Attention sinks and compression valleys have attracted significant attention as two puzzling phenomena in large language models, but have been studied in isolation. In this work, we present a surprising connection between attention sinks and compression valleys, tracing both to the formation of massive activations in the residual stream. We prove theoretically that massive activations necessarily produce representational compression and establish bounds on the resulting entropy reduction. Through experiments across several models (410M-120B parameters), we confirm that when the beginning-of-sequence token develops extreme activation norms in the middle layers, both compression valleys and attention sinks emerge simultaneously. This unified view motivates us to propose the Mix-Compress-Refine theory of information flow, as an attempt to explain how LLMs organize their computation in depth by controlling attention and representational compression via massive activations. Specifically, we posit that Transformer-based LLMs process tokens in three distinct phases: (1) broad mixing in the early layers, (2) compressed computation with limited mixing in the middle layers, and (3) selective refinement in the late layers. Our framework helps explain why embedding tasks perform best at intermediate layers, whereas generation tasks benefit from full-depth processing, clarifying differences in task-dependent representations. Large Language Models (LLMs) have become remarkably capable, yet how they process information through their layers remains poorly understood. Two phenomena have particularly puzzled researchers: attention sinks, where attention heads mysteriously collapse their focus onto semantically uninformative tokens (Xiao et al., 2024), and compression valleys, where intermediate representations show unexpectedly low entropy despite the model's high-dimensional space (Skean et al., 2025). These phenomena appear paradoxical: why would powerful models waste attention on meaningless tokens, and why would representations compress in the middle of processing? Previous work has explained attention sinks through positional biases (Gu et al., 2025) and over-mixing prevention (Barbero et al., 2025a), while compression valleys have been explained through an information bottleneck theory (Skean et al., 2025). However, the precise reasons why they emerge remain unclear and no formal link has been established between them.

This work investigates the reasoning and planning capabilities of foundation models and their scalability in complex, dynamic environments. We introduce PuzzlePlex, a benchmark designed to assess these capabilities through a diverse set of puzzles. PuzzlePlex consists of 15 types of puzzles, including deterministic and stochastic games of varying difficulty, as well as single-player and two-player scenarios. The PuzzlePlex framework provides a comprehensive environment for each game, and supports extensibility to generate more challenging instances as foundation models evolve. Additionally, we implement customized game-playing strategies for comparison. Building on this benchmark, we develop fine-grained metrics to measure performance and conduct an in-depth analysis of frontier foundation models across two settings: instruction-based and code-based. Furthermore, we systematically investigate their scaling limits. Our findings show that reasoning models outperform others in instruction-based settings, while code-based execution presents greater challenges but offers a scalable and efficient alternative. PuzzlePlex enables targeted evaluation and guides future improvements in reasoning, planning, and generalization for foundation models.

As large language models (LLMs) become ubiquitous in workplace tools and decision-making processes, ensuring explainability and fostering user trust are critical. Although advancements in LLM engineering continue, human-centered design is still catching up, particularly when it comes to embedding transparency and trust into AI interfaces. This study evaluates user experiences with two distinct AI interfaces - node-tree interfaces and chatbot interfaces - to assess their performance in exploratory, follow-up inquiry, decision-making, and problem-solving tasks. Our design-driven approach introduces a node-tree interface that visually structures AI-generated responses into hierarchically organized, interactive nodes, allowing users to navigate, refine, and follow up on complex information. In a comparative study with n=20 business users, we observed that while the chatbot interface effectively supports linear, step-by-step queries, it is the node-tree interface that enhances brainstorming. Quantitative and qualitative findings indicate that node-tree interfaces not only improve task performance and decision-making support but also promote higher levels of user trust by preserving context. Our findings suggest that adaptive AI interfaces capable of switching between structured visualizations and conversational formats based on task requirements can significantly enhance transparency and user confidence in AI-powered systems. This work contributes actionable insights to the fields of human-robot interaction and AI design, particularly for enterprise applications where trust-building is critical for teams.

