Tag
This paper introduces the readout-mediator angle to demonstrate that linear probes can decode information from language model activations that is orthogonal to the model's actual causal computation, undermining probe-based interpretability. The finding replicates across model scales and families, revealing a fundamental failure mode in using probes for mechanistic understanding or safety monitoring.
This paper introduces AsyncTool, a benchmark for evaluating LLM-based agents' asynchronous function calling abilities in multi-task scenarios with delayed tool responses. It proposes efficiency-oriented metrics and identifies key failure modes of current tool-using agents.
LiFT is a longitudinal instruction fine-tuning framework that unifies diverse temporal NLP tasks under a shared instruction schema with curriculum-based training. Evaluated across OLMo, LLaMA, and Qwen models, LiFT consistently outperforms base-model in-context learning, especially on out-of-distribution data and rare change events.
This paper introduces Zep, a temporal knowledge graph architecture for agent memory that outperforms MemGPT in benchmarks like DMR and LongMemEval. It highlights Zep's ability to handle dynamic knowledge integration and temporal reasoning for enterprise use cases.