The rapid development of large language models (LLMs) has driven the demand for more efficient optimization techniques. Among these, the Lookahead family of optimizers employs a two-loop framework, maintaining fast and slow sets of model weights. Multiple inner optimizer steps on the fast weights produce a trajectory - the pseudo-gradient - that is used to update the slow weights. DiLoCo, a notable example originally designed for distributed training, applies Nesterov momentum to the averaged pseudo-gradient from multiple workers, claiming to even outperform AdamW in a non-distributed setup. In this paper, we empirically show that DiLoCo's surprising effectiveness stems primarily from applying Nesterov momentum to the pseudo-gradient, which improves training in a non-distributed setting. We call this Lookahead variant the Step-$K$ Nesterov Outer Optimizer (SNOO). We demonstrate that SNOO achieves compute factor gains of 1.5 - 2.5$\times$ in a non-distributed setting up to a scale of 1e23 training FLOPs, with improvements that increase with model size. Because of its minimal compute and memory overhead and compatibility with model sharding, SNOO is a practical enhancement for a variety of inner optimizers, including AdamW and Muon.

Recent probing studies reveal that large language models exhibit linear subspaces that separate true from false statements, yet the mechanism behind their emergence is unclear. We introduce a transparent, one-layer transformer toy model that reproduces such truth subspaces end-to-end and exposes one concrete route by which they can arise. We study one simple setting in which truth encoding can emerge: a data distribution where factual statements co-occur with other factual statements (and vice-versa), encouraging the model to learn this distinction in order to lower the LM loss on future tokens. We corroborate this pattern with experiments in pretrained language models. Finally, in the toy setting we observe a two-phase learning dynamic: networks first memorize individual factual associations in a few steps, then -- over a longer horizon -- learn to linearly separate true from false, which in turn lowers language-modeling loss. Together, these results provide both a mechanistic demonstration and an empirical motivation for how and why linear truth representations can emerge in language models.

Recent image segmentation models have advanced to segment images into high-quality masks for visual entities, and yet they cannot provide comprehensive semantic understanding for complex queries based on both language and vision. This limitation reduces their effectiveness in applications that require user-friendly interactions driven by vision-language prompts. To bridge this gap, we introduce a novel task of omnimodal referring expression segmentation (ORES). In this task, a model produces a group of masks based on arbitrary prompts specified by text only or text plus reference visual entities. To address this new challenge, we propose a novel framework to "Refer to Any Segmentation Mask Group" (RAS), which augments segmentation models with complex multimodal interactions and comprehension via a mask-centric large multimodal model. For training and benchmarking ORES models, we create datasets MaskGroups-2M and MaskGroups-HQ to include diverse mask groups specified by text and reference entities. Through extensive evaluation, we demonstrate superior performance of RAS on our new ORES task, as well as classic referring expression segmentation (RES) and generalized referring expression segmentation (GRES) tasks. Project page: https://Ref2Any.github.io.

As AI chatbots gain adoption in clinical medicine, developing effective frameworks for complex, emerging diseases presents significant challenges. We developed and evaluated six Retrieval-Augmented Generation (RAG) corpus configurations for Long COVID (LC) clinical question answering, ranging from expert-curated sources to large-scale literature databases. Our evaluation employed an LLM-as-a-judge framework across faithfulness, relevance, and comprehensiveness metrics using LongCOVID-CQ, a novel dataset of expert-generated clinical questions. Our RAG corpus configuration combining clinical guidelines with high-quality systematic reviews consistently outperformed both narrow single-guideline approaches and large-scale literature databases. Our findings suggest that for emerging diseases, retrieval grounded in curated secondary reviews provides an optimal balance between narrow consensus documents and unfiltered primary literature, supporting clinical decision-making while avoiding information overload and oversimplified guidance. We propose Guide-RAG, a chatbot system and accompanying evaluation framework that integrates both curated expert knowledge and comprehensive literature databases to effectively answer LC clinical questions.

So-called `wicked problems', those involving complex multi-dimensional settings, non-verifiable outcomes, heterogeneous impacts and a lack of single objectively correct answers, have plagued humans throughout history. Modern examples include decisions over justice frameworks, solving environmental pollution, planning for pandemic resilience and food security. The use of state-of-the-art artificial intelligence systems (notably Large Language Model-based agents) collaborating with humans on solving such problems is being actively explored. While the abilities of LLMs can be improved by, for example, fine-tuning, hand-crafted system prompts and scaffolding with external tools, LLMs lack endogenous mechanisms to develop expertise through experience in such settings. This work address this gap with Dialectica, a framework where agents engage in structured dialogue on defined topics, augmented by memory, self-reflection, and policy-constrained context editing. Formally, discussion is viewed as an implicit meta-reinforcement learning process. The `dialogue-trained' agents are evaluated post-hoc using judged pairwise comparisons of elicited responses. Across two model architectures (locally run Qwen3:30b and OpenAI's o4-mini) results show that enabling reflection-based context editing during discussion produces agents which dominate their baseline counterparts on Elo scores, normalized Bradley-Terry-Davidson ability, and AlphaRank mass. The predicted signatures of learning are observed qualitatively in statement and reflection logs, where reflections identify weaknesses and reliably shape subsequent statements. Agreement between quantitative and qualitative evidence supports dialogue-driven context evolution as a practical path to targeted expertise amplification in open non-verifiable domains.

