Model editing enables targeted updates to the knowledge of large language models (LLMs) with minimal retraining. Among existing approaches, locate-then-edit methods constitute a prominent paradigm: they first identify critical layers, then compute residuals at the final critical layer based on the target edit, and finally apply least-squares-based multi-layer updates via residual distribution. While empirically effective, we identify a counterintuitive failure mode: residual distribution, a core mechanism in these methods, introduces weight shift errors that undermine editing precision. Through theoretical and empirical analysis, we show that such errors increase with the distribution distance, batch size, and edit sequence length, ultimately leading to inaccurate or suboptimal edits. To address this, we propose the Boundary Layer UpdatE (BLUE) strategy to enhance locate-then-edit methods. Sequential batch editing experiments on three LLMs and two datasets demonstrate that BLUE not only delivers an average performance improvement of 35.59%, significantly advancing the state of the art in model editing, but also enhances the preservation of LLMs' general capabilities. Our code is available at https://github.com/xpq-tech/BLUE.

Measuring progress is fundamental to the advancement of any scientific field. As benchmarks play an increasingly central role, they also become more susceptible to distortion. Chatbot Arena has emerged as the go-to leaderboard for ranking the most capable AI systems. Yet, in this work we identify systematic issues that have skewed the competitive landscape. Specifically, undisclosed private testing practices benefit a handful of providers who are able to test multiple variants before public release and selectively retract scores.

For an LLM to correctly respond to an instruction it must understand both the semantics and the domain (i.e., subject area) of a given task-instruction pair. However, syntax can also convey implicit information. Recent work shows that syntactic templates--frequent sequences of Part-of-Speech (PoS) tags--are prevalent in training data and often appear in model outputs. In this work we characterize syntactic templates, domain, and semantics in task-instruction pairs. We identify cases of spurious correlations between syntax and domain, where models learn to associate a domain with syntax during training; this can sometimes override prompt semantics.

With the growing demand for deploying large language models (LLMs) across diverse applications, improving their inference efficiency is crucial for sustainable and democratized access. However, retraining LLMs to meet new user-specific requirements is prohibitively expensive and environmentally unsustainable. In this work, we propose a practical and scalable alternative: composing efficient hybrid language models from existing pre-trained models. Our approach, Zebra-Llama, introduces a family of 1B, 3B, and 8B hybrid models by combining State Space Models (SSMs) and Multi-head Latent Attention (MLA) layers, using a refined initialization and post-training pipeline to efficiently transfer knowledge from pre-trained Transformers. Zebra-Llama achieves Transformer-level accuracy with near-SSM efficiency using only 7-11B training tokens (compared to trillions of tokens required for pre-training) and an 8B teacher. Moreover, Zebra-Llama dramatically reduces KV cache size--down to 3.9%, 2%, and 2.73% of the original for the 1B, 3B, and 8B variants, respectively--while preserving 100%, 100%, and >97% of average zero-shot performance on LMHarness tasks. Compared to models like MambaInLLaMA, X-EcoMLA, Minitron, and Llamba, Zebra-Llama consistently delivers competitive or superior accuracy while using significantly fewer tokens, smaller teachers, and vastly reduced KV cache memory.

Reasoning-enhanced large language models (RLLMs), whether explicitly trained for reasoning or prompted via chain-of-thought (CoT), have achieved state-ofthe-art performance on many complex reasoning tasks. However, we uncover a surprising and previously overlooked phenomenon: explicit CoT reasoning can significantly degrade instruction-following accuracy. Evaluating 20+ models on two benchmarks: IFEval (with simple, rule-verifiable constraints) and ComplexBench (with complex, compositional constraints), we consistently observe performance drops when CoT prompting is applied. Through large-scale case studies and an attention-based analysis, we identify common patterns where reasoning either helps (e.g., with formatting or lexical precision) or hurts (e.g., by neglecting simple constraints or introducing unnecessary content). We propose a metric, constraint attention, to quantify model focus during generation and show that CoT reasoning often diverts attention away from instruction-relevant tokens. To mitigate these effects, we introduce and evaluate four strategies: in-context learning, selfreflection, self-selective reasoning, and classifier-selective reasoning. Our results demonstrate that selective reasoning strategies, particularly classifier-selective reasoning, can substantially recover lost performance. To our knowledge, this is the first work to systematically expose reasoning-induced failures in instructionfollowing and offer practical mitigation strategies.

