We show that training a single $d$-dimensional steering vector per layer with reinforcement learning, while freezing all base weights, matches the accuracy of fully RL-tuned reasoning models on mathematical-reasoning tasks. On an 8 billion-parameter model this adds only $\approx 0.0016\%$ additional parameters and reproduces performance across a range of base models and mathematical-reasoning benchmarks. These results tighten the upper bound on the parameter budget required for high-level chain-of-thought reasoning, indicating that millions of adapter weights are unnecessary. The minimal trainable footprint reduces optimizer memory and inter-GPU communication, lowering the overall cost of fine-tuning. Moreover, a logit-lens analysis shows that the learned vectors amplify coherent token directions, providing clearer insight into the model's internal computations.

To enable this capability, we explore architectural adaptations, including interleaved Goal-State-Action modeling, context extension techniques, and Context Parallelism, to equip LLM-based agents for extreme long-context reasoning and interaction. Experimental results and analyses highlight the challenges posed by our benchmark and provide insights into training strategies and model behaviors under long-horizon conditions. Our work provides a foundation for the next generation of embodied AI systems capable of robust, long-term reasoning and planning. Real-world embodied reasoning is a sequential decision-making problem requiring long-horizon planning, where task success depends on both memorizing and reasoning over multiple events that occur far apart in time. Using large pre-trained vision-language-action (VLA) models as policies for such tasks requires surpassing the key challenge of long-context reasoning. We seek to answer questions pertaining to what design choices matter in terms of environments, model architectures, and training methods when using VLA models for long-horizon embodied tasks. To this end, we develop a new framework for long-horizon tasks designed to push the boundaries of long-context understanding in embodied AI. Our benchmark uniquely features tasks with a synthetic final goal, which involves multiple objects that appear at distant time steps, requiring multi-step reasoning across over hundreds of steps. Figure 1 illustrates an example: the agent observes the tomato at an early step (t=17) and the counter top much later (t=560). Then, the final task is given at t=670, which requires the agent to place the tomato on the counter top. This setup highlights the challenge of long-horizon dependency, where key objects and locations must be remembered and acted upon after hundreds of steps. Beyond these long-horizon dependencies, our framework also generates low-level robot-arm manipulation action sequences aligned with the trajectories, enabling agents to bridge from high-level reasoning to fine-grained physical execution.

With advancements in large audio-language models (LALMs), which enhance large language models (LLMs) with auditory capabilities, these models are expected to demonstrate universal proficiency across various auditory tasks. While numerous benchmarks have emerged to assess LALMs' performance, they remain fragmented and lack a structured taxonomy. To bridge this gap, we conduct a comprehensive survey and propose a systematic taxonomy for LALM evaluations, categorizing them into four dimensions based on their objectives: (1) General Auditory Awareness and Processing, (2) Knowledge and Reasoning, (3) Dialogue-oriented Ability, and (4) Fairness, Safety, and Trustworthiness. We provide detailed overviews within each category and highlight challenges in this field, offering insights into promising future directions. To the best of our knowledge, this is the first survey specifically focused on the evaluations of LALMs, providing clear guidelines for the community. We will release the collection of the surveyed papers and actively maintain it to support ongoing advancements in the field.

Domain models enable autonomous agents to solve long-horizon tasks by producing interpretable plans. However, in open-world environments, a single general domain model cannot capture the variety of tasks, so agents must generate suitable task-specific models on the fly. Large Language Models (LLMs), with their implicit common knowledge, can generate such domains, but suffer from high error rates that limit their applicability. Hence, related work relies on extensive human feed-back or prior knowledge, which undermines autonomous, open-world deployment. In this work, we propose LODGE, a framework for autonomous domain learning from LLMs and environment grounding. LODGE builds on hierarchical abstractions and automated simulations to identify and correct inconsistencies between abstraction layers and between the model and environment. Our framework is task-agnostic, as it generates predicates, operators, and their preconditions and effects, while only assuming access to a simulator and a set of generic, executable low-level skills. Experiments on two International Planning Competition ( IPC) domains and a robotic assembly domain show that LODGE yields more accurate domain models and higher task success than existing methods, requiring remarkably few environment interactions and no human feedback or demonstrations.

Legal Passage Retrieval (LPR) systems are crucial as they help practitioners save time when drafting legal arguments. However, it remains an underexplored avenue. One primary reason is the significant vocabulary mismatch between the query and the target passage. To address this, we propose a simple yet effective method, the Generative query REwriter (GuRE). We leverage the generative capabilities of Large Language Models (LLMs) by training the LLM for query rewriting. "Rewritten queries" help retrievers to retrieve target passages by mitigating vocabulary mismatch. Experimental results show that GuRE significantly improves performance in a retriever-agnostic manner, outperforming all baseline methods. Further analysis reveals that different training objectives lead to distinct retrieval behaviors, making GuRE more suitable than direct retriever fine-tuning for real-world applications. Codes are avaiable at github.com/daehuikim/GuRE.

