Safe Error Correction for Language Models: Frozen-Base Adjustment with Capability Preservation

• URL: http://arxiv.org/abs/2609.16145v1
• Date: Mon, 14 Sep 2026 18:00:15 GMT
• Title: Safe Error Correction for Language Models: Frozen-Base Adjustment with Capability Preservation
• Abstract summary: We propose a lightweight logit-level correction module atop a frozen language model.<n>CRN v2 corrects 53.3% of base-model errors while showing no degradation on tested capability benchmarks.<n>None of the alternative configurations we tested exceeded the rank-128 logit result.
• Score: 0.0
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: We study a practical question: can a small correction module fix errors in a frozen language model's outputs without degrading its base capabilities? We propose CRN v2, a lightweight logit-level correction module (~34M trainable parameters, 0.73% of the 4.65B text module) that sits atop a fully frozen Gemma 4 E2B model. The base model is never updated; only the correction module learns, via supervised fine-tuning followed by reference-free DPO on 83,400 error-correction pairs. On a 60-question domain exam (CEHRI: Certified Human-Robot Intelligence, covering facts, arithmetic, and implicit-goal reasoning), CRN v2 corrects 53.3% of base-model errors (reworded variant: 43.3%) while showing no degradation on tested capability benchmarks (MMLU/BoolQ N=200; car-wash N=8). A LoRA baseline at the matched CRN v1 budget (6.6M params, rank 19) achieves 83.3% correction but suffers 30-75% capability loss on the same benchmarks -- the correction-capability tradeoff. An ablation shows that the KL preservation term (lambda=0.1) is critical: lowering it to 0.01 degrades correction to 35.0%. A hidden-state injection variant at earlier layers (1.6M params, SFT-only) reaches 50.0%/55.8% but does not exceed logit correction; shallower injection (layer 4) drops to 30.0%/28.3%; multi-depth logit correction (~35M) reaches only 40%; and longer training (5,000 SFT + 2,000 DPO) stays at 53.3% -- none of the alternative configurations we tested exceeded the rank-128 logit result, consistent with a best-achieved result of ~53% rather than a floor. This is a study of a design principle (frozen base + logit correction + KL anchoring), not a claim of architectural novelty. All code, main-result weights, and evaluation scripts are released (deep variant as code only -- no trained deep checkpoints).

