KARMA updates advertising-agent knowledge without changing model weights
Original title: KARMA: Training-Free Knowledge-Atomic Harness Evolution for Industrial Advertising Agents
Authors report average hold-out score rising from 30.19 to 39.15 across 18 domain–model settings.
KARMA is a training-free framework that edits six typed knowledge modules for industrial advertising agents while model weights and harness code stay frozen. A trace verifier maps errors to modules, a critic compares successful and failed trajectories, and an evolver applies atomic edits kept only if a validation gate sees enough gain. On samples from about half a million operation logs across six e-commerce domains and three unnamed LLM families, authors report hold-out scores rising from 30.19 to 39.15, with cross-domain transfer averaging 8.54 points.
Unconfirmed: Score gains are author-reported; the three LLM families are not named.
- Weights and executable harness code remain frozen during evolution.
- Edits are kept only when measured gain meets a validation threshold.
- Evaluation covers 18 domain–model settings and six e-commerce domains.
- Removing successful or failed trajectories more than halves the reported gain.
- Cross-domain transfer averages 8.54 hold-out points without further adaptation.
- AI models
- KARMA
From the original article
1 min read at arxiv.orgIndustrial advertising agents must continually adapt their decision knowledge to changing product offerings, traffic patterns, and operating policies. We present KARMA (Knowledge-Atomic, Regression-Gated Modular Adaptation), a training-free framework that evolves six typed knowledge modules while keeping model weights and executable harness code frozen.
By Xinhua Miao · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.