IC-092All evaluated LLMs show consistent F1 degradation to at most 0.60 when two or more events match a retrieval cue

Alexis Huet, Zied Ben Houidi, Dario Rossi

SourceEpisodic Memories Generation and Evaluation Benchmark for Large Language Models

On the long book (100k tokens, 196 events), every model and memory strategy tested shows a sharp drop in recall accuracy once a cue matches more than one event. For single-event cues, top models reach F1 of 0.81–0.97, but for two or more matching events the best F1 across all models and strategies is 0.60 or below. The effect is consistent across in-context, RAG, and fine-tuning settings, and persists on the shorter 10k-token book. The authors attribute this to cue overload: the model cannot disambiguate among multiple similar events sharing a cue dimension.

Evidence
correlational
Key metric
f1 ≤0.60 for all models when 2+ events match the cue (long book); e.g. GPT-4o in-context: 0.81 (1 event) → 0.60 (2 events) → 0.57 (3-5) → 0.53 (6+); O1-mini in-context: 0.05 (1 event) → 0.12 (2) → 0.12 (3-5) → 0.24 (6+)
Caveat
The benchmark is synthetic and uses explicit temporal/spatial markers; the authors note it may not capture implicit or relative temporal references in natural language.
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]
Related work
Needle-in-a-haystack [context], Babi tasks [context], Michelangelo (LSQ) [context]
Related findings
IC-093, IC-094, IC-095
Extraction
automatic-extraction