Multimodal LLMs are turning their focus to video benchmarks, however most video benchmarks only provide outcome supervision, with no intermediate or interpretable reasoning steps. This makes it challenging to assess if models are truly able to combine perceptual and temporal information to reason about videos, or simply get the correct answer by chance or by exploiting linguistic biases. To remedy this, we provide a new video reasoning dataset called MINERVA for modern multimodal models. Each question in the dataset comes with 5 answer choices, as well as detailed, hand-crafted reasoning traces. Our dataset is multimodal, diverse in terms of video domain and length, and consists of complex multi-step questions. Extensive benchmarking shows that our dataset provides a challenge for frontier open-source and proprietary models. We perform fine-grained error analysis to identify common failure modes across various models, and create a taxonomy of reasoning errors. We use this to explore both human and LLM-as-a-judge methods for scoring video reasoning traces, and find that failure modes are primarily related to temporal localization, followed by visual perception errors, as opposed to logical or completeness errors. The dataset, along with questions, answer candidates and reasoning traces will be publicly available under https://github.com/google-deepmind/neptune?tab=readme-ov-file\#minerva.

As Large Language Models (LLMs) are rapidly evolving, providing accurate feedback and scalable oversight on their outputs becomes an urgent and critical problem. Leveraging LLMs as critique models to achieve automated supervision is a promising solution. In this work, we focus on studying and enhancing the math critique ability of LLMs. Current LLM critics provide critiques that are too shallow and superficial on each step, leading to low judgment accuracy and struggling to offer sufficient feedback for the LLM generator to correct mistakes. To tackle this issue, we propose a novel and effective two-stage framework to develop LLM critics that are capable of deliberately critiquing on each reasoning step of math solutions. In the first stage, we utilize Qwen2.5-72B-Instruct to generate 4.5K long-form critiques as seed data for supervised fine-tuning. Each seed critique consists of deliberate step-wise critiques that includes multi-perspective verifications as well as in-depth critiques of initial critiques for each reasoning step. Then, we perform reinforcement learning on the fine-tuned model with either existing human-labeled data from PRM800K or our automatically annotated data obtained via Monte Carlo sampling-based correctness estimation, to further incentivize its critique ability. Our developed critique model built on Qwen2.5-7B-Instruct not only significantly outperforms existing LLM critics (including the same-sized DeepSeek-R1-distill models and GPT-4o) on various error identification benchmarks, but also more effectively helps the LLM generator refine erroneous steps through more detailed feedback.

Large Language Models (LLMs) have shown impressive zero-shot performance across a variety of Natural Language Processing tasks, including document re-ranking. However, their effectiveness degrades on unseen tasks and domains, largely due to shifts in vocabulary and word distributions. In this paper, we investigate Task Arithmetic, a technique that combines the weights of LLMs pre-trained on different tasks or domains via simple mathematical operations, such as addition or subtraction, to adapt retrieval models without requiring additional fine-tuning. Our method is able to synthesize diverse tasks and domain knowledge into a single model, enabling effective zero-shot adaptation in different retrieval contexts. Extensive experiments on publicly available scientific, biomedical, and multilingual datasets show that our method improves state-of-the-art re-ranking performance by up to 18% in NDCG@10 and 15% in P@10. In addition to these empirical gains, our analysis provides insights into the strengths and limitations of Task Arithmetic as a practical strategy for zero-shot learning and model adaptation. We make our code publicly available at https://github.com/DetectiveMB/Task-Arithmetic-for-ZS-IR.

In response to the lack of trust in Artificial Intelligence (AI) for sequential planning, we design a Computational Tree Logic-guided large language model (LLM)-based natural language explanation framework designed for the Monte Carlo Tree Search (MCTS) algorithm. MCTS is often considered challenging to interpret due to the complexity of its search trees, but our framework is flexible enough to handle a wide range of free-form post-hoc queries and knowledge-based inquiries centered around MCTS and the Markov Decision Process (MDP) of the application domain. By transforming user queries into logic and variable statements, our framework ensures that the evidence obtained from the search tree remains factually consistent with the underlying environmental dynamics and any constraints in the actual stochastic control process. We evaluate the framework rigorously through quantitative assessments, where it demonstrates strong performance in terms of accuracy and factual consistency.

This study investigates whether large language models, specifically GPT4, can match human capabilities in analogical reasoning within strategic decision making contexts. Using a novel experimental design involving source to target matching, we find that GPT4 achieves high recall by retrieving all plausible analogies but suffers from low precision, frequently applying incorrect analogies based on superficial similarities. In contrast, human participants exhibit high precision but low recall, selecting fewer analogies yet with stronger causal alignment. These findings advance theory by identifying matching, the evaluative phase of analogical reasoning, as a distinct step that requires accurate causal mapping beyond simple retrieval. While current LLMs are proficient in generating candidate analogies, humans maintain a comparative advantage in recognizing deep structural similarities across domains. Error analysis reveals that AI errors arise from surface level matching, whereas human errors stem from misinterpretations of causal structure. Taken together, the results suggest a productive division of labor in AI assisted organizational decision making where LLMs may serve as broad analogy generators, while humans act as critical evaluators, applying the most contextually appropriate analogies to strategic problems.

