Language agents
PulseAugur coverage of Language agents — every cluster mentioning Language agents across labs, papers, and developer communities, ranked by signal.
1 day(s) with sentiment data
-
New framework tackles adaptive prompt injection attacks on AI agents
Researchers have developed CoRL, a novel framework for defending against and simulating adaptive indirect prompt injection (IPI) attacks on tool-augmented language agents. IPI attacks hide adversarial instructions withi…
-
KV cache flaw breaks language agent rollback consistency
A new research paper from arXiv details a critical vulnerability in language agents related to KV cache retention, which can lead to rollback inconsistencies. This means that even when an application believes it has dis…
-
New benchmark reveals frontier AI models fail catastrophically on misinformation
Researchers have developed the Synthetic Web Benchmark, a novel environment designed to test language agents' susceptibility to misinformation. This benchmark features procedurally generated hyperlinked articles with gr…
-
New framework analyzes how AI agents learn from game experience
A new research paper introduces a framework called experience-sensitive game learning to analyze how gameplay experience influences the decision-making behavior of both humans and language agents. The study found that h…
-
New framework ECHO enhances language agent adaptivity with turn-level credit
Researchers have introduced Epistemic Decision Processes (EDPs), a new framework for multi-turn language agents that explicitly models information-seeking behavior. This approach aims to improve agent adaptivity by focu…
-
New research identifies task insensitivity as a key weakness in language agents
Researchers have identified "task insensitivity" as a key reason for the weak out-of-distribution generalization in large language models acting as agents. This phenomenon occurs when models apply learned patterns to ne…
-
Language Agents Learn to Ask for Help More Effectively
A new research paper introduces ACTION-RATING, a method to integrate clarification-seeking directly into the action space of hierarchical language agents. This formulation allows agents to compete between acting and ask…
-
New framework co-trains policy and world modeling for language agents
Researchers have developed a new framework called PaW for training language agents. This method co-trains policy and world modeling components simultaneously during reinforcement learning. PaW leverages existing RL data…
-
Study reveals model-generated agent skills have mixed utility
Researchers have conducted a systematic study on the lifecycle of model-generated agent skills, from experience generation to skill consumption. Their findings indicate that while these skills generally improve agent pe…