Research map / Large language models / Reasoning and chain of thought

CyclicReflex: Improving Reasoning Models via Cyclical Reflection Token Scheduling

ICLR 2026 · Large language models › Reasoning and chain of thought · approach: Efficient and adaptive thinking · 0 citations (OpenAlex)

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 CyclicReflex, a training-free decoding method that adds a bidirectional triangular-wave logit bias to reflection tokens ("wait", "but") to balance under- and over-reflection in large reasoning models.

Problem

Large reasoning models suffer from under-reflection (premature answers) and over-reflection (looping on reflection tokens), and existing fixed-penalty decoding (TIP) helps some problems while hurting others.

Approach

Treat reflection tokens as a schedulable resource and analogize to cyclical learning-rate schedules: add a position-dependent triangular waveform to reflection-token logits, promoting them early and suppressing them later, at no extra compute cost.

Key results

• On DeepSeek-R1-Distill-Llama-8B, CyclicReflex raises AIME2024 accuracy from 0.42 (original) to 0.53 with comparable generation length.
• On DeepSeek-R1-Distill-Qwen-7B, CyclicReflex improves AMC2023 accuracy from 0.81 to 0.90 and MATH500 from 0.86 to 0.89.
• Across Qwen3-4B/8B/14B, CyclicReflex outperforms Original and TIP on all five benchmarks, e.g., AIME2024 0.43 vs 0.37 original for Qwen3-8B.

Surprising findings

• TIP, which penalizes reflection tokens, lowers accuracy on several settings (e.g., AIME2025 for Llama-8B drops 3 points) and on Easy/Medium MATH500 problems.
• Positive-only reflection boosting and random noise on reflection logits both underperform standard TIP.

Limitations

• Hyperparameters (amplitude A, period C) are tuned by grid search on the evaluation benchmarks, with separate ranges per dataset.
• Evaluation uses a fixed reflection-token list and rule-based answer matching; gains are a few percentage points on small benchmarks like AIME.

Experimental setup

Datasets: MATH500 AIME2024 AIME2025 AMC2023 GPQA Diamond LiveCodeBench

Benchmarks: MATH500 AIME2024 AIME2025 AMC2023 GPQA Diamond LiveCodeBench

Metrics: accuracy generation length pass@N

Baselines: Original decoding TIP S1 Best-of-N Beam Search Linear decay Random adjustment

Builds on: TIP (thought switching penalty) cyclical learning rates silver stepsize schedule landscape of thoughts

Closest papers by meaning

• Stop Unnecessary Reflection: Training LRMs for Efficient Reasoning with Adaptive Reflection and Length Coordinated Penalty (ICLR 2026 · similarity 0.91)
• Efficient Reasoning for Large Reasoning Language Models via Certainty-Guided Reflection Suppression (AAAI 2026 · similarity 0.88)
• Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach (NeurIPS 2025 · similarity 0.88)
• MM-HELIX: Boosting Multimodal Long-Chain Reflective Reasoning with Holistic Platform and Adaptive Hybrid Policy Optimization (ICLR 2026 · similarity 0.88)
• SRPO: Self-Reflective Policy Optimization for Long-Horizon Reasoning (ICML 2026 · similarity 0.87)
• Reasoning Can Be Restored by Correcting a Few Decision Tokens (ICML 2026 · similarity 0.87)