Preference optimization (PO) is indispensable for large language models (LLMs), with methods such as direct preference optimization (DPO) and proximal policy optimization (PPO) achieving great success. A common belief is that DPO is supervised learning while PPO is reinforcement learning, yet deeper analyses for the reasons underlying these differences remain lacking. To fill this gap, we analyze their optimization dynamics, revealing distinct algorithmic behaviors and comprehending their underlying causes. First, we examine the target directions of gradient-based updates and find that DPO follows stable targets, whereas PPO follows dynamic targets that balance exploration and exploitation, thus validating the common belief from a new perspective. Second, we examine the roles of positive learning, negative learning, and loss reweighting, which are three key components in PO methods. Our analyses reveal that these components play fairly different roles. In DPO, positive and negative learning jointly shape the learning targets meanwhile mutually offset each other. However, loss reweighting in DPO acts less as a reward signal but more as a regularizer to mitigate overfitting. In PPO, negative learning primarily supports exploration rather than determining the targets. Meanwhile, loss reweighting, related to absolute values of token-level advantages, indicates the distinct roles of token groups in updating targets. Given these findings, we conduct carefully designed ablation studies to further examine how controlling these dynamics impacts optimization efficiency and practical performance. The insights gained from our analyses not only deepen the understanding of PO methods but also inspire the development of more preference-aligned LLMs.

Recent advances have enabled Large Language Models (LLMs) to tackle reasoning tasks by generating chain-of-thought (CoT) rationales, yet these gains have largely applied to high-resource languages, leaving low-resource languages behind. In this work, we first investigate CoT techniques in extremely low-resource scenarios through previous prompting, model-editing, and fine-tuning approaches. We introduce English-Pivoted CoT Training, leveraging the insight that LLMs internally operate in a latent space aligned toward the dominant language. Given input in a low-resource language, we perform supervised fine-tuning to generate CoT in English and output the final response in the target language. Across mathematical reasoning benchmarks, our approach outperforms other baselines with up to 28.33% improvement in low-resource scenarios. Our analysis and additional experiments, including Mixed-Language CoT and Two-Stage Training, show that explicitly separating language understanding from reasoning enhances cross-lingual reasoning abilities. To facilitate future work, we also release \emph{LC2024}, the first benchmark for mathematical tasks in Irish, an extremely low-resource and endangered language. Our results and resources highlight a practical pathway to multilingual reasoning without extensive retraining in every extremely low-resource language, despite data scarcity.

As AI systems progress, we rely more on them to make decisions with us and for us. To ensure that such decisions are aligned with human values, it is imperative for us to understand not only what decisions they make but also how they come to those decisions. Reasoning language models, which provide both final responses and (partially transparent) intermediate thinking traces, present a timely opportunity to study AI procedural reasoning. Unlike math and code problems which often have objectively correct answers, moral dilemmas are an excellent testbed for process-focused evaluation because they allow for multiple defensible conclusions. To do so, we present MoReBench: 1,000 moral scenarios, each paired with a set of rubric criteria that experts consider essential to include (or avoid) when reasoning about the scenarios. MoReBench contains over 23 thousand criteria including identifying moral considerations, weighing trade-offs, and giving actionable recommendations to cover cases on AI advising humans moral decisions as well as making moral decisions autonomously. Separately, we curate MoReBench-Theory: 150 examples to test whether AI can reason under five major frameworks in normative ethics. Our results show that scaling laws and existing benchmarks on math, code, and scientific reasoning tasks fail to predict models' abilities to perform moral reasoning. Models also show partiality towards specific moral frameworks (e.g., Benthamite Act Utilitarianism and Kantian Deontology), which might be side effects of popular training paradigms. Together, these benchmarks advance process-focused reasoning evaluation towards safer and more transparent AI.

Reasoning language models improve performance on complex tasks by generating long chains of thought (CoTs), but this process can also increase harmful outputs in adversarial settings. In this work, we ask whether the long CoTs can be leveraged for predictive safety monitoring: do the reasoning traces provide early signals of final response alignment that could enable timely intervention? We evaluate a range of monitoring methods using either CoT text or activations, including highly capable large language models, fine-tuned classifiers, and humans. First, we find that a simple linear probe trained on CoT activations significantly outperforms all text-based baselines in predicting whether a final response is safe or unsafe, with an average absolute increase of 13 in F1 scores over the best-performing alternatives. CoT texts are often unfaithful and misleading, while model latents provide a more reliable predictive signal. Second, the probe can be applied to early CoT segments before the response is generated, showing that alignment signals appear before reasoning completes. Error analysis reveals that the performance gap between text classifiers and the linear probe largely stems from a subset of responses we call performative CoTs, where the reasoning consistently contradicts the final response as the CoT progresses. Our findings generalize across model sizes, families, and safety benchmarks, suggesting that lightweight probes could enable real-time safety monitoring and early intervention during generation.

