Language models can learn facts while misjudging what they know

The authors report that explicitly retraining answerability judgments helps keep them aligned with a language model’s changing knowledge.

Chinese Tech

Zichuan Fu · Xian Wu · Jingtong Gao · Wenlin Zhang · Jiaxuan Li · Binhao Wang · +6 more

City University of Hong Kong · Tencent Jarvis Lab · Xi’an Jiaotong University · Renmin University of China · University of Science and Technology of China

Research Digest·4 November 2026·2 min read

Illustration: Language models can learn facts while misjudging what they know

In one line

Large language models need explicit calibration of answerability judgments after knowledge updates to avoid confident wrong answers.

Before you read this

Three ideas this paper builds on

confidence calibrationConfidence calibration aligns a model's predicted confidence with its actual accuracy across predictions.76 papers →large language modelA neural network trained on massive text corpora to generate coherent text and perform diverse language tasks.261 papers →continual learningA learning paradigm where a model sequentially acquires knowledge from multiple tasks without significantly forgetting earlier ones.56 papers →

Read it as

When a (LLM) learns new facts, it might still confidently answer questions it cannot really answer, or refuse questions it now can. This mismatch between what the model knows and what it thinks it knows is a problem for reliable use.

Researchers trained two separate parts of the model: one that holds facts (knowledge) and one that judges if it can answer (feeling of knowing). They updated the knowledge part with corrected answers, then retrained the judgment part to match the new knowledge. Finally, they trained the model to check its own judgment before responding.

The main finding is that explicitly aligning the judgment with updated knowledge improves the model's accuracy in deciding whether to answer or not. This matters because it helps make LLMs more trustworthy when their knowledge changes over time.

How it works

before and after

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

Concepts

confidence calibration large language model supervised fine-tuning continual learning reinforcement learning low-rank adaptation