arXiv AI

Revisiting scaling laws for reward optimization

The paper presents a new scaling law for reward optimization in AI alignment, showing that performance scales as Θ(√min{log(M), K}), where M is the number of preference comparisons used to train a proxy reward model and K is the KL‑divergence budget relative to a reference policy. The authors derive this law using an information‑theoretic model, prove its tightness, and validate it with extensive experiments involving a 70B gold reward model and smaller proxy models (0.6B–4B). The empirical results demonstrate a strong fit (R² 97–99 %) across different model sizes, noise levels, and optimization methods, suggesting that reward optimization behaves like a simple selection task over IID Gaussian variables with noisy feedback.

arXiv Machine Learning
Jul 30

Post-Training at the Edge of Detectability: A Game-Theoretic Approach to Fine-Tuning

arXiv:2607. 26358v1 Announce Type: new Abstract: Reinforcement learning (RL) fine-tuning is widely used in language model training to improve model performance on a target task while limiting drift from a reference policy.

By Keegan Harris, Brian W. Lee, Ian Waudby-Smith, Philip Amortila, Nika Haghtalab, Michael I. Jordan
Hugging Face Trending Papers
Jul 29

Post-Training at the Edge of Detectability: A Game-Theoretic Approach to Fine-Tuning

Reinforcement learning (RL) fine-tuning is widely used in language model training to improve model performance on a target task while limiting drift from a reference policy. A standard way to balance this trade-off is via a KL-regularized RL objective, although this formulation does not by itself provide a principled way to set the regularization coefficient.

arXiv Machine Learning
Aug 10

Multiscale Reward Hedging from Correct Demonstrations

arXiv:2608. 06825v1 Announce Type: new Abstract: Learning from correct demonstrations is harder than supervised learning when many answers are correct: after predicting, the learner sees one valid answer but not whether its own answer was valid, nor any reward.

By Pahan Dewasurendra
arXiv AI
Oct 1

On the Complexity of Preference-Based Bandits

The paper investigates preference-based bandits where a learner selects pairs of arms and receives binary preference feedback modeled by Bradley–Terry. It introduces the locally sensitive eluder dimension, a new complexity measure for logistic preference feedback, and proposes the GINOP algorithm that uses log-loss confidence sets to balance optimism and exploration. The authors prove a first-order regret bound showing that learning with preference feedback can be as statistically efficient as learning from direct rewards, and they validate their theory with empirical experiments.

By Ahmed Ben Yahmed (CREST, ENSAE Paris, FAIRPLAY), Marc Abeille (FAIRPLAY), Cl\'ement Calauz\`enes (FAIRPLAY)
arXiv Statistics ML
Nov 9

High-Dimensional Off-Policy Evaluation without Sample Splitting using an Onsager Correction

The paper introduces an Onsager‑corrected, target‑policy‑weighted criterion that estimates the out‑of‑sample risk of a reward model trained on the full dataset using only its training data, thereby eliminating the need for sample splitting. This approach allows early stopping along a full‑data gradient‑descent trajectory and provides uniform consistency guarantees for neural networks under Gaussian contexts and single‑index rewards. The correction also yields a doubly robust policy‑value estimator that remains consistent even when reward‑model prediction error does not vanish, and experiments demonstrate reduced risk‑estimation error and improved reward‑model selection compared to held‑out baselines. whyItMatters":"By removing the need for sample splitting, the method enables more efficient use of data and improves the reliability of off‑policy evaluation and reward‑model selection in high‑dimensional settings."

By Saksham Jain, Guillaume Braun, Masaaki Imaizumi
arXiv Machine Learning
Aug 4

Meritocratic Fairness via $K$-Shapley Values in Budgeted Combinatorial Bandits with Full-Bandit Feedback

arXiv:2605. 00762v2 Announce Type: replace Abstract: We study meritocratic fairness in budgeted combinatorial multi-armed bandits with full-bandit feedback, where a learner selects at most $K$ arms per time step and observes only the noisy aggregate reward of the selected set.

By Shradha Sharma, Shweta Jain, Swapnil Dhamal