TRAAC paper: adaptive thinking budgets fix under- and overthinking in reasoning models
EliasEskin · x · 2026-10-06
- Problem: Reasoning models are under-adaptive — they underthink on hard problems (early termination causes errors) and overthink on easy ones, wasting tokens.
- Method: TRAAC combines difficulty calibration (dynamic thinking-budget allocation) with attention-based compression to prune redundant reasoning steps.
- Results: On AIME, AMC, GPQA-Diamond and BBEH, TRAAC built on Qwen3-4B improves average absolute accuracy while cutting test-time compute.
- Accepted at COLM 2026 (poster session Oct 6), code open-sourced; authors from UNC and collaborators.
More from Research
- New paper probes multimodal MoE expert structure to make model adaptation much faster — georgiagkioxari · 2026-10-06
- Developer Runs 1,200+ Blender Modeling Sessions to Blind-Rank LLMs and Agent Harnesses — Izolight · 2026-10-06
- FOVEATED Fixes Context Reliance in Unstructured Knowledge Editing Across 5 LLM Editors — Ding Wu · 2026-10-06
- Software-in-the-loop: self-supervised scaling of terminal environments lifts Terminal-Bench 2 to 53.56% — Zhongzhi Li · 2026-10-06
- CC-Bench at COLM 2026 finds LLMs still default to stereotypes over implicit cultural cues — MaartenSap · 2026-10-06
- UCLA PhD's LLM agents produce 126K-line Lean 4 proof of MIP* = RE core theorem in 63 days — siyan_zhao · 2026-10-06