A long-video answer is evidence-supported only when the frames decoded from the video cover every event the answer depends on. Existing evaluations score final-answer correctness or predicted evidence intervals, but the frames a method decodes before answering are rarely audited, so correct answers can still rest on incomplete observation. We introduce VES-Bench, a 600-question benchmark of Temporal Ordering and Event Counting items over 348 public long videos. Each item carries a jointly necessary set of evidence intervals, letting us audit at three strictness levels whether a method’s decoded frames cover every one of them. We also propose TRACE, a training-free agent that grounds answers in raw visual clips, builds an evidence bundle round by round, and stops only when the answer stabilises as the bundle grows and a final pass over the same clips returns the same answer. Under a same-backbone audit, TRACE answers 50.7% of questions correctly with at least two decoded frames inside every evidence interval, at 98.7 frames per question: over 10 points above uniform decoding at 128 frames (40.2%), and within 2.6 points of uniform decoding at 256 frames at 0.39× its frame cost, while reaching the highest answer accuracy in the audit (63.5%). TRACE also stays competitive on Video-MME (86.1), LVBench (75.6), and LongVideoBench (75.1).

We identify semantically coherent, context-consistent network components in large language models (LLMs) using coactivation of sparse autoencoder (SAE) features collected from just a handful of prompts. Focusing on concept-relation prediction tasks, we show that ablating these components for concepts (e.g., countries and words) and relations (e.g., capital city and translation language) changes model outputs in predictable ways, while amplifying these components induces counterfactual responses. Notably, composing relation and concept components yields compound counterfactual outputs. Further analysis reveals that while most concept components emerge from the very first layer, more abstract relation components are concentrated in later layers. Lastly, we show that extracted components more comprehensively capture concepts and relations than individual features while maintaining specificity. Overall, our findings suggest a modular organization of knowledge and advance methods for efficient, targeted LLM manipulation.

Cognitive tasks originally developed for humans are now increasingly used to study language models. While applying these tasks is often straightforward, interpreting their results can be challenging. In particular, when a model underperforms, it is often unclear whether this results from a limitation in the cognitive ability being tested or a failure to understand the task itself. A recent study argues that GPT 3.5’s declining performance on 2-back and 3-back tasks reflects a working memory capacity limit similar to humans (Gong et al., 2024). By analyzing a range of open-source language models of varying performance levels on these tasks, we show that the poor performance is due at least in part to a limitation in task comprehension and task set maintenance. We challenge the best-performing model with progressively harder versions of the task (up to 10-back) and experiment with alternative prompting strategies, before analyzing model attentions. Our larger aim is to contribute to the ongoing conversation around refining methodologies for the cognitive evaluation of language models.

Analogical reasoning is a fundamental capacity of human cognition that allows us to reason abstractly about novel situations by relating them to past experiences. While it is thought to be essential for robust reasoning in AI systems, conventional approaches require significant training and/or hard-coding of domain knowledge to be applied to benchmark tasks. Inspired by cognitive science research that has found connections between human language and analogy-making, we explore the use of intuitive language-based abstractions to support analogy in AI systems. Specifically, we apply large pre-trained language models (PLMs) to visual Raven’s Progressive Matrices (RPM), a common relational reasoning test. By simply encoding the perceptual features of the problem into language form, we find that PLMs exhibit a striking capacity for zero-shot relational reasoning, exceeding human performance and nearing supervised vision-based methods. We explore different encodings that vary the level of abstraction over task features, finding that higher-level abstractions further strengthen PLMs’ analogical reasoning. Our detailed analysis reveals insights on the role of model complexity, in-context learning, and prior knowledge in solving RPM tasks.