02 / 03
How can models learn from verifiers, rewards, and other models?
What the training signal should be — which reward, and how much each teacher's feedback should count.
Preprint 2026
Label routing decides which teacher supervises a prompt, but not how strongly that teacher's feedback counts. Normalizing the feedback scale per domain adds 1.2–3.1 points over label routing, across 2B, 4B and 9B students.
Preprint 2026
Rank-GRPO trains an LLM recommender on one human answer at a time, so 85% of its advantage variance records which answer was drawn. Rewarding each list once against all answers reaches the baseline's recall in 25–53% of the RL steps, across seven backbones and three datasets.
Preprint 2026
TMLR 2026
Trains 0.5B–14B models to write Dafny specifications with minimal human annotation — no natural-language chain-of-thought, auto-generated training data, rewards from the Dafny verifier; on 300 out-of-domain DafnyComp programs, the 14B RL model passes the verifier on 14.0% at pass@1, against 8.3% after SFT alone and 2.7% for Claude, the model that generated its training data.
npj Wireless Technology 2026
Small language models can learn physical reasoning from rewards. Fine-tuned with GRPO and physics-informed rewards, a 0.5B model reconstructs RSSI along UAV flight paths at 2.85 dB mean absolute error, 49% below its untrained baseline and on par with GPT-4o, transfers zero-shot to spatial interpolation, and runs at 3 ms latency.
EMNLP 2026 Industry
Turns thousands of review tuples into a ranked list of product insights while every tuple stays attached to the insight it supports: membership is computed outside the LLM, so coverage is 100% with no duplicates. An anti-copy reinforcement objective lifts the upstream extractor from 59.3% to 93.1% production quality.