Improve model card: add metadata and refine description

README.md

@@ -1,12 +1,36 @@
 ---
 license: apache-2.0
+pipeline_tag: text-generation
+library_name: transformers
 ---
 
+# LYNX: Learning Dynamic Exits for Confidence-Controlled Reasoning
+
 <!-- Badges -->
 [![arXiv](https://img.shields.io/badge/arXiv-2512.05325-b31b1b.svg)](https://arxiv.org/abs/2512.05325)
 [![GitHub](https://img.shields.io/badge/GitHub-Code-black?logo=github)](https://github.com/farukakgul/LYNX)
 
-<div style="font-size: 15px; line-height: 1.55;">
+LYNX turns a reasoning model’s own hidden states into **confidence‑controlled early exits**. At naturally occurring cue tokens (e.g., `hmm`, `wait`, `alternatively`), LYNX:
+1. Extracts features from a few intermediate layers.
+2. Uses a lightweight probe to predict whether the final answer will be correct if we stop now.
+3. Wraps the probe with split conformal prediction to get a **user‑tunable confidence level** and explicit guarantees.
+
+This repository contains a minimal, self‑contained implementation of that pipeline for open‑weight LMs (e.g., DeepSeek‑R1‑1.5B, QwQ‑32B, and Llama‑3.1‑Nemotron‑8B), featuring an HF‑only pipeline for training, calibration, and evaluation.
+
+For more details, refer to the [paper](https://huggingface.co/papers/2512.05325) and the [GitHub repository](https://github.com/farukakgul/LYNX).
+
+## Citation
+
+If you find LYNX useful, please cite the accompanying paper:
 
-Large reasoning models achieve strong performance on complex tasks by generating extended chains of thought, but they often “overthink”: continuing to reason long after they internally have enough information to answer correctly. This wastes inference-time compute and can even hurt accuracy. Existing attempts to stop early either manipulate decoding with extra sampling and heuristics, rely on auxiliary verifier models, or operate only as post-hoc analysis pipelines without formal guarantees. We introduce LYNX, an online early-exit mechanism that turns a model’s own hidden-state awareness into confidence-controlled stopping decisions. LYNX attaches exit decisions to naturally occurring reasoning cues (e.g., “hmm”, “wait”) during generation, trains a lightweight probe on hidden states at those cue tokens using supervision from forced exits, and wraps the resulting scores in split conformal prediction to obtain distribution-free control over the rate of premature exits. Crucially, we train and calibrate this probe once on a generic mathematical corpus and then reuse it unchanged across benchmarks, decoding temperatures, and even non-mathematical tasks. Across three model families spanning 1.5B to 32B parameters (DeepSeek-R1-1.5B, QwQ-32B, and Llama-3.1-Nemotron-8B), a single mathematically trained probe per base model yields strong accuracy–efficiency tradeoffs. On GSM8K, LYNX matches or improves baseline accuracy while reducing tokens by 40–65%; on MATH-500 it improves accuracy by up to 12 points with roughly 35–60% fewer tokens; on AIME 2024 it recovers baseline accuracy with more than 50% token savings; and on CommonsenseQA, a non-math benchmark, it transfers zero-shot with modest accuracy gains and up to 70% fewer tokens. Compared to state-of-the-art early-exit methods, LYNX offers competitive or superior Pareto frontiers while remaining fully online, requiring no proxy models at inference, and providing explicit, user-tunable confidence guarantees. Code is available at https://github.com/farukakgul/LYNX.
-</div>
+```bibtex
+@misc{akgül2025lynxlearningdynamicexits,
+      title={LYNX: Learning Dynamic Exits for Confidence-Controlled Reasoning},
+      author={Ömer Faruk Akgül and Yusuf Hakan Kalaycı and Rajgopal Kannan and Willie Neiswanger and Viktor Prasanna},
+      year={2025},
+      eprint={2512.05325},
+      archivePrefix={arXiv},
+      primaryClass={cs.CL},
+      url={https://arxiv.org/abs/2512.05325},
+}
+```