← Back to all news
arXiv AI

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.

Related stories

arXiv AI
Sep 21

GUARD: Natural Forgetting in Large Reasoning Models via Guided Answer-Reasoning Distillation

The paper introduces GUARD, a method for natural forgetting in large reasoning models that transforms unsafe disclosures into safe-exit trajectories using guided answer‑reasoning distillation. It aligns a frozen model with guidance tokens and distills this behavior into the parameters, aiming for a coherent, non‑disclosing chain of thought followed by a refusal‑style answer. The authors also propose the Natural Forgetting Reasoning Score (NFRS) to evaluate structural stability, fluency, and unsupported substitutes, and demonstrate GUARD’s effectiveness on R‑TOFU and a STAR‑1‑derived harmful‑intent setting.

By Zeyu Yan, Guanghao Zhou, Minghui Qiu, Ming Gao, Cen Chen
efficiencysafety
More like this →
arXiv AI
Sep 25

TWIST: A Proposed Benchmark for Intervention Quality in Conversational Memory, with a Human-Validated Draft-Alignment

TWIST is a new benchmark suite designed to evaluate the quality of interventions in conversational memory systems, focusing on how well these systems act correctly at belief change points. It includes four tracks—tension detection, draft vetting, belief-consistent answering, and sensitive recall governance—each paired with hard-negative controls to prevent gaming. The benchmark is rigorously validated through double annotation, adjudication, and a separability audit, revealing that current models face trade-offs between contradiction recall and false-positive rates.

By Subrat Panda
ragbenchmarkssafety
More like this →
arXiv AI
Sep 18

Correct Now, Insufficient Later: Auditing Update Sufficiency in Context Compression

The paper investigates how memory systems can answer a current query correctly yet fail to retain distinctions needed for later updates. Using a paired‑history audit, the authors evaluate 24 history pairs across six synthetic mechanisms and two model backends, achieving perfect reveal accuracy on DeepSeek and high accuracy on GLM. Record‑level audits reveal specific failures in structured reveal memories and frontier late‑reference adequacy, and the authors test a label‑equivariant repair that only partially restores correctness.

By Guangzhe Zhang
More like this →
arXiv AI
Sep 16

ReDraft, Don't Just Distill: Reference-Driven Revision for Continual VLLM Post-Training

ReDraft is a reference‑driven revision method for continual post‑training of large multimodal language models. It uses the model’s own incorrect outputs as references, revises them, verifies the revisions, and fine‑tunes on the accepted ones, thereby combining explicit supervision with policy proximity. On tasks such as Counting, Clock Reading, and Jigsaw, ReDraft outperforms standard supervised fine‑tuning and on‑policy methods, achieving higher target‑task gains while dramatically reducing forgetting.

By Zhihao Zhang, Mingqi Wu, Qiaole Dong, Enyu Zhou, Shuo Li, Boyang Liu, Jiazheng Zhang, Honglin Guo, Xin Guo, Shaofan Liu, Junzhe Wang, Dingwei Zhu, Zhiheng Xi, Minlong Peng, Yuan Hua, Qi Zhang, Tao Gui, Xuanjing Huang
fine-tuningefficiencymultimodal
More like this →
arXiv AI
Sep 2

Why Fine-Tuning Encourages Hallucinations and How to Fix It

The paper investigates why supervised fine‑tuning (SFT) of large language models leads to increased hallucinations of factual information. It proposes a self‑distillation SFT approach that regularizes output‑distribution drift to preserve pre‑training knowledge, and shows that freezing parameter groups can reduce hallucinations when new knowledge is unnecessary. Experiments attribute the main cause to interference among overlapping semantic representations, which self‑distillation mitigates, and an associative‑memory model explains the forgetting dynamics.

By Guy Kaplan, Zorik Gekhman, Zhen Zhu, Lotem Rozner, Yuval Reif, Swabha Swayamdipta, Derek Hoiem, Roy Schwartz
llmsfine-tuningefficiencysafety
More like this →