ICML 2026 · Seoul, South Korea · 5 July 2026

Adaptive Reasoning in LLMs: From Post-Training to Test-Time Learning

How LLMs learn to reason adaptively, from post-training methods to test-time learning that scales compute where it matters.

Overview

Reasoning in Large Language Models (LLMs) is increasingly understood not as a single forward pass but as an adaptive control loop, a process that decides how much to think, when to act, and how to revise. This tutorial traces that idea from classical search in AI through Chain-of-Thought prompting, building a shared vocabulary for reasoning as adaptation.

We then examine three practical mechanisms that make this loop adaptive: post-training as offline adaptation, test-time training and continual adaptation during inference, and agentic reasoning that adapts through external actions such as tools and environments.

Finally, we turn critical, looking at where adaptive reasoning fails and why, the open challenges in evaluating it, and the open problems that define the road ahead. The tutorial is aimed at researchers and practitioners who want a structured, current view of how LLMs learn to reason adaptively.

Outline

  1. Part 1

    Foundations of Adaptive Reasoning

    13:30–14:05 · 35 min

    Why reasoning benefits from adaptation, and the vocabulary we use to describe it.

    1.1 Motivation: why static reasoning fails 13:30–13:45 · 15 min

    Leads: Akhil Arora

    Why static prompting is often insufficient for complex, long-horizon reasoning, and when adaptive control helps or breaks down. Motivating settings include long-horizon reasoning, robustness under distribution shift, and interaction with external tools and environments.

    1.2 Concepts: reasoning as a control loop 13:45–14:05 · 20 min

    Leads: Akhil Arora

    Feedback, state, and adaptation as a unifying view across three strands of progress that have advanced in parallel, often with different assumptions, objectives, and evaluation practices: post-training refinement, test-time learning, and agentic inference. We ask what is fundamentally driving the gains, how the approaches relate, and where their limits lie.

  2. Part 2

    Control Mechanisms in Practice

    14:05–15:05 · 45 min

    Three practical mechanisms for adapting the reasoning control loop, looking at how adaptation operates at training and inference time and what feedback signals enable or hinder effective reasoning control.

    2.1 Post-training refinement as offline adaptation 14:05–14:20 · 15 min

    Leads: Vishrav Chaudhary

    Shaping the model's reasoning behaviour offline, before deployment.

    2.2 Test-time training & continual adaptation 14:20–14:35 · 15 min

    Leads: Vishrav Chaudhary

    Adapting the model on the fly, with continual adaptation during inference.

    ☕ Break / Q&A · 14:35–14:50 · 15 min

    2.3 Agentic reasoning: coordination, delegation & control 14:50–15:05 · 15 min

    Leads: Akhil Arora

    Adaptation through external action: coordination, delegation, and control across agentic systems.

  3. Part 3

    Failures, Evaluation, and Open Problems

    15:05–15:45 · 40 min

    Where adaptive reasoning breaks down, how to evaluate it, and what remains open.

    3.1 Credit assignment, instability & agent failure modes 15:05–15:20 · 15 min

    Leads: Vishrav Chaudhary

    A critical discussion of how adaptive and agentic systems fail: instability under continual adaptation, challenges in credit assignment, and mismatches between optimization signals and reasoning quality.

    3.2 Evaluating adaptive reasoning beyond accuracy 15:20–15:30 · 10 min

    Leads: Akhil Arora

    A critical discussion of evaluation challenges, and open questions in evaluation and deployment as adaptive reasoning systems approach the limits of existing benchmarks.

    3.3 Open challenges & future directions 15:30–15:45 · 15 min · Q&A

    Leads: Akhil Arora & Vishrav Chaudhary

    Closing discussion of open research directions, including verification and world-model approaches, followed by open Q&A.

Team

Materials

Citation

@inproceedings{arora2026adaptive-icml,
  title     = {Adaptive Reasoning in LLMs: From Post-Training to Test-Time Learning},
  author    = {Arora, Akhil and Chaudhary, Vishrav},
  booktitle = {Tutorials at the 43rd International Conference on Machine Learning (ICML)},
  year      = {2026}
}