REVEAL: Reverse Engineering Agentic Latents from Frontier-Model Traces
Inspired by protocol reverse engineering, we introduce $Agentic Latents$, comprising anchored states and dynamic profiles of feedback--response cycles elicited by replay in the target model.
ProofPaper ↗
Key points
- Frontier LLM agents perform complex interactive tasks, but their inference costs motivate transferring their experience to smaller open-source agents.
- We study how textual frontier-model trajectories can provide reusable local controls for a frozen agent.
- Our training-free framework, $REVEAL$ ($REV$erse $E$ngineering $A$gentic $L$atents), uses these profiles and trajectory outcomes to match and select local events from successful and failed rollouts of the same task and source model.
- Representation analyses and behavioral case studies support reusing frontier-model experience as local activation control.
Sources (1)
- [1]REVEAL: Reverse Engineering Agentic Latents from Frontier-Model TracesarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 12, 07:14 AM
Inspired by protocol reverse engineering, we introduce $Agentic Latents$, comprising anchored states and dynamic profiles of feedback--response cycles elicited by replay in the target model.
Frontier LLM agents perform complex interactive tasks, but their inference costs motivate transferring their experience to smaller open-source agents.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 12, 2026Avoiding What Your Teacher Already Knows: Entity Familiarity for Deep-Search Distillation
- Oct 11, 2026A Sovereign, Indic-First, Reasoning Model
- Oct 10, 2026VC-OPD: Visual Counterfactual On-Policy Distillation for Grounded Vision-Language Reasoning
- Oct 9, 2026X-AVDD: Cross-Attentive Audio-Visual Dataset Distillation
- Oct 9, 2026Knowing Whether, Not Just Where: Anchor-Probing Distillation for Open-Set Remote Sensing Visual Grounding
- Oct 9, 2026FPD-Drive: Future Policy Distillation for End-to-End Autonomous Driving