Language models use internal signals to reduce factual mistakes
Look Before You Leap: Factual Decoding with Internal Attribution Signals
Computation and LanguageArtificial Intelligence
Summary
Large language models sometimes make up wrong facts while writing, causing errors to build up quickly. The authors found that certain internal parts of these models show signs when they start making these mistakes. They created a method called DescaPE that watches these signals during writing to avoid risky, likely wrong continuations. Their tests show this improves the model’s accuracy with only a small speed cost.
What this means in practice
- •For chatbot developers: Integrate DescaPE to reduce the likelihood of hallucinated facts in chatbot responses for more reliable user interactions.
- •For content generation teams: Apply internal factuality signals during text generation to improve the trustworthiness of AI-written articles and reports.
Authors
Hayeong Ryu, JungMin Yun, Byeonggeuk Lim, Sunhee Jo, YoungBin Kim
Abstract
Hallucination remains a critical challenge in large language models (LLMs), where early factual errors compound through autoregressive generation in a snowballing effect that neither post-hoc correction nor weight-level intervention can effectively preempt. We propose DescaPE (DEcoding Signal Control Against Path Error-snowballing), a decoding framework that leverages internal model signals to suppress hallucination-prone trajectories at inference time. Through sliding-window MLP ablation, we identify a factual-salient layer span within LLMs whose derived signal is selectively elevated for factual tokens and exhibits anomalous spikes at hallucination-prone steps. We train a lightweight probe to approximate this signal from a single forward pass and integrate it into candidate scoring to penalize high-risk continuations while rewarding factually grounded ones. Experiments across five factuality benchmarks on three LLMs demonstrate that DescaPE achieves factuality improvements over decoding-time baselines in multiple settings, while incurring only 1.10x latency overhead in our efficiency evaluation. Our code is available at https://github.com/hayeonggg/DESCAPE.