Speculative decoding is a promising approach for accelerating large language models. The primary idea is to use a lightweight draft model to speculate the output of the target model for multiple subsequent timesteps, and then verify them in parallel to determine whether the drafted tokens should be accepted or rejected. To enhance acceptance rates, existing frameworks typically construct token trees containing multiple candidates in each timestep. However, their reliance on token-level verification mechanisms introduces two critical limitations: First, the probability distribution of a sequence differs from that of individual tokens, leading to suboptimal acceptance length. Second, current verification schemes begin from the root node and proceed layer by layer in a top-down manner.

Simulating collective decision-making involves more than aggregating individual behaviors; it emerges from dynamic interactions among individuals. While large language models (LLMs) offer strong potential for social simulation, achieving quantitative alignment with real-world data remains a key challenge. To bridge this gap, we propose the Mean-Field LLM (MF-LLM) framework, the first to incorporate mean field theory into LLM-based social simulation. MF-LLM models bidirectional interactions between individuals and the population through an iterative process, generating population signals to guide individual decisions, which in turn update the signals. This interplay produces coherent trajectories of collective behavior. To improve alignment with real-world data, we introduce IB-Tune, a novel fine-tuning method inspired by the Information Bottleneck principle, which retains population signals most predictive of future actions while filtering redundant history. Evaluated on a real-world social dataset, MF-LLM reduces KL divergence to human population distributions by 47% compared to nonmean-field baselines, enabling accurate trend forecasting and effective intervention planning. Generalizing across 7 domains and 4 LLM backbones, MF-LLM provides a scalable, high-fidelity foundation for social simulation.

We introduce IntrinsiX, a novel method that generates high-quality intrinsic images from text description. In contrast to existing text-to-image models whose outputs contain baked-in scene lighting, our approach predicts physically-based rendering (PBR) maps. This enables the generated outputs to be used for content creation scenarios in core graphics applications that facilitate re-lighting, editing, and texture generation tasks. In order to train our generator, we exploit strong image priors, and pre-train separate models for each PBR material component (albedo, roughness, metallic, normals). We then align these models with a new cross-intrinsic attention formulation that concatenates key and value features in a consistent fashion.

Why do AI models struggle with online hate speech detection? Hate speech that once circulated in person now travels farther and faster via anonymous online accounts behind a screen. As the United Nations marks the International Day for Countering Hate Speech on June 18, UN Secretary-General Antonio Guterres has warned that social platforms are amplifying the threat. With artificial intelligence (AI) increasingly tasked with detecting and removing hate speech online, Al Jazeera looks at where these systems fall short compared with human judgement. How is hate speech defined?

Large language models (LLMs) demonstrate advanced reasoning abilities, enabling robots to understand natural language instructions and generate high-level plans with appropriate grounding. However, LLM hallucinations present a significant challenge, often leading to overconfident yet potentially misaligned or unsafe plans. While researchers have explored uncertainty estimation to improve the reliability of LLM-based planning, existing studies have not sufficiently differentiated between epistemic and intrinsic uncertainty, limiting the effectiveness of uncertainty estimation. In this paper, we present Combined Uncertainty estimation for Reliable Embodied planning (CURE), which decomposes the uncertainty into epistemic and intrinsic uncertainty, each estimated separately. Furthermore, epistemic uncertainty is subdivided into task clarity and task familiarity for more accurate evaluation. The overall uncertainty assessments are obtained using random network distillation and multi-layer perceptron regression heads driven by LLM features.