First systematic study of chat-to-agent transfer of additive activation steering
Researchers present the first systematic study of how additive activation steering—calibrated in single-turn chat—transfers to tool-using ReAct agents. Using a matched-information design, they compare behavioral and representational effects of steering vectors in plain chat versus ReAct episodes, controlling for KV-cache contamination.
Entities
Related
LLM/VLA models enable prompt-driven exploration in RL
A new research paper proposes using large language models (LLMs) and vision-language-action (VLA) models to drive exploration in reinforcement learning by modifying natural language prompts, which induce global behavioral changes beyond standard action noise. This approach could help policies escape weak local optima more effectively.
KV-PRM: Efficient Process Reward Modeling via KV-Cache Transfer for Multi-Agent Test-Time Scaling
A new paper introduces KV-PRM, a process reward model that reuses KV-cache from the base model to avoid re-encoding entire trajectories, reducing scoring cost from quadratic to linear in sequence length. This enables efficient test-time scaling for long-context multi-agent systems.
Researchers propose memory agent to combat behavioral state decay in long-horizon tasks
A new arXiv paper identifies 'behavioral state decay' as a failure mode where decision-relevant information gets buried in long trajectories. The authors propose a separate memory agent that actively updates a structured memory bank alongside an unmodified action agent, rather than relying on passive retrieval.
GATS framework eliminates LLM calls during agent planning inference
Researchers propose GATS (Graph-Augmented Tree Search), a planning framework that uses a layered world model and UCB1-based tree search to avoid LLM inference during planning, reducing computational cost and stochasticity. The approach outperforms LATS and ReAct on multi-step planning tasks.
CLAP paper proposes direct VLM-to-VLA adaptation to isolate semantic transfer
A new arXiv paper introduces CLAP, a method that converts pretrained vision-language models (VLMs) into vision-language-action models (VLAs) with minimal architectural changes, avoiding the distribution mismatch caused by predicting actions as bare numeric tokens. The approach aims to clarify how VLM capabilities transfer across model scales for robot control.