the ridge teacher for every regularization level.<n>We propose a consistent one-shot tuning method to estimate $star$ without grid search, sample splitting, or refitting.
  arXiv  Detail & Related papers  (2026-02-19T17:21:15Z)
• CARE What Fails: Contrastive Anchored-REflection for Verifiable Multimodal [84.71254539482369]
  Group-relative reinforcement learning with verifiable rewards (RLVR) often wastes the most informative data it already has the failures.<n>We present CARE, a failure-centric post-training framework for multimodal reasoning that turns errors into supervision.<n> CARE improves accuracy and training smoothness while explicitly increasing the share of learning signal that comes from failures.
  arXiv  Detail & Related papers  (2025-12-22T16:34:21Z)
• Honesty over Accuracy: Trustworthy Language Models through Reinforced Hesitation [12.503662455234954]
  We show that modern language models produce confident hallucinations even when wrong answers carry catastrophic consequences.<n>We propose Reinforced Hesitation (RH): a modification to Reinforcement Learning from Verifiable Rewards (RLVR) to use ternary rewards instead of binary.
  arXiv  Detail & Related papers  (2025-11-14T17:20:45Z)
• Reinforcement Learning with Verifiable yet Noisy Rewards under Imperfect Verifiers [90.50039419576807]
  Reinforcement Learning with Verifiable Rewards (RLVR) trains policies against automated verifiers to avoid costly human labeling.<n>To reduce vulnerability to verifier hacking, many RLVR systems collapse rewards to binary $0,1$ during training.<n>This choice carries a cost: it introduces textitfalse negatives (rejecting correct answers, FNs) and textitfalse positives (accepting incorrect ones, FPs)
  arXiv  Detail & Related papers  (2025-10-01T13:56:44Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.