💬 NLP

FROST: Filtering Reasoning Outliers with Attention for Efficient Reasoning

FROST is an attention-aware method that filters reasoning outliers to prune uncritical paths, significantly improving both the efficiency (69.68% token reduction) and accuracy (26.70% gain) of reasoning models across multiple benchmarks.

Original authors: Haozheng Luo, Zhuolin Jiang, Md Zahid Hasan, Yan Chen, Soumalya Sarkar

Published 2026-02-25

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Haozheng Luo, Zhuolin Jiang, Md Zahid Hasan, Yan Chen, Soumalya Sarkar

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a brilliant but overly chatty student named Reasoning-Model. When you ask this student a hard math problem, they don't just solve it; they write a whole novel about it. They start with the solution, then doubt themselves, then re-calculate, then check their work, then check their check, and then wander off to think about the history of mathematics before finally circling back to the answer.

While this "overthinking" sometimes helps, it often wastes a massive amount of time (and computer energy) on steps that don't actually help solve the problem. This is what the paper calls "Reasoning Outliers."

The authors introduce a new method called FROST (Filtering Reasoning Outliers with Attention) to help this student think clearly and efficiently. Here is how it works, using simple analogies:

1. The Problem: The "Noisy Room"

Imagine the student's brain is a crowded room where 100 people are talking at once.

Critical thoughts (the actual math steps) are like a few people shouting the right answer.
Outliers (redundant self-checks, "wait, let me try again," or repetitive loops) are like the other 95 people whispering, coughing, or repeating the same sentence over and over.

The student tries to listen to everyone, which makes them slow and confused. They might even get the answer wrong because they got distracted by the noise.

2. The Solution: The "Smart Filter" (FROST)

FROST acts like a super-smart sound engineer in that room. It has a special tool called Softmax1 (a mathematical tweak to how the model pays attention).

How it works: Instead of listening to everyone equally, FROST looks at the "volume" (attention weight) of each thought.
The Magic: It realizes that the critical steps are loud and clear, while the "outliers" are quiet and repetitive.
The Action: FROST turns the volume down on the quiet, useless chatter until it's silent, while keeping the loud, important steps crystal clear.

Think of it like using a noise-canceling headphone that only blocks out the background hum but lets the music play perfectly.

3. The Result: A Faster, Smarter Student

After applying FROST, the student changes dramatically:

Shorter Stories: Instead of writing a novel, they write a concise, clear paragraph.
Better Accuracy: Because they aren't distracted by their own internal monologue, they actually get the right answer more often.
Less Energy: The computer doesn't have to process thousands of useless words.

The Real-World Impact

The researchers tested this on some of the smartest AI models available (like Phi-4 and GPT-OSS). The results were impressive:

69% Less Talking: The models used 70% fewer words (tokens) to solve problems.
27% Smarter: Despite talking less, they got the right answer more often than before.
Speed: The models solved problems much faster because they didn't waste time on the "noise."

In a Nutshell

FROST is a technique that teaches AI models to stop "overthinking." It filters out the mental clutter and redundant self-doubt, allowing the AI to focus only on the steps that actually matter. It's the difference between a student who panics and scribbles everything down, and a calm expert who knows exactly what to write to get the job done.

1. Problem Statement

Large Reasoning Models (LRMs) have demonstrated exceptional performance in complex tasks like mathematical problem-solving, coding, and scientific QA. However, they suffer from "overthinking," generating excessive reasoning chains filled with irrelevant steps, redundant self-verification, and repetitive loops. These uncritical steps, termed reasoning outliers, lead to:

High Latency and Cost: Significant token usage increases inference time and computational costs.
Reduced Reliability: Long reasoning paths increase the probability of hallucinations and logical errors.
Inefficiency of Current Solutions:
- Token-level pruning (e.g., TALE) risks removing essential reasoning steps because reasoning is naturally sentence-based.
- Sentence-level iterative refinement (e.g., DRP, GRPO-S) often incurs high computational overhead and latency.
- Prompt-based methods are unreliable on complex problems.

2. Methodology: FROST

The authors propose FROST (Filtering Reasoning Outliers with Attention), an attention-aware method designed to prune uncritical reasoning paths at the sentence level while preserving critical ones.

