LM Playschool (LMP 2026): Improving Language Models through Learning from Dialogue Interaction .
Co-located with EMNLP 2026 — 24–29 October 2026
Website: https://lm-playschool.github.io/
The LM Playschool Workshop invites submissions exploring the frontier of language agents that learn, adapt, and improve through situated interaction. We focus on conversational, collaborative, goal-oriented, and multi-turn environments—moving beyond static training and evaluation datasets and asking whether interactive learning can yield more data-efficient, robust, and adaptable language agents.
Workshop Vision
Current LLM training and evaluation often still rely on static snapshots of linguistic experience. However, human language acquisition and use are inherently social, grounded, and iterative. LMP 2026 aims to bridge the gap between three ongoing research trends:
- Game-based Benchmarking: Using interactive games and game-like environments to measure functional linguistic competence and verbal interaction skills (Qiao et al., 2023; Momentè et al., 2025)
- Machine Language Acquisition: Drawing inspiration from the data efficiency of human children, who learn through social interaction rather than just observation of massive text corpora [Ma et al., 2024; Charpentier et al., 2025).
- Interactive Learning Signals: Moving away from training objectives that reward the local appropriateness of single tokens toward utterance-level signals of communicative success (Lightman et al., 2023; Gul & Artzi, 2024; Horst et al., 2025).
Topics of Interest
We welcome original research and work-in-progress on topics including, but not limited to:
- Architectures and Training Regimes for Interactive Agents: Models and training procedures designed for situated, multi-turn, goal-directed language interaction.
- Intrinsic Rewards & Learning Signals: Reinforcement learning from game-state success and interaction-relevant signals rather than human preference models (RLHF).
- Benchmarking via Dialogue Games: New environments that challenge model reasoning and collaborative skills.
- Data Efficiency & Social Interaction: Insights from developmental linguistics applied to machine learning.
- Social Cognition in Interactive Systems: Developing agents that can model their partner's intent to achieve shared goals (e.g., Theory of Mind modelling).
- Human-Agent Collaboration: Evaluating how models adapt to human quirks in real-time play.
- Communicative and Perceptual Grounding: Models and benchmarks investigating the interplay and synergies between communicative and perceptual (e.g., visual) grounding.
Additionally, we will feature technical reports developed by the participants of the LM Playschool Challenge.
The LM Playschool Challenge (Shared Task)
LMP 2026 will feature a new shared task where participants are challenged to post-train LLMs to master communicative skills in unseen dialogue games while retaining original language capabilities. We will announce the shared task rules and settings soon, so please stay tuned.
Submission Tracks
We accept two types of submissions:
- Challenge track: Technical report produced by the participants to the LM-Playschool challenge (8 pages for long, 4 pages for short), to be published in the EMNLP 2026 Workshop Proceedings.
- Paper-only track: Work-in-progress, or recently published papers (8 pages for long, 4 pages for short). Work-in-progress submissions may be submitted either as archival or non-archival papers. Recently published papers are non-archival and will be presented at the workshop but not included in the proceedings.
Submission Link: Available soon on our website
Formatting: Submissions must follow the official EMNLP 2026 LaTeX style files. Page limits do not include references.
Important Dates
- ARR paper submission deadline: May 25, 2026
- Challenge submission deadline: July 10, 2026 (TBC)
- Direct paper submission deadline (challenge track and paper-only track): July 17, 2026 (TBC)
- Notification of acceptance: August 20, 2026
- Camera ready due: September 20, 2026
- Workshop at EMNLP in Budapest: October 24-29, 2026 (Exact date TBD)
Organisers
- Raffaella Bernardi, Free University of Bozen-Bolzano
- Raquel Fernández, University of Amsterdam
- Mario Giulianelli, University College London
- Sherzod Hakimov, University of Potsdam
- Alexander Koller, Saarland University
- Dieu-Thu Le, Amazon - AGI, Germany
- Oliver Lemon, Heriot-Watt University
- Davide Mazzaccara, University of Trento
- Sabrina McCallum, University of Edinburgh
- David Schlangen, University of Potsdam
- Alessandro Suglia, University of Edinburgh
Website: https://lm-playschool.github.io/