Low overhead features improve detection of unsupported ai answers
FlashAttention with Statistics: Low-Overhead Attention Features for Compact Hallucination Detectors
Artificial Intelligence
Summary
Large language models sometimes give answers that don't match the facts they were given. The authors show how smaller, fixed models can spot these unsupported answers by looking at how the model pays attention to information. They developed a way to collect helpful information about the model's attention in a smarter, faster way that uses less memory. Their tests show that these attention-based features help detectors find false or made-up answers better than other existing methods.
What this means in practice
- •For ai developers: Build more efficient tools to catch unsupported answers produced by AI assistants by adding low-cost attention-based features to proxy detectors.
- •For cloud service engineers: Optimize memory usage and speed for AI response verification when deploying large language models with long input sequences.
Authors
Andrei Volodichev, Maksim Savichev, Nikita Sviridov, Artem Yuliy, Dmitry Simakov, Alexey Zaytsev
Abstract
Large language models can produce fluent answers that the supplied context does not support. A compact frozen proxy model can detect such answers by reading the context and another model's response, even from a proprietary generator. Attention-based detectors use summary statistics of the proxy's attention weights, but standard pipelines materialize the attention matrix, whose memory grows quadratically with sequence length. We compute these statistics without materializing attention, by accumulation in the FlashAttention pass, sparse lookups after it, or a second tiled pass over the scores, and classify them by this path into nearly free and costlier ones. On an H100, the statistics of five published detectors add 16% to the latency of a Qwen3-8B forward at 2K tokens, falling to 1.4% at 35K, and store 22.5 KiB per response token. On three benchmarks, probes trained on frozen Qwen3 proxies of 0.6B to 8B parameters prove strong detectors: a 4B proxy outperforms the strongest of nine zero-shot 4-9B LLM judges in F1 on two of the three, the proxies match or outperform specialized grounding checkers of up to 7B parameters, probing each proxy beats prompting it as a judge, and adding attention features significantly improves F1 in half of the 12 benchmark-proxy pairs, never significantly lowering it.