Core Concepts

Reasoning Outliers: Defined as reasoning steps characterized by low attention weights and low sentence entropy. These steps contribute negligibly to the final answer prediction, often representing self-checking or repetition.
Softmax1 Activation: Instead of the standard Softmax, FROST replaces the attention mechanism with Softmax1.
- Formula: $\text{Softmax1}(x_i) = \frac{\exp(x_i)}{\sum_j \exp(x_j) + 1}$
- Mechanism: This modification effectively drives low attention weights toward zero (suppressing outliers) while preserving or slightly enhancing high attention weights (critical steps). It acts as a "tail contraction" mechanism.

Training Strategy

Supervised Fine-Tuning (SFT): FROST integrates Softmax1 into the attention layers of pre-trained LRMs (e.g., Phi-4-Reasoning, GPT-oss-20B).
Efficiency: Unlike previous methods requiring training from scratch or complex reinforcement learning, FROST achieves effective outlier removal with only a few steps of fine-tuning using LoRA (Low-Rank Adaptation).
Objective: The model is trained on math problems with detailed reasoning steps, learning to focus attention on critical sentences and suppress low-weight outliers during the generation process.

3. Theoretical Analysis

The paper provides theoretical proofs supporting the efficacy of Softmax1:

Sentence-Level Suppression: The authors prove that Softmax1 satisfies "tail contraction," meaning it reduces the relative dominance of outliers in the attention distribution at the sentence level.
Deployment-Time Impact: Theoretical bounds show that sentences with low attention weights ( $\alpha_i \le \epsilon$ ) have a bounded, negligible contribution to the final output logits. This ensures that low-attention sentences are effectively skipped during inference without disrupting the model's reasoning capacity.
Preservation of Capacity: The analysis demonstrates that removing these outliers does not degrade the model's ability to solve problems; rather, it sharpens the focus on critical reasoning traces.

4. Key Contributions

Concept of Reasoning Outliers: Introduced a formal definition for uncritical reasoning steps based on attention weights and entropy, identifying them as the primary cause of inefficiency in LRMs.
FROST Framework: Developed an attention-based pruning mechanism using Softmax1 that operates at the sentence level, avoiding the pitfalls of token-level pruning.
Theoretical Guarantees: Provided mathematical proofs showing that Softmax1 suppresses outliers while maintaining or enhancing reasoning capacity.
Efficient Training Strategy: Demonstrated that combining Softmax1 with lightweight LoRA-based SFT yields efficient reasoning models without the need for expensive Reinforcement Learning (RL) or full fine-tuning.

5. Experimental Results

The authors evaluated FROST on four mathematical benchmarks (GSM8K, MATH500, AIME24, Minerva) using three backbone models (Phi-4-Reasoning, GPT-oss-20B, Magistral-Small-1.1).

Performance Gains:
- Accuracy: Achieved an average 26.70% improvement in Pass@1 accuracy compared to base models.
- Efficiency: Reduced token usage by an average of 69.68%, significantly shortening reasoning trajectories.
- Comparison: Outperformed state-of-the-art baselines like TALE, DRP, SelfBudgeter, and ThinkLess in the trade-off between accuracy and token efficiency.
Outlier Metrics:
- Reduced the maximum infinity norm ( $\|x\|_\infty$ ) of activation tensors by 15.97%.
- Reduced the average kurtosis (a measure of outlier presence) by 91.09%.
- Increased average sentence entropy by 13.28%, indicating a shift toward more critical, high-entropy reasoning steps.
Generalization: The method generalized well to out-of-domain tasks (coding and physics), maintaining or improving performance on LeetCode, LiveCodeBench, and UGPhysics.
Time Efficiency: FROST reduced training time by 42.2% and inference time by at least 28.6% compared to other SFT baselines.

6. Significance

FROST represents a significant advancement in making Large Reasoning Models practical for real-world deployment. By addressing the "overthinking" phenomenon through a theoretically grounded, attention-based pruning mechanism, it offers a solution that is:

Computationally Efficient: Drastically reduces inference costs and latency.
Reliable: Improves accuracy by removing noise and hallucination-prone redundant steps.
Accessible: Requires only lightweight fine-tuning (LoRA), making it feasible for organizations without massive compute resources to optimize their reasoning models.

The paper concludes that FROST successfully transforms LRMs from "over-thinkers" into efficient, precise reasoners, setting a new standard for balancing reasoning depth with computational efficiency.