Themis introduces the largest open code preference dataset with over 350k pairs and trains multilingual reward models from 600M to 32B parameters that support flexible multi-criteria scoring, with experiments showing scaling trends and cross-lingual transfer.
Helping or herding? reward model ensembles mitigate but do not eliminate reward hacking,
15 Pith papers cite this work. Polarity classification is still indexing.
15
Pith papers citing it
citation-role summary
background 1
method 1
citation-polarity summary
representative citing papers
TOMPA performs black-box adversarial optimization in token space to discover non-linguistic patterns that nearly double the reward scores of GPT-5 answers on Skywork-Reward-V2 while producing gibberish text.
OPPO is an evidence-aware preference optimization objective that contrasts faithful responses under varying visual evidence strengths to reduce hallucinations in MLLMs.
HARVE removes the component of the reward-head vector aligned with a multi-directional hacking subspace from residual streams using a small set of contrastive examples, improving robustness on RewardHackBench across eight models without fine-tuning while preserving general capability.
Introduces MOOD benchmark for OOD LLM alignment failures and shows guard models plus Mahalanobis and perplexity OOD detectors improve recall from 39% to 45% with positive scaling.
AdaScope adaptively selects optimal RL intervention points during diffusion denoising by monitoring structural and semantic changes, delivering 66% higher performance at 59% lower cost than full-trajectory RL baselines.
Response times modeled as drift-diffusion processes enable consistent estimation of population-average preferences from heterogeneous anonymous binary choices.
The power distribution is the target of power sampling, the closed-form solution to self-reward KL-regularized RL, and the basis for power self-distillation that matches sampling performance at lower cost.
Current video models succeed on basic understanding but achieve under 25% success on logically grounded generation and near 0% on interactive generation, exposing gaps in multimodal reasoning.
FUSE ensembles verifiers unsupervisedly by controlling their conditional dependencies to improve spectral ensembling algorithms, matching or exceeding semi-supervised baselines on benchmarks including GPQA Diamond and Humanity's Last Exam.
A factored causal representation learning method improves robustness of reward models in RLHF by isolating causal factors from biases like length and sycophancy using adversarial gradient reversal.
PROF curates RL training data via PRM-ORM consistency to improve both final-answer accuracy and intermediate reasoning quality while reducing reliance on strong process reward models.
CoLD mitigates length bias in process reward models for mathematical reasoning via counterfactual guidance, length penalties, bias estimation, and joint training, improving step selection accuracy and conciseness on MATH500 and GSM-Plus while boosting downstream RL performance.
Individual adaptive delegation-verification strategies aggregate via three extrapolation principles into sociotechnical lock-in modeled as a prisoner's dilemma that degrades epistemic standards unless mitigated by communicative standards and institutional norms.
citing papers explorer
No citing papers match the current filters.