Hallucinations in large language models (LLMs) present a growing challenge across real-world applications, from healthcare to law, where factual reliability is essential. Despite advances in alignment and instruction tuning, LLMs can still generate outputs that are fluent yet fundamentally untrue. Understanding the cognitive dynamics that underlie these hallucinations remains an open problem. In this study, we propose a prompt-based framework to systematically trigger and quantify hallucination: a Hallucination-Inducing Prompt (HIP), which synthetically fuses semantically distant concepts (e.g., periodic table of elements and tarot divination) in a misleading way, and a Hallucination Quantifying Prompt (HQP), which scores the plausibility, confidence, and coherence of the output. Controlled experiments across multiple LLMs revealed that HIPs consistently produced less coherent and more hallucinated responses than their null-fusion controls. These effects varied across models, with reasoning-oriented LLMs showing distinct profiles from general-purpose ones. Our framework provides a reproducible testbed for studying hallucination vulnerability, and opens the door to developing safer, more introspective LLMs that can detect and self-regulate the onset of conceptual instability.

Generalizing language-conditioned multi-task imitation learning (IL) models to novel long-horizon 3D manipulation tasks remains a significant challenge. To address this, we propose DeCo (Task Decomposition and Skill Composition), a model-agnostic framework compatible with various multi-task IL models, designed to enhance their zero-shot generalization to novel, compositional, long-horizon 3D manipulation tasks. DeCo first decomposes IL demonstrations into a set of modular atomic tasks based on the physical interaction between the gripper and objects, and constructs an atomic training dataset that enables models to learn a diverse set of reusable atomic skills during imitation learning. At inference time, DeCo leverages a vision-language model (VLM) to parse high-level instructions for novel long-horizon tasks, retrieve the relevant atomic skills, and dynamically schedule their execution; a spatially-aware skill-chaining module then ensures smooth, collision-free transitions between sequential skills. We evaluate DeCo in simulation using DeCoBench, a benchmark specifically designed to assess zero-shot generalization of multi-task IL models in compositional long-horizon 3D manipulation. Across three representative multi-task IL models (RVT-2, 3DDA, and ARP), DeCo achieves success rate improvements of 66.67%, 21.53%, and 57.92%, respectively, on 12 novel compositional tasks. Moreover, in real-world experiments, a DeCo-enhanced model trained on only 6 atomic tasks successfully completes 9 novel long-horizon tasks, yielding an average success rate improvement of 53.33% over the base multi-task IL model. Video demonstrations are available at: https://deco226.github.io.

A growing intuition in machine learning suggests a link between sparsity and in-terpretability. We introduce a novel self-ablation mechanism to investigate this connection ante-hoc in the context of language transformers. Our approach dynamically enforces a k-winner-takes-all constraint, forcing the model to demonstrate selective activation across neuron and attention units. Unlike post-hoc methods that analyze already-trained models, our approach integrates interpretabil-ity directly into model training, promoting feature localization from inception. Training small models on the TinyStories dataset and employing interpretabil-ity tests, we find that self-ablation leads to more localized circuits, concentrated feature representations, and increased neuron specialization without compromising language modelling performance. Surprisingly, our method also decreased overall sparsity, indicating that self-ablation promotes specialization rather than widespread inactivity. This reveals a complex interplay between sparsity and in-terpretability, where decreased global sparsity can coexist with increased local specialization, leading to enhanced interpretability. To facilitate reproducibility, we make our code available at https://github.com/keenanpepper/ As machine learning systems are entrusted with increasingly complex tasks, our ability to understand their decision-making processes lags behind their growing capabilities (OpenAI, 2024; Gemini Team, Google, 2024). Much of the current research in interpretability for LLMs focuses on developing post-hoc methods, attempting to explain the behaviour of already-trained models (Ribeiro et al., 2016; Conmy et al., 2023; Foote et al., 2023b; Bills et al., 2023; Huben et al., 2024). While valuable, these approaches often provide only an approximate or incomplete understanding of the underlying mechanisms (Rudin, 2019). A more fundamental yet less studied approach involves designing models to be inherently more interpretable, an ante-hoc approach, where transparency is woven into the architecture itself (Slavin et al., 2018; Tamkin et al., 2023; Cloud et al., 2024; Liu et al., 2024).

As large language models (LLMs) are increasingly deployed in high-stakes domains, detecting hallucinated content$\unicode{x2013}$text that is not grounded in supporting evidence$\unicode{x2013}$has become a critical challenge. Existing benchmarks for hallucination detection are often synthetically generated, narrowly focused on extractive question answering, and fail to capture the complexity of real-world scenarios involving multi-document contexts and full-sentence outputs. We introduce the HalluMix Benchmark, a diverse, task-agnostic dataset that includes examples from a range of domains and formats. Using this benchmark, we evaluate seven hallucination detection systems$\unicode{x2013}$both open and closed source$\unicode{x2013}$highlighting differences in performance across tasks, document lengths, and input representations. Our analysis highlights substantial performance disparities between short and long contexts, with critical implications for real-world Retrieval Augmented Generation (RAG) implementations. Quotient Detections achieves the best overall performance, with an accuracy of 0.82 and an F1 score of 0.84.

Identifying regulatory statements in legislation is useful for developing metrics to measure the regulatory density and strictness of legislation. A computational method is valuable for scaling the identification of such statements from a growing body of EU legislation, constituting approximately 180,000 published legal acts between 1952 and 2023. Past work on extraction of these statements varies in the permissiveness of their definitions for what constitutes a regulatory statement. In this work, we provide a specific definition for our purposes based on the institutional grammar tool. We develop and compare two contrasting approaches for automatically identifying such statements in EU legislation, one based on dependency parsing, and the other on a transformer-based machine learning model. We found both approaches performed similarly well with accuracies of 80% and 84% respectively and a K alpha of 0.58. The high accuracies and not exceedingly high agreement suggests potential for combining strengths of both approaches.