Related papers

• Target-Checked Reliability Score Refinement for Video Question Answering [3.8271803328378677]
  We study whether answer-level reliability scores can be improved under target shift without retraining the models or changing their answers.<n>We collect option-probability lists from three fixed video-language models under four deterministic video samplings.<n>We compare the original score, defined as the probability assigned to the chosen answer, with a histogram-based gradient-boosting (HGB) score trained on the development datasets.<n>A candidate replaces the original score only when repeated video-level checks indicate a positive, stable improvement.
  arXiv  Detail & Related papers  (2026-09-09T08:29:00Z)
• SS-ESOAP: Self-Scaled Adaptive Preconditioning for Physics-Informed Learning [70.71795707610961]
  method augments SOAP-style preconditioning with a scalar secant-energy correction adapted to Kronecker geometry.<n>Method attains the lowest final residual on six of eight PDE benchmarks.<n>Three-seed $L2$ and $H1$ errors on four representative PDEs support the link between lower residuals and improved solution accuracy.
  arXiv  Detail & Related papers  (2026-08-29T21:34:47Z)
• Capability-Stratified Degradation in Ternary Language Models [0.0]
  Ternary language models restrict weights to $-1,0,+1$, approaching the limit of $log 3 approx 1.585$ bits/weight.<n>We convert Qwen3.5-0.8B (752M parameters) to ternary weights using 72.4M tokens of quantisation-aware training (QAT)<n>The resulting model, Cloe, is evaluated across 29 benchmarks, representation diagnostics, and downstream fine-tuning.
  arXiv  Detail & Related papers  (2026-08-28T19:22:38Z)
• Stuck on "A": Diagnosing and Repairing Interface Injury in Attention-to-KDA Linearization of a 0.6B Language Model [0.0]
  We convert 21 of 28 full-attention layers of Qwen3-0.6B-Base into KDA (Kimi Delta Attention)<n>After surgery, hidden-state alignment and end-to-end KL distillation drive the student close to its teacher in perplexity.<n>We release code, weights, recipes, and the full audit trail, and distill the engineering lessons.
  arXiv  Detail & Related papers  (2026-08-03T08:54:18Z)
• Reason, Reward, Refine: Step-Level Errors Corrections with Structured Feedback for Physics Reasoning in Small Language Models [32.7643564019866]
  We propose a step-level reward framework that identifies the first reasoning error, generates targeted structured feedback, and trains the model to revise its solution via policy gradient with KL regularization.<n>Across five physics benchmarks, our framework delivers accuracy gains of 17-20% over CoT prompting and 10-16% over the strongest baseline, reduces calculation errors from 56.9% to 23.5%, and reduces miscomprehension errors from 22.3% to 12.0% in the best observed cases.
  arXiv  Detail & Related papers  (2026-07-06T15:16:10Z)
• Knowledge Index of Noah's Ark [63.143852586221534]
  We introduce KINA, an 899-item benchmark across 261 fine-grained disciplines.<n>We prove a bonus-on-bar tournament weakly FOSD-dominates flat payment in released-review quality.<n>The top model, Gemini-3.1-Pro-Preview, reaches 53.17%, followed by Claude-Opus-4.6 at 49.92% and GPT-5.4 at 48.55%.
  arXiv  Detail & Related papers  (2026-06-03T17:06:49Z)
• ARC-STAR: Auditable Post-Hoc Correction for PDE Foundation Models [47.9430487327673]
  ARC-STAR organizes correction into three stages: a global corrector removes broad solver bias, a blockwise local refiner cleans the post-global residual, and, at deployment, a label-free score routes refinement to high-risk blocks under a compute budget.<n>ARC-STAR is the only method that cuts velocity rollout error by at least 36x over raw Poseidon on every cell.
  arXiv  Detail & Related papers  (2026-05-21T09:26:16Z)
• Step-wise Rubric Rewards for LLM Reasoning [72.17879367869503]
  Reinforcement Learning with Verifiable Rewards (RLVR) is widely used to improve reasoning in large language models.<n>We find 18.2% of steps in correct responses are wrong yet positively rewarded.<n>Across six mathematical reasoning benchmarks, SRaR improves average accuracy over RaR by 3.57 points.
  arXiv  Detail & Related papers  (2026-05-17T07:08:14Z)
• Mechanistically Guided LoRA Improves Paraphrase Consistency in Medical Vision-Language Models [2.064612766965483]
  Medical Vision-Language Models can give different yes or no answers to rephrasings of the same clinical question.<n>We study this in MedGemma-4B using PSF-Med Sadanandan and Behzadan (2025), which provides paraphrase pairs for systematic consistency evaluation on medical VQA.<n>We fine-tune Low-Rank Adaptation (LoRA) adapters with a combined loss that balances paraphrase consistency with answer accuracy.
  arXiv  Detail & Related papers  (2026-02-24T23:30:50Z)
• Squeeze the Soaked Sponge: Efficient Off-policy Reinforcement Finetuning for Large Language Model [56.92219181993453]
  We propose Reincarnating Mix-policy Proximal Policy Gradient (ReMix) to enable on-policyRFT methods like PPO and GRPO to leverage off-policy data.<n>ReMix consists of three major components: (1) Mix-policy proximal policy gradient with an increased Update-To-Data (UTD) ratio for efficient training; (2) KL-Convex policy constraint to balance the trade-off between stability and flexibility; (3) Policy reincarnation to achieve a seamless transition from efficient early-stage learning to steady improvement.
  arXiv  Detail & Related papers  (2025-07-09T14:29:45Z)
• High Accuracy, Less Talk (HALT): Reliable LLMs through Capability-Aligned Finetuning [84.52940628494879]
  Large Language Models (LLMs) currently respond to every prompt.<n>LLMs can produce incorrect answers when they lack knowledge or capability.<n>We propose post-training an LLM to generate content only when confident in its correctness and to otherwise abstain.
  arXiv  Detail & Related papers  (2025-06-04T15:16:21Z)

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.