ACL 2026 · San Diego, USA · 2 July 2026

Current Advances in LLM Reasoning

A unified, hands-on tour of how well LLMs reason, how to make them reason better, and where the field is heading next.

Overview

As Large Language Models (LLMs) increasingly tackle reasoning-heavy tasks, from mathematics to commonsense to multilingual understanding, researchers face three pressing questions: How well do models reason? How can we make them reason better? And what are the next frontiers in LLM reasoning?

This tutorial answers these questions through a unified view of LLM reasoning. We explore comprehensive evaluation strategies to assess the reasoning abilities of models and discuss two families of methods that improve reasoning: advanced inference-time methods and post-training methods.

The tutorial is designed for both researchers and practitioners seeking actionable insight into LLM reasoning.

Outline

  1. Part 1

    How well can models reason?

    14:00–14:40 · 40 min

    Leads: Akhil Arora, Lars Klein, Julia Kreutzer

    Foundations for evaluating reasoning beyond accuracy. Multilingual, low-resource, and robustness benchmarks; evaluation dimensions including consistency, faithfulness, calibration, and variance-aware scoring. The hands-on session uses the accompanying code to assess reasoning variance across languages and seeds, compute calibration error, and visualize consistency metrics.

  2. Part 2

    How do we make models reason better?

    14:40–17:00

    2.1 Inference-time Reasoning Strategies 14:40–15:10 · 30 min

    Leads: Niket Tandon, Akhil Arora, Nearchos Potamitis

    Prompt-based, structured, and agentic reasoning under a single lens: iterative and self-refinement methods, structured prompting via graphs/trees, tool-augmented and agentic reasoning, and dynamic compute allocation. Hands-on: compare iterative, self-consistent, and agentic strategies under latency budgets.

    2.2 Post-training & RL-based Reasoning 15:10–15:30 & 16:00–16:20 · 40 min

    Leads: Nouha Dziri, Vishrav Chaudhary

    Reinforcement-learning and preference-based post-training (RLHF, DPO, reward modeling) for robustness and controllability. We relate these to inference-time strategies and cover recent alignment variants. Hands-on: how reward-model calibration influences reasoning confidence and abstention. Continues after the coffee break.

    ☕ Coffee Break — 15:30–16:00 · 30 min

    2.3 Hands-on Exercises & Q&A 16:20–17:00 · 30 min + 10 min Q&A

    Guided hands-on exercises using the accompanying code, followed by open Q&A.

  3. Part 3

    What are the next frontiers in reasoning?

    17:00–17:30 · 30 min

    3.1 Efficient Reasoning 10 min

    Leads: Akhil Arora, Niket Tandon

    Sustainable, cost-effective reasoning via adaptive compute reuse, caching, batching, and scalable orchestration.

    3.2 Reasoning in High-Stakes Domains 10 min

    Leads: Lars Klein

    From "reason as hard as possible and guess" to calibrated abstention: training and evaluating models to refuse when in doubt, prioritizing safety and factual reliability.

    3.3 Low-Resource & Multilingual Reasoning 10 min

    Leads: Julia Kreutzer, Akhil Arora

    Advances and open challenges in multilingual and culturally grounded reasoning — cross-lingual transfer, low-resource adaptation, and inclusive benchmark design.

Speakers

Materials

Citation

@inproceedings{arora2026current,
  title     = {Current Advances in LLM Reasoning},
  author    = {Arora, Akhil and Chaudhary, Vishrav and Kreutzer, Julia and Potamitis, Nearchos and Dziri, Nouha and Tandon, Niket},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 5: Tutorial Abstracts)},
  pages     = {7--8},
  year      = {2026}
}