IC-095Evaluated LLMs achieve at most 36% latest-state accuracy and 18% full-set accuracy on multi-event entity tracking, with low Kendall's tau on chronological ordering

Alexis Huet, Zied Ben Houidi, Dario Rossi

SourceEpisodic Memories Generation and Evaluation Benchmark for Large Language Models

On questions requiring tracking an entity's state across multiple events (ground truth length ≥ 2), no model exceeds 36% on identifying the correct latest state, and no model exceeds 18% on retrieving the complete set of all states. Even when models do retrieve the correct events, they frequently fail to order them chronologically: Kendall's tau coefficients range from 0.33 to 1.00, with most models at or below 0.60. Llama 3.1-405B and O1-mini achieve 0% on full-set recall. The authors conclude this reveals a significant gap in models' ability to process temporal relationships between events.

Evidence
correlational
Key metric
Latest state: max 36% (GPT-4o-mini RAG, GPT-4o in-context); All states: max 18% (Claude 3 Haiku RAG); Kendall's tau: 0.43 (Claude 3 Haiku RAG), 0.93 (GPT-4o-mini RAG), 0.60 (Claude 3.5 Sonnet RAG), 0.50 (GPT-4o RAG, GPT-4o in-context), 0.33 (GPT-4o-mini in-context); Llama 3.1-405B and O1-mini: 0% all states, n.a. Kendall's tau
Caveat
Kendall's tau is computed only on answers that fully match the ground truth set, so it measures ordering accuracy conditional on correct retrieval; the 0% all-states scores for Llama 3.1 and O1-mini mean no ordering could be assessed.
Model
GPT-4o mini, Claude 3 Haiku, Claude 3.5 Sonnet, Llama 3.1 405B Instruct, O1 / OpenAI-o1-preview O1-mini
Concepts
Failure mode
Datasets
Episodic Memory Benchmark / Episodic Memory Benchmark (short book) [eval]
Methods
In-Context Learning / In-context learning prompt [primary], Retrieval-Augmented Generation / Retrieval augmentation (top-5 chunks) [primary], Kendall's tau / Kendall-tau [eval]
Related findings
IC-092, IC-093, IC-094
Extraction
automatic-extraction