Ideal or real - that is the question.In this work, we explore whether principles from game theory can be effectively applied to the evaluation of large language models (LLMs). This inquiry is motivated by the growing inadequacy of conventional evaluation practices, which often rely on fixed-format tasks with reference answers and struggle to capture the nuanced, subjective, and open-ended nature of modern LLM behavior. To address these challenges, we propose a novel alternative: automatic mutual evaluation, where LLMs assess each other's output through self-play and peer review. These peer assessments are then systematically compared with human voting behavior to evaluate their alignment with human judgment. Our framework incorporates game-theoretic voting algorithms to aggregate peer reviews, enabling a principled investigation into whether model-generated rankings reflect human preferences. Empirical results reveal both convergences and divergences between theoretical predictions and human evaluations, offering valuable insights into the promises and limitations of mutual evaluation. To the best of our knowledge, this is the first work to jointly integrate mutual evaluation, game-theoretic aggregation, and human-grounded validation for evaluating the capabilities of LLMs.

Reasoning models have gained significant attention due to their strong performance, particularly when enhanced with retrieval augmentation. However, these models often incur high computational costs, as both retrieval and reasoning tokens contribute substantially to the overall resource usage. In this work, we make the following contributions: (1) we propose a retrieval-augmented reasoning model that dynamically adjusts the length of the retrieved document list based on the query and retrieval results; (2) we develop a cost-aware advantage function for training of efficient retrieval-augmented reasoning models through reinforcement learning; and (3) we explore both memory- and latency-bound implementations of the proposed cost-aware framework for both proximal and group relative policy optimization algorithms. We evaluate our approach on seven public question answering datasets and demonstrate significant efficiency gains, without compromising effectiveness. In fact, we observed that the model latency decreases by ~16-20% across datasets, while its effectiveness increases by ~5% on average, in terms of exact match.

Reinforcement Learning from Human Feedback (RLHF) has become central to aligning large language models with human values, typically by first learning a reward model from preference data which is then used to update the model with reinforcement learning. Recent alternatives such as Direct Preference Optimization (DPO) simplify this pipeline by directly optimizing on preferences. However, both approaches often assume uniform annotator preferences and rely on binary comparisons, overlooking two key limitations: the diversity of human evaluators and the limitations of pairwise feedback. In this work, we address both these issues. First, we connect preference learning in RLHF with the econometrics literature and show that binary comparisons are insufficient for identifying latent user preferences from finite user data and infinite users, while (even incomplete) rankings over three or more responses ensure identifiability. Second, we introduce methods to incorporate heterogeneous preferences into alignment algorithms. We develop an Expectation-Maximization adaptation of DPO that discovers latent annotator types and trains a mixture of LLMs accordingly. Then we propose an aggregation algorithm using a min-max regret fairness criterion to produce a single generative policy with equitable performance guarantees. Together, these contributions establish a theoretical and algorithmic framework for fairness and personalization for diverse users in generative model alignment.

Neural theorem proving has advanced rapidly in the past year, reaching IMO gold-medalist capabilities and producing formal proofs that span thousands of lines. Although such proofs are mechanically verified by formal systems like Lean, their excessive length renders them difficult for humans to comprehend and limits their usefulness for mathematical insight. Proof simplification is therefore a critical bottleneck. Yet, training data for this task is scarce, and existing methods -- mainly agentic scaffolding with off-the-shelf LLMs -- struggle with the extremely long proofs generated by RL-trained provers. We introduce ProofOptimizer, the first language model trained to simplify Lean proofs without requiring additional human supervision. ProofOptimizer is trained via expert iteration and reinforcement learning, using Lean to verify simplifications and provide training signal. At inference time, it operates within an iterative proof-shortening workflow, progressively reducing proof length. Experiments show that ProofOptimizer substantially compresses proofs generated by state-of-the-art RL-trained provers on standard benchmarks, reducing proof length by 87% on miniF2F, 57% on PutnamBench, and 49% on Seed-Prover's IMO 2025 proofs. Beyond conciseness, the simplified proofs check faster in Lean and further improve downstream prover performance when reused as training data for supervised finetuning.

This research introduces a novel approach to textual and multimodal Hate Speech Detection (HSD), using Large Language Models (LLMs) as dynamic knowledge bases to generate background context and incorporate it into the input of HSD classifiers. Two context generation strategies are examined: one focused on named entities and the other on full-text prompting. Four methods of incorporating context into the classifier input are compared: text concatenation, embedding concatenation, a hierarchical transformer-based fusion, and LLM-driven text enhancement. Experiments are conducted on the textual Latent Hatred dataset of implicit hate speech and applied in a multimodal setting on the MAMI dataset of misogynous memes. Results suggest that both the contextual information and the method by which it is incorporated are key, with gains of up to 3 and 6 F1 points on textual and multimodal setups respectively, from a zero-context baseline to the highest-performing system, based on embedding concatenation.