The authors built a proposer-solver loop in which a model generates an open-ended task and a weighted rubric.
Adversarial rubric repair improves self-training on open-ended expert tasks
RRI tests model-written grading rubrics against deliberately unhelpful answers, repairs exploitable criteria, and then uses the revised rewards for reinforcement learning.
Big Tech
Wei Dai · Qianchu Liu · Guanghui Qin · Xing Han · Tristan Naumann · Paul Pu Liang · +1 more
Massachusetts Institute of Technology · Microsoft Research · Johns Hopkins University
Research Digest··3 min read
Thread:Open-Ended RL Reward Design
Dai et al.
Why this paper
From Microsoft Research and 2 others · Part of Open-Ended RL Reward Design, now 12 papers
In one line
A reward hacking adversary supervises and repairs model-written rubrics, improving self-improvement on open-ended tasks.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ✓Compute or model size stated (params 9B)
- ✓Limitations stated by the authors
- ✓Reports numbers on named benchmarks (4 benchmarks)
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§