We propose a novel architecture for integrating large language models (LLMs) with a persistent, interactive Lisp environment. This setup enables LLMs to define, invoke, and evolve their own tools through programmatic interaction with a live REPL. By embedding Lisp expressions within generation and intercepting them via a middleware layer, the system allows for stateful external memory, reflective programming, and dynamic tool creation. We present a design framework and architectural principles to guide future implementations of interactive AI systems that integrate symbolic programming with neural language generation.

Law has long been a domain that has been popular in natural language processing (NLP) applications. Reasoning (ratiocination and the ability to make connections to precedent) is a core part of the practice of the law in the real world. Nevertheless, while multiple legal datasets exist, none have thus far focused specifically on reasoning tasks. We focus on a specific aspect of the legal landscape by introducing a corporate governance reasoning benchmark (CHANCERY) to test a model's ability to reason about whether executive/board/shareholder's proposed actions are consistent with corporate governance charters. This benchmark introduces a first-of-its-kind corporate governance reasoning test for language models - modeled after real world corporate governance law. The benchmark consists of a corporate charter (a set of governing covenants) and a proposal for executive action. The model's task is one of binary classification: reason about whether the action is consistent with the rules contained within the charter. We create the benchmark following established principles of corporate governance - 24 concrete corporate governance principles established in and 79 real life corporate charters selected to represent diverse industries from a total dataset of 10k real life corporate charters. Evaluations on state-of-the-art (SOTA) reasoning models confirm the difficulty of the benchmark, with models such as Claude 3.7 Sonnet and GPT-4o achieving 64.5% and 75.2% accuracy respectively. Reasoning agents exhibit superior performance, with agents based on the ReAct and CodeAct frameworks scoring 76.1% and 78.1% respectively, further confirming the advanced legal reasoning capabilities required to score highly on the benchmark. We also conduct an analysis of the types of questions which current reasoning models struggle on, revealing insights into the legal reasoning capabilities of SOTA models.

We introduce SealQA, a new challenge benchmark for evaluating SEarch-Augmented Language models on fact-seeking questions where web search yields conflicting, noisy, or unhelpful results. SealQA comes in three flavors: (1) Seal-0 (main) and (2) Seal-Hard, which assess factual accuracy and reasoning capabilities, with Seal-0 focusing on the most challenging questions where chat models (e.g., GPT-4.1) typically achieve near-zero accuracy; and (3) LongSeal, which extends SealQA to test long-context, multi-document reasoning in "needle-in-a-haystack" settings. Our evaluation reveals critical limitations in current models: Even frontier LLMs perform poorly across all SealQA flavors. On Seal-0, frontier agentic models equipped with tools like o3 and o4-mini achieve only 17.1% and 6.3% accuracy, respectively, at their best reasoning efforts. We find that advanced reasoning models such as DeepSeek-R1-671B and o3-mini are highly vulnerable to noisy search results. Notably, increasing test-time compute does not yield reliable gains across o3-mini, o4-mini, and o3, with performance often plateauing or even declining early. Additionally, while recent models are less affected by the "lost-in-the-middle" issue, they still fail to reliably identify relevant documents in LongSeal when faced with numerous distractors. To facilitate future work, we release SealQA at huggingface.co/datasets/vtllms/sealqa.

Recent advancements in Large Language Models (LLMs) and multimodal counterparts have spurred significant interest in developing web agents -- AI systems capable of autonomously navigating and completing tasks within web environments. While holding tremendous promise for automating complex web interactions, current approaches face substantial challenges due to the fundamental mismatch between human-designed interfaces and LLM capabilities. Current methods struggle with the inherent complexity of web inputs, whether processing massive DOM trees, relying on screenshots augmented with additional information, or bypassing the user interface entirely through API interactions. This position paper advocates for a paradigm shift in web agent research: rather than forcing web agents to adapt to interfaces designed for humans, we should develop a new interaction paradigm specifically optimized for agentic capabilities. To this end, we introduce the concept of an Agentic Web Interface (AWI), an interface specifically designed for agents to navigate a website. We establish six guiding principles for AWI design, emphasizing safety, efficiency, and standardization, to account for the interests of all primary stakeholders. This reframing aims to overcome fundamental limitations of existing interfaces, paving the way for more efficient, reliable, and transparent web agent design, which will be a collaborative effort involving the broader ML community.

We introduce~\textsc{Domain2Vec}, a novel approach that decomposes any dataset into a linear combination of several \emph{meta-domains}, a new concept designed to capture the key underlying features of datasets. \textsc{Domain2Vec} maintains a vocabulary of meta-domains and uses a classifier to decompose any given dataset into a domain vector that corresponds to a distribution over this vocabulary. These domain vectors enable the identification of the optimal data mixture for language model (LM) pretraining in a training-free manner under the \emph{\textbf{D}istribution \textbf{A}lignment \textbf{A}ssumption} (DA$^{2}$), which suggests that when the data distributions of the training set and the validation set are better aligned, a lower validation loss is achieved. Moreover, \textsc{Domain2vec} can be seamlessly integrated into previous works to model the relationship between domain vectors and LM performance, greatly enhancing the efficiency and scalability of previous methods. Extensive experiments demonstrate that \textsc{Domain2Vec} helps find the data mixture that enhances downstream task performance with minimal computational overhead. Specifically, \textsc{Domain2Vec} achieves the same validation loss on Pile-CC using only $51.5\%$ of the computation required when training on the original mixture of The Pile dataset. Under equivalent compute budget, \textsc{Domain2Vec} improves downstream performance by an average of $2.83\%$.