Ambiguity in natural language is a significant obstacle for achieving accurate text to structured data mapping through large language models (LLMs), which affects the performance of tasks such as mapping text to agentic tool calling and text-to-SQL queries. Existing methods to ambiguity handling either rely on the ReACT framework to obtain correct mappings through trial and error, or on supervised fine-tuning to bias models toward specific tasks. In this paper, we adopt a different approach that characterizes representation differences of ambiguous text in the latent space and leverages these differences to identify ambiguity before mapping them to structured data. To detect sentence-level ambiguity, we focus on the relationship between ambiguous questions and their interpretations. Unlike distances calculated by dense embeddings, we introduce a new distance measure based on a path kernel over concepts. With this measurement, we identify patterns to distinguish ambiguous from unambiguous questions. Furthermore, we propose a method for improving LLM performance on ambiguous agentic tool calling through missing concept prediction. Both achieve state-of-the-art results.

Reinforcement learning with verifiable rewards (RLVR) has emerged as the leading approach for enhancing reasoning capabilities in large language models. However, it faces a fundamental compute and memory asymmetry: rollout generation is embarrassingly parallel and memory-light, whereas policy updates are communication-heavy and memory-intensive. To address this, we introduce PODS (Policy Optimization with Down-Sampling), which decouples rollout generation from policy updates by training only on a strategically selected subset of rollouts, maintaining learning quality while dramatically reducing update costs. We propose a principled subset selection criterion, max-variance down-sampling, that maximizes reward diversity, and provide an efficient $O(n\log n)$ implementation. Empirically, Group Relative Policy Optimization (GRPO) with PODS achieves the peak test accuracy of vanilla GRPO at least $\mathbf{1.7\times}$ faster across the different reasoning benchmarks and hardware configurations we tested.

Large Language Models (LLMs) are revolutionizing the development of AI assistants capable of performing diverse tasks across domains. However, current state-of-the-art LLM-driven agents face significant challenges, including high operational costs and limited success rates on complex benchmarks like GAIA. To address these issues, we propose Knowledge Graph of Thoughts (KGoT), an innovative AI assistant architecture that integrates LLM reasoning with dynamically constructed knowledge graphs (KGs). KGoT extracts and structures task-relevant knowledge into a dynamic KG representation, iteratively enhanced through external tools such as math solvers, web crawlers, and Python scripts. Such structured representation of task-relevant knowledge enables low-cost models to solve complex tasks effectively while also minimizing bias and noise. For example, KGoT achieves a 29% improvement in task success rates on the GAIA benchmark compared to Hugging Face Agents with GPT-4o mini. Moreover, harnessing a smaller model dramatically reduces operational costs by over 36x compared to GPT-4o. Improvements for other models (e.g., Qwen2.5-32B and Deepseek-R1-70B) and benchmarks (e.g., SimpleQA) are similar. KGoT offers a scalable, affordable, versatile, and high-performing solution for AI assistants.

Memory consumption of the Key-Value (KV) cache represents a major bottleneck for efficient large language model inference. While attention-score-based KV cache pruning shows promise, it faces critical practical limitations: attention scores from future tokens are unavailable during compression, and modern implementations like Flash Attention do not materialize the full attention matrix, making past scores inaccessible. To overcome these challenges, we introduce $\textbf{Expected Attention, a training-free compression method}$ that estimates KV pairs importance by predicting how future queries will attend to them. Our approach leverages the distributional properties of LLM activations to compute expected attention scores in closed form for each KV pair. These scores enable principled ranking and pruning of KV pairs with minimal impact on the residual stream, achieving effective compression without performance degradation. Importantly, our method operates seamlessly across both prefilling and decoding phases, consistently outperforming state-of-the-art baselines in both scenarios. Finally, $\textbf{we release KVPress, a comprehensive library to enable researchers to implement and benchmark KV cache compression methods, already including more than 20 techniques}$.

Large audio-language models (LALMs) are often used in tasks that involve reasoning over ordered options. An open question is whether their predictions are influenced by the order of answer choices, which would indicate a form of selection bias and undermine their reliability. In this paper, we identify and analyze this problem in LALMs. We demonstrate that no model is immune to this bias through extensive experiments on six LALMs across three widely used benchmarks and their spoken counterparts. Shuffling the order of answer options can cause performance fluctuations of up to 24% and even change model rankings, raising concerns about the reliability of current evaluation practices. We also study permutation-based strategies and show that they can mitigate bias in most cases. Our work represents the first systematic investigation of this issue in LALMs, and we hope it raises awareness and motivates further research in this direction.