Add PUMA (Preprint'26) + create 2026 section under Reasoning in LLMs

[ZhishanQ]

Two small changes to 🔤 Reasoning in Large Language Models:

1. Create new `### 2026` year section at the top (above the existing `### 2025`), matching the year-ordered structure used throughout the repo.
2. Add PUMA as the inaugural 2026 entry, following the existing format (**[Title.](url)** [[code](url)] then *Authors.* Venue'YY).

PUMA (Stop When Reasoning Converges: Semantic-Preserving Early Exit for Reasoning Models) is a plug-and-play inference-time early-exit framework for Large Reasoning Models that uses reasoning-level semantic redundancy as a complementary stopping signal to answer-level confidence/consistency. Architecture: lightweight Redundancy Detector (fine-tuned Qwen3-Embedding-0.6B) + Answer Verification window + Loop Breaker fallback.

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!