We explore the potential of self-play training for large language models (LLMs) in a two-player adversarial language game called Adversarial Taboo. In this game, an attacker and a defender communicate around a target word only visible to the attacker. The attacker aims to induce the defender to speak the target word unconsciously, while the defender tries to infer the target word from the attacker's utterances. To win the game, both players must have sufficient knowledge about the target word and high-level reasoning ability to infer and express in this information-reserved conversation. Hence, we are curious about whether LLMs' reasoning ability can be further enhanced by Self-Playing this Adversarial language Game (SPAG). With this goal, we select several open-source LLMs and let each act as the attacker and play with a copy of itself as the defender on an extensive range of target words. Through reinforcement learning on the game outcomes, we observe that the LLMs' performances uniformly improve on a broad range of reasoning benchmarks. Furthermore, iteratively adopting this self-play process can continuously promote LLMs' reasoning abilities. The code is available at https://github.com/Linear95/SPAG.

In this game, an attacker and a defender communicate around a target word only visible to the attacker.

Fine-grained word meaning resolution remains a critical challenge for neural language models (NLMs) as they often overfit to global sentence representations, failing to capture local semantic details. We propose a novel adversarial training strategy, called LANE, to address this limitation by deliberately shifting the model's learning focus to the target word. This method generates challenging negative training examples through the selective marking of alternate words in the training set. The goal is to force the model to create a greater separability between same sentences with different marked words. Experimental results on lexical semantic change detection and word sense disambiguation benchmarks demonstrate that our approach yields more discriminative word representations, improving performance over standard contrastive learning baselines. We further provide qualitative analyses showing that the proposed negatives lead to representations that better capture subtle meaning differences even in challenging environments. Our method is model-agnostic and can be integrated into existing representation learning frameworks.

Recent studies have revealed that LLMs can exhibit behavioral self-awareness: the ability to accurately describe or predict their own learned behaviors without explicit supervision. This capability raises safety concerns as it may, for example, allow models to better conceal their true abilities during evaluation. We attempt to characterize the minimal conditions under which such self-awareness emerges, and the mechanistic processes through which it manifests. Through controlled finetuning experiments on instruction-tuned LLMs with low-rank adapters (LoRA), we find: (1) that self-awareness can be reliably induced using a single rank-1 LoRA adapter; (2) that the learned self-aware behavior can be largely captured by a single steering vector in activation space, recovering nearly all of the fine-tune's behavioral effect; and (3) that self-awareness is non-universal and domain-localized, with independent representations across tasks. Together, these findings suggest that behavioral self-awareness emerges as a domain-specific, linear feature that can be easily induced and modulated.

Large language models demonstrate proficiency on phonetic tasks, such as rhyming, without explicit phonetic or auditory grounding. In this work, we investigate how \verb|Llama-3.2-1B-Instruct| represents token-level phonetic information. Our results suggest that Llama uses a rich internal model of phonemes to complete phonetic tasks. We provide evidence for high-level organization of phoneme representations in its latent space. In doing so, we also identify a ``phoneme mover head" which promotes phonetic information during rhyming tasks. We visualize the output space of this head and find that, while notable differences exist, Llama learns a model of vowels similar to the standard IPA vowel chart for humans, despite receiving no direct supervision to do so.

We investigate a new setting for foreign language learning, where learners infer the meaning of unfamiliar words in a multimodal context of a sentence describing a paired image. We conduct studies with human participants using different image-text pairs. We analyze the features of the data (i.e., images and texts) that make it easier for participants to infer the meaning of a masked or unfamiliar word, and what language backgrounds of the participants correlate with success. We find only some intuitive features have strong correlations with participant performance, prompting the need for further investigating of predictive features for success in these tasks. We also analyze the ability of AI systems to reason about participant performance, and discover promising future directions for improving this reasoning ability.

In this game, an attacker and a defender communicate around a target word only visible to the attacker.

Autoregressive Large Language Models (LLMs) demonstrate exceptional performance in language understanding and generation. However, their application in text embedding tasks has been relatively slow, along with the analysis of their semantic representation in probing tasks, due to the constraints of the unidirectional attention mechanism. This paper aims to explore whether such constraints can be overcome by enabling bidirectional attention in LLMs. We tested different variants of the Llama architecture through additional training steps, progressively enabling bidirectional attention and unsupervised/supervised contrastive learning.

Despite their strong performance, large language models (LLMs) face challenges in real-world application of lexical simplification (LS), particularly in privacy-sensitive and resource-constrained environments. Moreover, since vulnerable user groups (e.g., people with disabilities) are one of the key target groups of this technology, it is crucial to ensure the safety and correctness of the output of LS systems. To address these issues, we propose an efficient framework for LS systems that utilizes small LLMs deployable in local environments. Within this framework, we explore knowledge distillation with synthesized data and in-context learning as baselines. Our experiments in five languages evaluate model outputs both automatically and manually. Our manual analysis reveals that while knowledge distillation boosts automatic metric scores, it also introduces a safety trade-off by increasing harmful simplifications. Importantly, we find that the model's output probability is a useful signal for detecting harmful simplifications. Leveraging this, we propose a filtering strategy that suppresses harmful simplifications while largely preserving beneficial ones. This work establishes a benchmark for efficient and safe LS with small LLMs. It highlights the key trade-offs between performance, efficiency, and safety, and demonstrates a promising approach for safe real-world deployment.

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 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 the promising new direction of learning in (synthetic) interaction.