Instruction-tuning fundamentally transforms how language models process linguistic input and interact with the user. Through the lens of speech act theory, we investigate whether instruction-tuning causes models to shift from prioritizing syntactical form to pragmatic intent. We create a controlled dataset of 400 sentences systematically varying along two dimensions: syntactical structure (declarative vs. interrogative) and communicative intent (assertive vs. request). Using Principal Component Analysis on hidden state representations from Qwen2.5 (1.5B-7B) and models from two other families (Gemma3-1B, and Llama3.2-3B), we reveal a consistent pattern: base models cluster sentences by syntactical form, while instruction-tuned models reorganize representations around pragmatic intent. This syntactic-to-pragmatic shift occurs in middle layers, with declarative requests and interrogative requests—maximally separated in base models—becoming the most similar categories after instruction-tuning. The phenomenon explains how instruction-tuned models correctly interpret indirect speech acts, treating polite declaratives like I’d appreciate corrections" as functionally equivalent to direct interrogatives. Our findings demonstrate that instruction-tuning teaches models to prioritize the communicative dimension over surface form, a fundamental reorganization consistent across model scales and architectures.

Children acquire language from far less input than language models, and social interaction is widely thought to be part of the reason. Dialogue games cannot fully replicate human interaction, but they offer a controllable environment for simulating situated language use and for studying how models learn from and through interaction. The LM Playschool Challenge builds on this idea, shifting language model training away from observational data towards multi-turn language use. Participants were free to post-train a language model of any size, using any data or method, with one objective in mind — to improve its performance on an unseen set of dialogue games. Our findings suggest that post-training on dialogue games effectively teaches game rules and structure and can support generalisation, especially for small models — the best system achieved gains of roughly seven and three times its baseline score on held-out instances of seen games and on unseen games, respectively. However, gains on unseen games were concentrated in a few games, and larger models improved mainly by achieving better completion rates. General ability was largely retained across submitted systems, owing to the fact that fine-tuning on dialogue games produces shorter responses — and in turn, trigger fewer formatting errors on static benchmarks. Nearly all submissions relied on supervised fine-tuning, often followed by preference optimisation, and none trained on the dense turn-level feedback offered by Playpen, leaving this an open direction for future iterations of the challenge.

Interaction between learner and feedback-giver has come into focus recently for post-training of Large Language Models (LLMs), through the use of reward models that judge the appropriateness of a model’s response. In this paper, we investigate whether Dialogue Games—goal-directed and rule-governed activities driven predominantly by verbal actions—can also serve as a source of feedback signals for learning.We introduce Playpen, an environment for off- and online learning through Dialogue Game self-play, and investigate a representative set of post-training methods: supervised fine-tuning; direct alignment (DPO); and reinforcement learning with Group Relative Policy Optimization (GRPO). We experiment with post-training a small LLM (Llama-3.1-8B-Instruct), evaluating performance on unseen instances of training games as well as unseen games, and on standard benchmarks. We find that imitation learning through SFT improves performance on unseen instances, but negatively impacts other skills, while interactive learning with GRPO shows balanced improvements without loss of skills. We release the framework and the baseline training setups to foster research in this promising new direction of “learning in (synthetic) interaction”.

Questions are essential tools for acquiring the necessary information to complete information-seeking tasks. However, large language models (LLMs), especially open-source models, often perform poorly in generating informative questions, as measured by expected information gain (EIG). In this paper, we propose a method to enhance the informativeness of LLM-generated questions in 20-question game dialogues. We sample multiple questions from the same model (LLaMA 2-Chat 7B) for each game and create pairs of low-EIG and high-EIG questions to apply a Direct Preference Optimization (DPO) algorithm. Our results show that this method produces more effective questions (in terms of EIG), even in domains different from those used to train the DPO model.

Given a word in context, the task of VisualWord Sense Disambiguation consists of select-ing the correct image among a set of candidates. To select the correct image, we propose a so-lution blending text augmentation and multi-modal models. Text augmentation leverages thefine-grained semantic annotation from Word-Net to get a better representation of the tex-tual component. We then compare this sense-augmented text to the set of image using pre-trained multimodal models CLIP and ViLT. Oursystem has been ranked 16th for the Englishlanguage, achieving 68.5 points for hit rate and79.2 for mean reciprocal rank.

Large Language Models, and ChatGPT in particular, have recently grabbed the attention of the community and the media. Having reached high language proficiency, attention has been shifting toward its reasoning capabilities. In this paper, our main aim is to evaluate ChatGPT’s question generation in a task where language production should be driven by an implicit reasoning process. To this end, we employ the 20-Questions game, traditionally used within the Cognitive Science community to inspect the information seeking-strategy’s development. This task requires a series of interconnected skills: asking informative questions, stepwise updating the hypothesis space, and stopping asking questions when enough information has been collected. We build hierarchical hypothesis spaces, exploiting feature norms collected from humans vs. ChatGPT itself, and we inspect the efficiency and informativeness of ChatGPT’s strategy. Our results show that ChatGPT’s performance gets closer to an optimal agent only when prompted to explicitly list the updated space stepwise.