Process Reward Models (PRMs) have recently emerged as a powerful framework for supervising intermediate reasoning steps in large language models (LLMs). Previous PRMs are primarily trained on model final output responses and struggle to evaluate intermediate thinking trajectories robustly, especially in the emerging setting of trajectory-response outputs generated by frontier reasoning models like Deepseek-R1. In this work, we introduce ReasonFlux-PRM, a novel trajectory-aware PRM explicitly designed to evaluate the trajectory-response type of reasoning traces. ReasonFlux-PRM incorporates both step-level and trajectory-level supervision, enabling fine-grained reward assignment aligned with structured chain-of-thought data. We adapt ReasonFlux-PRM to support reward supervision under both offline and online settings, including (i) selecting high-quality model distillation data for downstream supervised fine-tuning of smaller models, (ii) providing dense process-level rewards for policy optimization during reinforcement learning, and (iii) enabling reward-guided Best-of-N test-time scaling. Empirical results on challenging downstream benchmarks such as AIME, MATH500, and GPQA-Diamond demonstrate that ReasonFlux-PRM-7B selects higher quality data than strong PRMs (e.g., Qwen2.5-Math-PRM-72B) and human-curated baselines. Furthermore, our derived ReasonFlux-PRM-7B yields consistent performance improvements, achieving average gains of 12.1% in supervised fine-tuning, 4.5% in reinforcement learning, and 6.3% in test-time scaling. We also release our efficient ReasonFlux-PRM-1.5B for resource-constrained applications and edge deployment. Project: https://github.com/Gen-Verse/ReasonFlux

Recent advancements in Large Language Models (LLMs) have transformed many fields including scientific discovery, content generation, biomedical text mining, and educational technology. However, the substantial requirements for training data, computational resources, and energy consumption pose significant challenges for their practical deployment. A promising alternative is to leverage smaller, specialized language models and aggregate their outputs to improve overall response quality. In this work, we investigate a networked LLM system composed of multiple users, a central task processor, and clusters of topic-specialized LLMs. Each user submits categorical binary (true/false) queries, which are routed by the task processor to a selected cluster of $m$ LLMs. After gathering individual responses, the processor returns a final aggregated answer to the user. We characterize both the information accuracy and response timeliness in this setting, and formulate a joint optimization problem to balance these two competing objectives. Our extensive simulations demonstrate that the aggregated responses consistently achieve higher accuracy than those of individual LLMs. Notably, this improvement is more significant when the participating LLMs exhibit similar standalone performance.

Large language models (LLMs) can effectively elicit human preferences through multi-turn dialogue. Complex tasks can be accomplished through iterative clarifying questions and final responses generated by an LLM acting as a questioner (STaR-GATE; Andukuri et al., 2024}). However, existing approaches based on self-taught reasoning struggle to identify optimal dialogue trajectories and avoid irrelevant questions to the tasks. To address this limitation, we propose TO-GATE, a novel framework that enhances question generation through trajectory optimization, which consists of two key components: a clarification resolver that generates optimal questioning trajectories, and a summarizer that ensures task-aligned final responses. The trajectory optimization enables the model to produce effective elicitation questions and summary responses tailored to specific tasks. Experimental results demonstrate that TO-GATE significantly outperforms baseline methods, achieving a 9.32% improvement on standard preference elicitation tasks.

Supporting learners' understanding of taught skills in online settings is a longstanding challenge. While exercises and chat-based agents can evaluate understanding in limited contexts, this challenge is magnified when learners seek explanations that delve into procedural knowledge ( how things are done) and reasoning ( why things happen). We hypothesize that an intelligent agent's ability to understand and explain learners' questions about skills can be significantly enhanced using the TMK (Task-Method-Knowledge) model, a Knowledge-based AI framework. We introduce Ivy, an intelligent agent that leverages an LLM and iterative refinement techniques to generate explanations that embody teleological, causal, and compositional principles. Our initial evaluation demonstrates that this approach goes beyond the typical shallow responses produced by an agent with access to unstructured text, thereby substantially improving the depth and relevance of feedback. This can potentially ensure learners develop a comprehensive understanding of skills crucial for effective problem-solving in online environments.

As robots increasingly operate in dynamic human-centric environments, improving their ability to detect, explain, and recover from action-related issues becomes crucial. Traditional model-based and data-driven techniques lack adaptability, while more flexible generative AI methods struggle with grounding extracted information to real-world constraints. We introduce RAIDER, a novel agent that integrates Large Language Models (LLMs) with grounded tools for adaptable and efficient issue detection and explanation. Using a unique "Ground, Ask& Answer, Issue" procedure, RAIDER dynamically generates context-aware precondition questions and selects appropriate tools for resolution, achieving targeted information gathering. Our results within a simulated household environment surpass methods relying on predefined models, full scene descriptions, or standalone trained models. Additionally, RAIDER's explanations enhance recovery success, including cases requiring human interaction. Its modular architecture, featuring self-correction mechanisms, enables straightforward adaptation to diverse scenarios, as demonstrated in a real-world human-assistive task. This showcases RAIDER's potential as a versatile agentic AI solution for robotic issue detection and explanation, while addressing the problem of grounding generative AI for its effective application in embodied agents. Project website: https://raider-llmagent.github.io/

As our world digitizes, web agents that can automate complex and monotonous tasks are becoming essential in streamlining workflows. This paper introduces an approach to improving web agent performance through multi-modal validation and self-refinement. We present a comprehensive study of different modalities (text, vision) and the effect of hierarchy for the automatic validation of web agents, building upon the state-of-the-art Agent-E web automation framework. We also introduce a self-refinement mechanism for web automation, using the developed auto-validator, that enables web agents to detect and self-correct workflow failures. Our results show significant gains on Agent-E's (a SOTA web agent) prior state-of-art performance, boosting task-completion rates from 76.2\% to 81.24\% on the subset of the WebVoyager benchmark. The approach presented in this paper paves the way for more reliable digital assistants in complex, real-world scenarios.