MATE
A deterministic post-retrieval memory adaptation framework for embodied agents. It introduces a dedicated adaptation stage between memory retrieval and embodied execution, transforming historical trajectories into executable representations through action normalization and condition-action-effect decomposition, without requiring additional LLM inference.
Timeline (last 90 days)
MATE's current validation is primarily focused on the ALFWorld benchmark, and its generalization capability in more complex and open-ended real-world environments remains to be further examined.
MATE argues that memory adaptation should exist as an independent stage between retrieval and embodied execution.
MATE's processing pipeline includes removing outdated control context, extracting condition-action-effect transitions, applying verified action normalization, selecting task-relevant representations, and serializing output within a fixed budget.
MATE's entire processing pipeline does not require additional LLM inference.
The research team validated MATE's effectiveness on 134 ALFWorld tasks.
MATE paired with the Qwen2.5-14B model achieves a task success rate of 81.3%.
The number of tokens consumed by MATE is approximately one-tenth that of the original trajectory.
Ablation analysis confirms that verified action normalization is the primary mechanism by which MATE restores the usability of retrieved experience.
Trail of Bits maintains the MATE framework for binary analysis and the smart contract analysis tool Echidna.
All Facts (9)
MATE's current validation is primarily focused on the ALFWorld benchmark, and its generalization capability in more complex and open-ended real-world environments remains to be further examined.
50%UnverifiedMATE argues that memory adaptation should exist as an independent stage between retrieval and embodied execution.
50%UnverifiedMATE's processing pipeline includes removing outdated control context, extracting condition-action-effect transitions, applying verified action normalization, selecting task-relevant representations, and serializing output within a fixed budget.
50%UnverifiedMATE's entire processing pipeline does not require additional LLM inference.
50%UnverifiedThe research team validated MATE's effectiveness on 134 ALFWorld tasks.
50%UnverifiedMATE paired with the Qwen2.5-14B model achieves a task success rate of 81.3%.
50%UnverifiedThe number of tokens consumed by MATE is approximately one-tenth that of the original trajectory.
50%UnverifiedAblation analysis confirms that verified action normalization is the primary mechanism by which MATE restores the usability of retrieved experience.
50%UnverifiedTrail of Bits maintains the MATE framework for binary analysis and the smart contract analysis tool Echidna.
50%