Large language models excel on math benchmarks, but their math reasoning robustness to linguistic variation is underexplored. While recent work increasingly treats high-difficulty competitions like the IMO as the gold standard for evaluating reasoning, we believe in comprehensive benchmarking of high school-level math problems in real educational settings. We introduce MathRobust-LV, a test set and evaluation methodology that mirrors how instructors rephrase problems across assessments while keeping difficulty constant: we change surface details (names, contexts, variables) while preserving numerical structure and answers. In contrast to prior efforts that alter problem content or emphasize IMO-level tasks, we focus on high-school-level dataset problems at the difficulty level where models are currently deployed in educational settings: tutoring and assessment systems. In these applications, instructors rephrase identical concepts in varied ways, making linguistic robustness essential for reliable deployment. Although MATH data benchmarking is often regarded as saturated, our experiment on 34 models reveals that accuracy declines when moving from the baseline to the variants. These drops are severe for smaller models (9-11%) while stronger models also show measurable degradation. Frontier models like GPT-5, Gemini-2.5pro remain comparatively stable. Our results highlight that robustness to linguistic variation is a fundamental challenge, exposing reasoning vulnerabilities in models.

Large Language Models (LLMs) frequently hallucinate to long-form questions, producing plausible yet factually incorrect answers. A common mitigation strategy is to provide attribution to LLM outputs. However, existing benchmarks primarily focus on simple attribution that retrieves supporting textual evidence as references. We argue that in real-world scenarios such as financial applications, attribution goes beyond reference retrieval. We introduce FinLFQA, a benchmark designed to evaluate the ability of LLMs to generate long-form answers to complex financial questions with reliable and nuanced attributions. FinLFQA evaluates three critical aspects of attribution through human annotations: (1) supporting evidence extracted from financial reports, (2) intermediate numerical reasoning steps, and (3) domain-specific financial knowledge that informs the reasoning process. We further provide an automatic evaluation framework covering both answer quality and attribution quality. Through extensive experiments on eight LLMs across multiple attribution-generation paradigms, we find that fine-grained metrics are important to distinguish model capabilities, that end-to-end generation achieves comparable performance to post-hoc approaches, and that iterative refinement only helps when guided by external feedback.

Virtual Labs offer valuable opportunities for hands-on, inquiry-based science learning, yet teachers often struggle to adapt them to fit their instructional goals. Third-party materials may not align with classroom needs, and developing custom resources can be time-consuming and difficult to scale. Recent advances in Large Language Models (LLMs) offer a promising avenue for addressing these limitations. In this paper, we introduce a novel alignment framework for instructional goal-aligned question generation, enabling teachers to leverage LLMs to produce simulation-aligned, pedagogically meaningful questions through natural language interaction. The framework integrates four components: instructional goal understanding via teacher-LLM dialogue, lab understanding via knowledge unit and relationship analysis, a question taxonomy for structuring cognitive and pedagogical intent, and the TELeR taxonomy for controlling prompt detail. Early design choices were informed by a small teacher-assisted case study, while our final evaluation analyzed over 1,100 questions from 19 open-source LLMs. With goal and lab understanding grounding questions in teacher intent and simulation context, the question taxonomy elevates cognitive demand (open-ended formats and relational types raise quality by 0.29-0.39 points), and optimized TELeR prompts enhance format adherence (80% parsability, >90% adherence). Larger models yield the strongest gains: parsability +37.1%, adherence +25.7%, and average quality +0.8 Likert points.

