REDE paper filters noisy reasoning traces to improve hallucination detection
SharonYixuanLi · x · 2026-07-29
REDE filters noisy reasoning traces to improve hallucination detection
Large reasoning models often produce long traces that include irrelevant or repetitive steps, and those noisy steps can obscure signals useful for hallucination detection.
The paper identifies two common noise types — irrelevant steps and repetitive steps — and shows that simple confidence scores or embedding-based filtering do not reliably separate noisy from informative reasoning steps. REDE addresses this by using final-answer attention as automatic supervision to shape step-level representations, making noisy steps easier to identify and remove.
In experiments across multiple reasoning benchmarks, REDE consistently improves hallucination detection over competitive baselines, and it can be plugged into different detectors by operating on the filtered trajectory.
Related event: REDE Denoises Reasoning Traces to Detect LLM Hallucinations(2 posts)→
More from Research
- A proposal to make KV cache portable across machines, data centers, and the WAN — knowrohit07 · 2026-07-29
- TransluceAI proposes oversight foundation models to catch reward hacking at scale — JacobSteinhardt · 2026-07-29
- Why tokenizers still resist end-to-end optimization despite years of pretraining — seanmcdonaldxyz · 2026-07-29
- Student builds a local coding agent on molab and beats two 7B coder baselines — S_Conradi · 2026-07-29
- Why tokenizer optimization is hard: expensive pretraining, slow feedback, and non-differentiable design — paul_cal · 2026-07-29
- TILT improves compositional text-to-image generation with a model-intrinsic reward — Debottam Dutta · 2026-07-29