Add PUMA: Semantic-Preserving Early Exit for Reasoning Models

[ZhishanQ]

Adding PUMA (arXiv 2026-05) to Efficient Reasoning during Inference → Length Budgeting at the top of the chronological list, alongside DEER ("Dynamic Early Exit in Reasoning Models").

PUMA is a plug-and-play inference-time early-exit framework for Large Reasoning Models. Unlike answer-level early-exit methods (DEER uses confidence; others use trial-answer consistency), PUMA introduces reasoning-level semantic redundancy as a complementary stopping signal — detected by a lightweight Redundancy Detector (fine-tuned Qwen3-Embedding-0.6B with a contrastive objective) — followed by an Answer Verification window confirming the candidate exit.

Results across 5 LRMs (DeepSeek-R1-Distill-Qwen-7B/14B/32B, Llama-3.1-Nemotron-Nano-8B, Qwen3-30B-A3B-Thinking) and 5 reasoning benchmarks (MATH-500, AIME24, AIME25, OlympiadBench, GPQA-Diamond): 26.2% average token reduction with preserved (slightly improved) accuracy; 1.40× / 1.28× wall-clock speedup on DS-7B/14B. Also generalizes to LiveCodeBench (code) and zero-shot VLM reasoning (MathVista, MathVision).

• Paper: https://arxiv.org/abs/2605.17672
• Code: https://github.com/giovanni-vaccarino/PUMA

Thanks!