We show that reinforcement learning with verifiable rewards (RLVR) can elicit strong mathematical reasoning in certain models even with spurious rewards that have little, no, or even negative correlation with the correct answer. For example, RLVR improves MATH-500 performance for Qwen2.5-Math-7B in absolute points by 21.4% (random reward), 13.8% (format reward), 24.1% (incorrect label), 26.0% (1-shot RL), and 27.1% (majority voting) -- nearly matching the 29.1% gained with ground truth rewards. However, the spurious rewards that work for Qwen often fail to yield gains with other model families like Llama3 or OLMo2. In particular, we find code reasoning -- thinking in code without actual code execution -- to be a distinctive Qwen2.5-Math behavior that becomes significantly more frequent after RLVR, from 65% to over 90%, even with spurious rewards. Overall, we hypothesize that, given the lack of useful reward signal, RLVR must somehow be surfacing useful reasoning representations learned during pretraining, although the exact mechanism remains a topic for future work. We suggest that future RLVR research should possibly be validated on diverse models rather than a single de facto choice, as we show that it is easy to get significant performance gains on Qwen models even with completely spurious reward signals.

In-context image editing aims to modify images based on a contextual sequence comprising text and previously generated images. Existing methods typically depend on task-specific pipelines and expert models (e.g., segmentation and inpainting) to curate training data. In this work, we explore whether an in-context image editing model can be learned directly from videos. We introduce a scalable approach to annotate videos as interleaved multimodal sequences. To effectively learn from this data, we design a block-causal diffusion transformer trained on three proxy tasks: next-image prediction, current segmentation prediction, and next-segmentation prediction. Additionally, we propose a novel multi-turn image editing benchmark to advance research in this area. Extensive experiments demonstrate that our model exhibits strong in-context image editing capabilities and achieves state-of-the-art results on two multi-turn image editing benchmarks. Despite being trained exclusively on videos, our model also shows promising abilities in multi-concept composition, story generation, and chain-of-editing applications.

Reduced social connectedness increasingly poses a threat to mental health, life expectancy, and general well-being. Generative AI (GAI) technologies, such as large language models (LLMs) and image generation tools, are increasingly integrated into applications aimed at enhancing human social experiences. Despite their growing presence, little is known about how these technologies influence social interactions. This scoping review investigates how GAI-based applications are currently designed to facilitate social interaction, what forms of social engagement they target, and which design and evaluation methodologies designers use to create and evaluate them. Through an analysis of 30 studies published since 2020, we identify key trends in application domains including storytelling, socio-emotional skills training, reminiscence, collaborative learning, music making, and general conversation. We highlight the role of participatory and co-design approaches in fostering both effective technology use and social engagement, while also examining socio-ethical concerns such as cultural bias and accessibility. This review underscores the potential of GAI to support dynamic and personalized interactions, but calls for greater attention to equitable design practices and inclusive evaluation strategies.

Large language models (LLMs) are increasingly deployed in high-stakes hiring applications, making decisions that directly impact people's careers and livelihoods. While prior studies suggest simple anti-bias prompts can eliminate demographic biases in controlled evaluations, we find these mitigations fail when realistic contextual details are introduced. We address these failures through internal bias mitigation: by identifying and neutralizing sensitive attribute directions within model activations, we achieve robust bias reduction across all tested scenarios. Across leading commercial (GPT-4o, Claude 4 Sonnet, Gemini 2.5 Flash) and open-source models (Gemma-2 27B, Gemma-3, Mistral-24B), we find that adding realistic context such as company names, culture descriptions from public careers pages, and selective hiring constraints (e.g.,``only accept candidates in the top 10\%") induces significant racial and gender biases (up to 12\% differences in interview rates). When these biases emerge, they consistently favor Black over White candidates and female over male candidates across all tested models and scenarios. Moreover, models can infer demographics and become biased from subtle cues like college affiliations, with these biases remaining invisible even when inspecting the model's chain-of-thought reasoning. To address these limitations, our internal bias mitigation identifies race and gender-correlated directions and applies affine concept editing at inference time. Despite using directions from a simple synthetic dataset, the intervention generalizes robustly, consistently reducing bias to very low levels (typically under 1\%, always below 2.5\%) while largely maintaining model performance. Our findings suggest that practitioners deploying LLMs for hiring should adopt more realistic evaluation methodologies and consider internal mitigation strategies for equitable outcomes.

A central goal for mechanistic interpretability has been to identify the right units of analysis in large language models (LLMs) that causally explain their outputs. While early work focused on individual neurons, evidence that neurons often encode multiple concepts has motivated a shift toward analyzing directions in activation space. A key question is how to find directions that capture interpretable features in an unsupervised manner. Current methods rely on dictionary learning with sparse autoencoders (SAEs), commonly trained over residual stream activations to learn directions from scratch. However, SAEs often struggle in causal evaluations and lack intrinsic interpretability, as their learning is not explicitly tied to the computations of the model. Here, we tackle these limitations by directly decomposing MLP activations with semi-nonnegative matrix factorization (SNMF), such that the learned features are (a) sparse linear combinations of co-activated neurons, and (b) mapped to their activating inputs, making them directly interpretable. Experiments on Llama 3.1, Gemma 2 and GPT-2 show that SNMF derived features outperform SAEs and a strong supervised baseline (difference-in-means) on causal steering, while aligning with human-interpretable concepts. Further analysis reveals that specific neuron combinations are reused across semantically-related features, exposing a hierarchical structure in the MLP's activation space. Together, these results position SNMF as a simple and effective tool for identifying interpretable features and dissecting concept representations in LLMs.