LYNX: Learning Dynamic Exits for Confidence-Controlled Reasoning
Ömer Faruk Akgül, Yusuf Hakan Kalaycı, Rajgopal Kannan, Willie Neiswanger, Viktor Prasanna
2025-12-10
Summary
This paper introduces a new method called LYNX that helps large language models (like those used for answering complex questions) become more efficient. These models often spend too much time 'thinking' even after they already know the answer, which slows them down and can sometimes make them less accurate.
What's the problem?
Large language models are really good at solving hard problems by breaking them down into steps, but they frequently continue reasoning even *after* they have enough information to give the correct answer. This is a waste of computing power and can actually lower the final accuracy. Previous attempts to stop these models early were either complicated, needed extra models to check the answers, or couldn't guarantee they were stopping at the right time.
What's the solution?
LYNX works by teaching the model to recognize its own 'thinking cues' – things like the words 'hmm' or 'wait' that appear during its reasoning process. It then trains a small, separate component to look at the model's internal state when these cues appear and predict how confident the model is in its answer. This confidence score is used to decide whether to stop the reasoning process early. Importantly, this 'confidence checker' is trained on math problems but can then be used on *any* type of problem without needing to be retrained. It uses a technique called 'split conformal prediction' to ensure it doesn't stop too early and give a wrong answer.
Why it matters?
LYNX is important because it makes large language models much more efficient without sacrificing accuracy. It can significantly reduce the number of steps a model takes to solve a problem, saving time and computing resources. It’s also versatile, working well on math problems, general knowledge questions, and even tasks it wasn’t specifically trained for. Plus, it provides a way to control how confident the model needs to be before stopping, giving users more control over the balance between speed and accuracy.
Abstract
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 have enough information to answer correctly. This wastes inference-time compute and can 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 premature exits. Crucially, we train and calibrate this probe once on a generic mathematical corpus and reuse it unchanged across benchmarks, decoding temperatures, and even non-mathematical tasks. Across three model families spanning 1.5B to 32B parameters, 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.