Jasin Cekinmez, Addison J. Wu, Thomas L. Griffiths
原文摘要
Positional biases such as recency and primacy effects have been documented in large language models (LLMs), yet the underlying mechanism by which these models make their evaluations remains poorly understood. Both primacy and recency biases have been observed in human judgments in response to evidence, but recent work suggest that when the listener updates their beliefs - during the presentation of evidence or only at the end - influences the presence of such effects. We investigate whether a similar phenomenon holds for LLMs, finding divergence from human behavior. These biases are more pronounced in newer models compared to their predecessors.
Chunru Lin, Hongxin Zhang, Fenghao Yu, Zhehuan Chen, Thomas L. Griffiths, Yejin Choi, David Held, Chuang Gan
原文摘要
The ability to reason, adapt, and creatively solve problems under unexpected challenges is essential for robots operating in real-world environments. However, current robotic benchmarks primarily emphasize skill-level execution and provide limited insight into such cognitive reasoning capabilities. We introduce RoboWits, a bi-manual robotic benchmark designed to systematically evaluate cognitive reasoning, creative tool use, and robustness to unexpected conditions. To enable scalable construction of high-quality reasoning-centric unexpected scenarios, we propose an automated task generation pipeline formulated as a multi-agent cooperative framework, comprising agents for seed task generation and verification, metric generation, scene generation, and task mutation. Using the pipeline, we curated 30 diverse seed tasks and 200 tasks with mutations and graded difficulty across geometry, material, and assembly-based reasoning. We benchmark popular robot policies, pre-trained VLAs, a zero-shot policy built on a frontier VLM, and oracle-state planners. Our results reveal a significant performance gap: while pre-trained VLAs exhibit preliminary success on seed tasks after fine-tuning, they struggle to perform on mutated tasks, implying their brittleness in manipulation tasks requiring reasoning, strategy adaptation, and robustness to deceptive or constrained environments. A zero-shot policy built on GPT-6 Astra outperforms all fine-tuned VLAs, yet still loses about a fifth of its success on mutated tasks, showing that stronger foundation models alone do not close the robustness gap. See our project page https://umass-embodied-agi.github.io/RoboWits/ for videos.
Ryan Liu, David Lee, Icey Siyi Ai, Marwa Abdulhai, Rachit Dubey, Thomas L. Griffiths
原文摘要
Today's large language models (LLMs) increasingly interact with time: AI agents schedule tasks, coordinate with people and other AI systems, and even decide how long to think. Yet LLMs only ever encounter time secondhand---in descriptions rather than experience. Does this hinder their reasoning and decisions that involve time? Our paper traces this question from a model's knowledge to its actions through four successive evaluations. We find that models know the rules of time, which can be read, but struggle to predict durations, which must be lived; while they are highly accurate (92.0-97.5\%) on a collection of questions built from the theorems of various time-related literatures, they underestimate their own completion times by up to 99.5\% and fail to identify which tasks will take them longer to run. And while they are better at estimating durations of human tasks, their duration judgments change based on how an event is worded, magnifying similar human biases by up to 1.3$\times$. Moving from estimation to action, we find that some LLMs disregard time costs completely when answering user queries, ignoring a 1,200-fold increase in wait time entirely. Together, our experiments reveal new challenges posed to LLMs as they navigate a dimension they cannot directly perceive, and highlight how their blind spots can even be simple cognitive abilities that humans take for granted.
Large language models (LLMs) solve complex problems yet fail on simpler variants, suggesting they achieve correct outputs through mechanisms fundamentally different from human reasoning. To understand this gap, we synthesize cognitive science research into a taxonomy of 28 cognitive elements spanning reasoning invariants, meta-cognitive controls, representations for organizing reasoning and knowledge, and transformation operations. We introduce a fine-grained evaluation framework and apply it to 192K traces from 18 models across text, vision, and audio, alongside 1,346 human think-aloud traces on a new, domain-agnostic dataset for human-model comparison. We find that models are able to deploy cognitive elements seen in human reasoning but differ in how they use them. Leveraging these patterns, we develop test-time reasoning guidance that automatically scaffolds successful reasoning topologies, improving the accuracy of the strongest models by 9.7 percentage points on ill-structured problems. Meta-analysis of 1.6K LLM reasoning papers reveals the research community concentrates on easily quantifiable elements but rarely targets self-awareness, whose use differs between human and model reasoning. By establishing a shared vocabulary between cognitive science and LLM research, our framework enables diagnosis of reasoning failures and principled development of models that reason through robust cognitive mechanisms rather than spurious shortcuts, while providing tools to test theories of human cognition at scale.
Kaiqu Liang, Jaime Fernández Fisac, Thomas L. Griffiths, Andrew Lampinen
原文摘要
Language agents increasingly use persistent memory to record user preferences and procedures across sessions. We show that memory from benign interactions can cause unsafe or unreliable behavior on later tasks. We call this failure the memory trap. To study this issue, we compare behavior in sessions with and without memory across six frontier models (including Claude Sonnet, Opus, and Fable models and GPT-5.6). We evaluate these models in a range of settings. First, in autonomy experiments, we find that requests to skip confirmation during low-stakes coding chores lead agents to skip confirmation and verification on simulated high-stakes tasks in both coding and everyday life. We next study memories formed when users seek emotional support, or describe harm from uncritical feedback, and find these memories bias later judgments of mathematical work, arguments, and poems toward sycophancy or over-critique (respectively). Memory notes written during legitimate stale-test repairs lead agents to weaken correct tests that catch broken safeguards; these memories also increase cheating on ImpossibleBench, with larger increases for stronger Sonnet and Opus models. We show that memory traps can arise both from overgeneralizations in memory writing, and when acting on memory. These findings demonstrate that even benign interactions can lead to serious risks from memory systems, and thus motivate the need for new research into ensuring that agents preserve the scope of their memories.
Ruairidh M. Battleday, Kai Sandbrink, Jimi Cullen-Drohan, Zihan Yan, Timothy Muller, Clare Maguire, Ales Kubicek, Fraser Greenlee-Scott 等 16 位作者
原文摘要
Discovery---formulating novel generalizations---is a central part of the scientific process. Despite its importance, there is a gap in the current AI benchmark landscape, with few benchmarks directly probing the capacity for discovering new knowledge with experimentation in controlled environments where the objective is unknown. To address this gap, we release a new benchmark: DiG-bench (Discovery in Games). DiG-bench consists of a set of 70 independent games. Each game is encoded as a short string and has unique transformation rules that must be discovered through interaction and experimentation. The levels of the game present a series of challenges to test whether the rules have been discovered, where the win conditions for each level are also unknown. We provide games at seven tiers of difficulty for AI agents. The lowest tier is routinely solvable by multiple models, while the highest tier challenges the best models in agentic harnesses. All 70 games were solved by at least one human on first attempt. A subset of 21 games is released publicly, and the remainder is held private for secure evaluation.