Research map / Large language models / Reasoning and chain of thought
FROST: Filtering Reasoning Outliers with Attention for Efficient Reasoning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes FROST, which replaces softmax with softmax_1 in attention and applies LoRA-based SFT to suppress low-attention "reasoning outlier" steps, yielding shorter reasoning traces on math benchmarks.
Problem
Large reasoning models generate many uncritical reasoning steps (redundant self-verification), causing inefficient and sometimes inaccurate long reasoning traces.
Approach
Defines "reasoning outliers" as low-attention, low-contribution reasoning steps and uses the softmax_1 (off-by-one softmax) attention function during LoRA-based SFT so attention suppresses these low-weight steps while keeping high-weight ones.
Key results
- FROST reduces average token usage by 69.68% and improves average accuracy by 26.70% over base models across Phi-4-Reasoning, GPT-oss-20B and Magistral-Small-1.1.
- On AIME24 with Phi-4-Reasoning, FROST reduces maximum infinity norm by 15.97% and average kurtosis by 91.09% versus base.
- On Phi-4-Reasoning, FROST reaches 0.9311 Pass@1 on GSM8K with 154.33 tokens versus base 0.9242 with 1017.70 tokens.
Surprising findings
- Sparsemax and Entmax, which also sharpen attention, do not significantly degrade accuracy and sometimes exceed the base model, despite FROST's advantage in some settings.
- On Minerva, Entmax attains higher accuracy than FROST at similar token count.
Limitations
- Restricted to mathematical reasoning in main experiments; coding and other domains are future work.
- Relies only on SFT without GRPO or RL.
- Some low-attention steps that are important may be pruned, causing lower accuracy than TALE on some tasks.
Experimental setup
Datasets: OpenR1 GSM8K MATH500 AIME24 Minerva LeetCode LiveCodeBench UGPhysics
Benchmarks: GSM8K MATH500 AIME24 Minerva LeetCode LiveCodeBench UGPhysics
Metrics: pass@1 token usage maximum infinity norm average kurtosis average sentence entropy training time inference time
Baselines: TALE DRP SelfBudgeter ThinkLess Sparsemax Entmax Softmax
Builds on: softmax_1 (outlier-free attention) TALE DRP ThinkLess LoRA
Closest papers by meaning
- Mitigating Overthinking in Large Reasoning Language Models via Reasoning Path Deviation Monitoring (NeurIPS 2026 · similarity 0.88)
- Stop When Further Reasoning Won’t Help: Attention-State Adaptive Generation in Reasoning Models (ICML 2026 · similarity 0.87)
- How Far Are We from Optimal Reasoning Efficiency? (NeurIPS 2025 · similarity 0.87)
- Efficient Reasoning via Constrained Optimization in Latent Space (NeurIPS 2026 · similarity 0.87)
- Hint Tuning: Less Data Makes Better Reasoners (NeurIPS 2026 · similarity 0.87)
- OS-Pruner: Pruning Chains-of-Thought of Reasoning Models via Optimal Stopping (NeurIPS 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.