Chelsea Zou

Emmy Liu, Varun Gangal, Chelsea Zou et al. · cmu

Despite numerous attempts at mitigation since the inception of language models, hallucinations remain a persistent problem even in today's frontier LLMs. Why is this? We review existing definitions of hallucination and fold them into a single, unified definition wherein prior definitions are subsumed. We argue that hallucination can be unified by defining it as simply inaccurate (internal) world modeling, in a form where it is observable to the user. For example, stating a fact which contradicts a knowledge base OR producing a summary which contradicts the source. By varying the reference world model and conflict policy, our framework unifies prior definitions. We argue that this unified view is useful because it forces evaluations to clarify their assumed reference "world", distinguishes true hallucinations from planning or reward errors, and provides a common language for comparison across benchmarks and discussion of mitigation strategies. Building on this definition, we outline plans for a family of benchmarks using synthetic, fully specified reference world models to stress-test and improve world modeling components.

Long Phan, Alice Gatti, Ziwen Han et al. · amazon-science, apple-ml

Benchmarks are important tools for tracking the rapid advancements in large language model (LLM) capabilities. However, benchmarks are not keeping pace in difficulty: LLMs now achieve over 90\% accuracy on popular benchmarks like MMLU, limiting informed measurement of state-of-the-art LLM capabilities. In response, we introduce Humanity's Last Exam (HLE), a multi-modal benchmark at the frontier of human knowledge, designed to be the final closed-ended academic benchmark of its kind with broad subject coverage. HLE consists of 2,500 questions across dozens of subjects, including mathematics, humanities, and the natural sciences. HLE is developed globally by subject-matter experts and consists of multiple-choice and short-answer questions suitable for automated grading. Each question has a known solution that is unambiguous and easily verifiable, but cannot be quickly answered via internet retrieval. State-of-the-art LLMs demonstrate low accuracy and calibration on HLE, highlighting a significant gap between current LLM capabilities and the expert human frontier on closed-ended academic questions. To inform research and policymaking upon a clear understanding of model capabilities, we publicly release HLE at https://lastexam.ai.

Chelsea Zou, Yiheng Yao, Selena She et al.

We introduce CalBench, a controlled evaluation environment for studying multi-agent coordination through calendar scheduling. In CalBench, N agents each manage a private calendar containing pre-existing commitments and must coordinate to schedule a stream of M incoming meetings while minimizing disruption costs. Because agents observe only their own calendars, successful scheduling requires communication across private information boundaries. Each scenario is generated with an oracle solution, enabling precise measurement of coordination quality via realized-to-optimal cost, as well as a Distributed Constraint Optimization (DCOP) baseline to provide a fair comparison under the same private-information constraints. CalBench enables precise verification of task success, communication efficiency, and fairness in the distribution of disruption costs. Our environment also studies privacy-preserving coordination by augmenting calendar entries with private semantic contexts of varying sensitivity and measuring whether agents reveal task-irrelevant private information during negotiation. Unlike multi-agent benchmarks where a single capable agent can often substitute for the group, CalBench is inherently decentralized: no agent has access to another agent's private calendar, yet agents must still reach mutually consistent decisions over shared meeting scheduling. CalBench therefore provides a practical and verifiable setting for studying coordination protocols, communication efficiency, negotiation strategies, fairness, and privacy leakage in multi-agent systems.

Yiheng Yao, Chelsea Zou, Robert D. Hawkins

Grounding is the collaborative process of establishing mutual belief sufficient for the current communicative purpose. While static grounding maps language to a shared, externally observable context, dynamic grounding is a joint activity where meaning is negotiated through interaction. Current multi-agent Large Language Model (LLM) benchmarks focus on static, one-shot tasks, overlooking the ability to repair grounding breakdowns across turns. We introduce an iterated, multi-turn negotiation game in which two agents allocate shared resources toward private projects with verifiable jointly optimal outcomes. While individual agents can identify Pareto-optimal allocations in isolation, agent dyads consistently fail to reach them across open- and closed-source models. Our investigation reveals four failure modes: (1) coordination degrades when shared interaction history is absent; (2) yet accumulated context can itself become a liability through stubborn anchoring, where initial proposals are treated as axiomatic rather than negotiable; (3) a reliance on perfunctory fairness (equal resource splits) over reward-maximizing coordination; and (4) failures in referential binding, where agents lose track of commitments across turns. These results highlight dynamic grounding as a critical and understudied axis of multi-agent coordination. Our framework decomposes the coordination gap into measurable components: the oracle baseline establishes that the gap is not attributable to individual reasoning limitations; the no-talk baseline establishes that communication is necessary; and a full-transparency intervention establishes that information exchange alone is insufficient: the bottleneck lies in the interactive processes of joint plan formation, commitment, and execution that constitute dynamic grounding.

Chelsea Zou, Kenneth J. Kurtz

We introduce a cluster-based generative image segmentation framework to encode higher-level representations of visual concepts based on one-shot learning inspired by the Omniglot Challenge. The inferred parameters of each component of a Gaussian Mixture Model (GMM) represent a distinct topological subpart of a visual concept. Sampling new data from these parameters generates augmented subparts to build a more robust prototype for each concept, i.e., the Abstracted Gaussian Prototype (AGP). This framework addresses one-shot classification tasks using a cognitively-inspired similarity metric and addresses one-shot generative tasks through a novel AGP-VAE pipeline employing variational autoencoders (VAEs) to generate new class variants. Results from human judges reveal that the generative pipeline produces novel examples and classes of visual concepts that are broadly indistinguishable from those made by humans. The proposed framework leads to impressive, but not state-of-the-art, classification accuracy; thus, the contribution is two-fold: 1) the system is low in theoretical and computational complexity yet achieves the standard of 'true' one-shot learning by operating in a fully standalone manner unlike existing approaches that draw heavily on pre-training or knowledge engineering; and 2) in contrast with existing neural network approaches, the AGP approach addresses the importance of broad task capability emphasized in the Omniglot challenge (successful performance on classification and generative tasks). These two points are critical in advancing our understanding of how learning and reasoning systems can produce viable, robust, and flexible concepts based on literally no more than a single example.

Chelsea Zou, Yiheng Yao, Basant Khalil

This project develops a self correcting framework for large language models (LLMs) that detects and mitigates hallucinations during multi-step reasoning. Rather than relying solely on final answer correctness, our approach leverages fine grained uncertainty signals: 1) self-assessed confidence alignment, and 2) token-level entropy spikes to detect unreliable and unfaithful reasoning in real time. We design a composite reward function that penalizes unjustified high confidence and entropy spikes, while encouraging stable and accurate reasoning trajectories. These signals guide a reinforcement learning (RL) policy that makes the model more introspective and shapes the model's generation behavior through confidence-aware reward feedback, improving not just outcome correctness but the coherence and faithfulness of their intermediate reasoning steps. Experiments show that our method improves both final answer accuracy and reasoning calibration, with ablations validating the individual contribution of each signal.