Call for papers

FoRLM 2025 submission guidances

We welcome contributions that advance the scientific foundations of reasoning in language models, exemplified by

  • Rich insights driven by rigorous experimentation and mathematical analyses (including negative results)
  • Well-tested frameworks, interpretable testbeds, and theory capturing core challenges and capabilities
  • Principled algorithmic interventions supported by systematic evaluation

Submission instructions (TLDR): Submissions may have at most 9 content pages (within that range, all lengths are welcome), excluding references and supplementary material. Papers will be submitted to this OpenReview portal. Works that have been published at archival venues, including NeurIPS 2025, will not be considered. The submission deadline is Wednesday, September 3, 2025.

Topics and Scope

Topics include, but are not limited to:

  1. Models of Reasoning
    Frameworks, theories, analysis of reasoning phenomena
    • Novel paradigms for generalization (ex: compositional, length, diversity, multi-task)
    • Behaviors such as backtracking, “aha moments,” increasing length of reasoning trace
    • Abstractions of reasoning problems, e.g., formal theorem proving as graph problems
    • Characterizing effective inductive biases (ex: pivotal tokens)
  2. Principled Algorithmic Interventions
    Innovations targeting core reasoning challenges, grounded in theoretical and empirical insights
    • Algorithmic innovations in pre-training, post-training, RL, test-time methods
    • Learning or incorporating feedback models (verifiers, rewards, process rewards)
    • Exploration strategies
    • (Automated) collection or design of demonstrations and reasoning traces
  3. Diagnostics and Evaluations
    Dedicated tasks, metrics, and evaluation methods that probe reasoning behaviors
    • Interpretable benchmarks targeting reasoning challenges
    • Tasks simulating aspects of large-scale LM development (ex: pre-training)
    • Rigorous scientific evaluation of reasoning models and algorithms
    • Mechanistic interpretability studies
  4. Representational and Architectural Questions in Reasoning
    How model architectures, learned representations, and generation methods enable or limit capabilities
    • Representation learning, feature acquisition, compositional reasoning
    • Impacts of depth, dimension, chain-of-thought, decoding methods
    • Alternative architectures like state-space models, text diffusion models, deep equilibrium models
  5. Training Paradigms and the Emergence of Reasoning
    How reasoning abilities arise and evolve across pre-training, post-training, and inference
    • Interplay between pre-training and post-training methods (SFT, expert iteration, RL)
    • Test-time strategies that best support reasoning for an LM
    • Curriculum and continual learning
    • Self-improvement and self-verification

Important Dates

  • Submission deadline: Wednesday, September 3, 2025, AoE
  • Author notification: Monday, September 22, 2025, AoE
  • Workshop date: (TBD) Weekend of December 6-7, 2025

Submission Instructions

Format: Submitted papers may have at most 9 content pages, excluding references and supplementary material. Within those constraints, all page lengths are welcome. Previously published papers, including those at NeurIPS 2025, will not be permitted (see “Dual Submission Policy” below).

Style files: Submissions must use the NeurIPS 2025 style files.

Submitting your paper: Please upload your submission as a single PDF to this OpenReview portal.

Anonymization: Reviews will be single-blind, therefore submissions need not be anonymized.

Dual Submission Policy

We do not permit submissions that have been published at archival venues, including those accepted to NeurIPS 2025.

This policy is in accordance with the NeurIPS 2025 workshop guidances, which state that “workshops are not a venue for work that has been previously published in other conferences on machine learning or related fields. Work that is presented at the main NeurIPS conference should not appear in a workshop, including as part of an invited talk.” It ensures that FoRLM 2025 best facilitates the presentation and discussion of fresh contributions best suited to the workshop’s format.