arXiv AI

When Should LLMs Trust Their Own Revisions? A Risk-Aware Study of Intrinsic Self-Correction

The paper investigates intrinsic self‑correction, where a language model revises its own answer without new evidence. Across 29 open‑weight LLMs on BoolQ, GSM8K, and Corr2Cause, the study tracks how revisions change correctness, revealing that while some models improve significantly, others lose a notable fraction of correct answers. The authors compare three runtime strategies—keeping the initial answer, always accepting the revision, and selectively gating revisions—and find that the best approach depends on the model and task, suggesting that self‑correction should be treated as a revision policy rather than a uniformly beneficial second pass.

arXiv AI
3d ago

When Expert Disagreement Hurts: Prestige and Selectivity in LLM Decision Revision

The study introduces a two‑pass audit to evaluate how language models revise decisions when faced with disagreement. By fixing the question, evidence, and initial answer, the audit isolates the effect of different revision contexts—neutral reflection, anonymous disagreement, or disagreement from a domain expert—on accuracy. Across five models and 1,989 evidence‑grounded decisions in medicine, science, and law, expert disagreement rarely improves accuracy, often reducing it by up to 16.1 points, and the number of reversals is a poor indicator of harm.

By Yupeng Tang, Mingfeng Lin
arXiv Machine Learning
Sep 25

Return or Revise? Learning When Revision Helps Retrieval-Augmented QA

The paper investigates when it is better to return an existing draft answer or revise it using retrieved evidence in retrieval‑augmented QA systems. By grading both the draft and its candidate revision with the same correctness judge, the authors define a paired effect called recoverability and train policies to predict it before revision. Experiments on 25,870 open‑domain questions show that a recoverability‑based scorer outperforms a draft‑correctness scorer across multiple Llama setups, improving accuracy–revision trade‑offs and closing a significant portion of the oracle gap, though it still applies harmful revisions in a substantial fraction of cases.

By Nicholas Kashani Motlagh, Tim Anderson, Jeremy Gwinnup, Grant Erdmann
arXiv AI
Oct 2

Beyond Final Accuracy: Auditing Communication in LLM Multi-Agent Systems

The paper introduces Independent–Communicate–Revise (ICR), a framework that isolates communication effects in large language model multi‑agent systems by fixing initial reasoning and measuring how messages influence answer revision. ICR evaluates correction, preservation, and selectivity across four reasoning benchmarks, revealing that similar overall accuracy can mask divergent revision behaviors. The study shows that richer messages can both improve and harm outcomes, and that receiver policies can shift preservation and correction dynamics differently across tasks.

By Shixuan Li, Wei Yang, Peiyu Zhang, Anzhe Cheng, Heng Ping, Paul Bogdan
arXiv Computation and Language
Sep 24

Exact Feedback Is Not Control: Evaluating Text-based Closed-Loop Revision in LLMs

The paper introduces a fixed‑budget revision protocol that uses deterministic verifiers to expose all remaining violations across exact‑length, lexical, and compositional constraints, thereby isolating model‑side revision behavior. Experiments on 19 open‑ and closed‑source LLMs show wide variability in controller‑level success, with some models achieving up to 99.8% success while others remain below 20%. Controlled studies reveal that post‑training and scale affect model responses to exact feedback, but do not consistently improve exact correction, and that recurrence of earlier outputs is linked to lower recoverability.

By Haitong Jiang, Chunlin Liu, Yile Wang, Yuhong Feng
arXiv AI
Oct 21

Correcting Attribution Patching with a Few Exact Measurements

The paper introduces Guided-AP, a method that improves attribution patching by using a few exact measurements to identify and correct its errors. On a single prompt, Guided-AP reduces error by up to 26% across three behaviors, while on many prompts it cuts average error on factual completion by 75% to 79%. The approach leverages shared error signs across components and tasks to efficiently guide verification and correction.

By Luyang Zhang, Minxing Zheng, Jialu Wang
arXiv AI
Aug 24

Knowing but Not Saying: Preventing Factual Access Failures in LLM SFT via Recall-Anchored Distillation

The paper identifies a specific issue in supervised fine‑tuning (SFT) of large language models called factual access failure, where models can recognize correct facts under constrained tests but fail to generate them in open‑ended settings. It demonstrates that SFT can cause both genuine wrong answers and expression‑level errors such as verbosity or formatting mismatches. To mitigate this, the authors propose Recall‑Anchored Distillation (RAD), a self‑distillation method that aligns the fine‑tuned model with the base model’s soft output distribution on unlabeled out‑of‑distribution text, thereby recovering lost factual recall without needing labeled data.

By Haodong Chen, Yadong Wang, Shengtao Wen, Dong Liang, Xiang Chen
arXiv AI
Oct 12

Inject a Turn, Not a Token: Learning Where to Redirect LLM Reasoning for Self-Correction

The paper introduces UPTURN, a method that improves large language model self‑correction by learning where to intervene in a failed reasoning rollout and injecting a new user turn to redirect the model, rather than fine‑tuning the LLM. A small 130K‑parameter MLP head is trained on a frozen encoder using only final‑answer correctness to select intervention points. Experiments on four reasoners and six math benchmarks show that UPTURN consistently raises accuracy over a 32‑rollout baseline and outperforms existing self‑correction and reflection‑token injection techniques while transferring to new tasks without retraining.

By Jiwon Jeon, Myungsik Cho, Jeonghye Kim, Youngchul Sung
arXiv Machine Learning
6d ago

Auditing the Judge: Internal Signals for LLM Reasoning Assurance

The paper investigates how internal signals from large language models can be used to improve error detection and budget allocation for answer review. By training a supervised correctness probe on attention-head activations, the authors show that this probe can identify 1.3–22.5 percentage points more errors at a 5% review budget across several 7–12B models on GSM8K and MATH, outperforming the model’s own self-judgment in many settings. While the probe improves in-domain allocation, its effectiveness drops when transferred across datasets, indicating that internal signals are not universally reliable confidence estimators.

By Mayank Gulati, Gerhard Wunder