REDE denoises reasoning traces to boost hallucination detection in large reasoning models
NanyangTechnologicalUniversity · hf · 2026-07-28
REDE cleans noisy reasoning traces to improve hallucination detection
Large reasoning models often produce long chains of thought that contain irrelevant or repetitive steps, which makes hallucination detection harder. The paper identifies these two noise types and shows they can significantly hurt detection accuracy.
To solve this, the authors propose REDE, a framework that uses final-answer attention as supervision to reshape step-level representations. After filtering noisy steps, REDE can be plugged into different hallucination detectors. Across multiple reasoning benchmarks, it consistently improves performance over strong baselines.
Related event: REDE Denoises Reasoning Traces to Detect LLM Hallucinations(2 posts)→
More from Research
- Tsinghua professor's ApexIntelligence raises $50M to build self-evolving foundation models — DeryaTR_ · 2026-10-16
- Fluids Don't Compute—They Contribute the Nonlinear Step, Sandberg Argues — anderssandberg · 2026-10-16
- Anders Sandberg on fluid computation: error budgets tighten quadratically over time — anderssandberg · 2026-10-16
- How the fluid computation works: force field advects particle to 0/1 boxes — anderssandberg · 2026-10-16
- FlashAttention creator Tri Dao joins Radical Numerics as scientific advisor — exnx · 2026-10-16
- SAIR Launches Open Math OS, Offering Free Compute and 100+ Tools for Community-Owned Math AI — Thom_Wolf · 2026-10-16