Reasoning LLMs are trained to verbalize their reasoning process, yielding strong gains on complex tasks. This transparency also opens a promising direction: multiple reasoners can directly collaborate on each other's thinking within a shared trajectory, yielding better inference efficiency and exploration. A key prerequisite, however, is the ability to assess the usefulness and build on another model's partial thinking--we call this off-trajectory reasoning. Our paper investigates a critical question: can standard solo-reasoning training pipelines deliver desired off-trajectory behaviors? We propose twin tests that capture the two extremes of the off-trajectory spectrum, namely Recoverability, which tests whether LLMs can backtrack from "distractions" induced by misleading reasoning traces, and Guid-ability, which tests their ability to build upon correct reasoning from stronger collaborators. Our study evaluates 15 open-weight LLMs (1.5B-32B) and reveals a counterintuitive finding--"stronger" LLMs on benchmarks are often more fragile under distraction. Moreover, all models tested fail to effectively leverage guiding steps from collaborators on problems beyond their inherent capabilities with solve rates remaining under 9.2%. Finally, we conduct control studies to isolate the effects of three factors in post-training on these behaviors: the choice of distillation teacher, the use of RL, and data selection strategy. Our results provide actionable insights for training natively strong reasoning collaborators; e.g., we find that suboptimal recoverability behaviors of teacher models are transferred to distilled students even if the distillation trajectories are correct. Taken together, this work lays the groundwork for evaluating multi-model collaborations in shared reasoning trajectories and highlights the limitations of off-the-shelf reasoning LLMs. Figure 1: Comparison of solo (left) vs. collaborative reasoning (right) setting. LLMs of different sizes and functionalities collaborate on a shared trajectory.

Reward models (RMs) are central to the alignment of language models (LMs). An RM often serves as a proxy for human preferences to guide downstream LM behavior. However, our understanding of RM behavior is limited. Our work (i) formalizes a framework for measuring the alignment of opinions captured by RMs, (ii) investigates the extent to which RMs demonstrate sociodemographic biases, and (iii) explores the effects of prompting to steer rewards towards the preferences of a target group. We study the subjective and diverse perspectives on controversial topics, which allows us to quantify RM perspectives in terms of their opinions, attitudes, and values. We show that RMs are poorly aligned with several demographic groups and can systematically reward harmful stereotypes, and steering alone is not enough to overcome these limitations. Our findings underscore the need for more careful consideration of RM behavior in model alignment during preference learning to prevent the propagation of unwanted social biases in the language technologies that we use.

Generating stylistic text with specific attributes is a key problem in controllable text generation. Recently, diffusion models have emerged as a powerful paradigm for both visual and textual generation. Existing approaches can be broadly categorized into classifier-free guidance (CFG) and classifier guidance (CG) methods. While CFG effectively preserves semantic content, it often fails to provide effective attribute control. In contrast, CG modifies the denoising trajectory using classifier gradients, enabling better attribute alignment but incurring high computational costs during sampling and suffering from classifier generalization issues. In this work, we propose RegDiff, a regularized diffusion framework that leverages attribute features without requiring a pretrained classifier during sampling, thereby achieving controllable generation with reduced computational costs. Specifically, RegDiff employs a VAE-based encoder--decoder architecture to ensure reconstruction fidelity and a latent diffusion model trained with attribute supervision to enable controllable text generation. Attribute information is injected only during training. Experiments on five datasets spanning multiple stylistic attributes demonstrate that RegDiff outperforms strong baselines in generating stylistic texts. These results validate the effectiveness of RegDiff as an efficient solution for attribute-controllable text diffusion. Our code, datasets, and resources will be released upon publication at https://github.com/xxxx.

While de-identification models can help conceal the identity of the individual(s) mentioned in a document, they fail to address linkage risks, defined as the potential to map the de-identified text back to its source. One straightforward way to perform such linkages is to extract phrases from the de-identified document and then check their presence in the original dataset. This paper presents a method to counter search-based linkage attacks while preserving the semantic integrity of the text. The method proceeds in two steps. We first construct an inverted index of the N-grams occurring in the document collection, making it possible to efficiently determine which N-grams appear in less than $k$ documents (either alone or in combination with other N-grams). An LLM-based rewriter is then iteratively queried to reformulate those spans until linkage is no longer possible. Experimental results on a collection of court cases show that the method is able to effectively prevent search-based linkages while remaining faithful to the original content.