New Jersey Institute of Technology

3 related articles

Subscribe to updates

arXiv Oct 5, 2026
MLCommons releases Jailbreak Benchmark v1.0: unsafe-response rate across eight open-weight models rises from 11.08% to 18.65% under jailbreak conditions, with an average Resilience Gap of 7.57%
MLCommons Jailbreak Benchmark v1.0 introduces an end-to-end methodology that evaluates eight open-weight large language models with paired baseline and adversarial conditions using 264 seed prompts spanning eleven hazard categories and representative attacks drawn from the MLCommons Jailbreak Taxonomy, judging responses with the AILuminate Assessment Standard v1.4 and measuring robustness through the Resilience Gap; the unsafe-response rate rose from 11.08% at baseline to 18.65% under jailbreak conditions, yielding an average Resilience Gap of 7.57%, with accessible systems showing a larger mean gap and attack effectiveness varying substantially across attack categories and hazards.
MLCommons Jailbreak Benchmark v1.0 introduces an end-to-end methodology that evaluates eight open-weight large language models with paired baseline and adversarial conditions using 264 seed prompts spanning eleven hazard categories and representative attacks drawn from the MLCommons Jailbreak Taxonomy, judging responses with the AILuminate Assessment Standard v1.4 and measuring robustness through the Resilience Gap; the unsafe-response rate rose from 11.08% at baseline to 18.65% under jailbreak conditions, yielding an average Resilience Gap of 7.57%, with accessible systems showing a larger mean gap and attack effectiveness varying substantially across attack categories and hazards.
Qualcomm (United States)Evolve Restorative Center New York University University of Warwick Mohamed I University Clarkson University University of Illinois Urbana-Champaign New Jersey Institute of Technology
Carsten Maple（First author）
MLCommons Jailbreak Benchmark v1.0 jailbreak robustness evaluation
arXiv May 23, 2026
HyperGuide guides LLM reasoning along a single trajectory via virtual tokens in hyperbolic space, matching or exceeding search and verifier baselines on competition math and code generation with far fewer tokens
HyperGuide treats reasoning as movement through a tree of states embedded in hyperbolic space (the Poincaré ball), first training a state encoder and then a guidance head that predicts the direction of a minimum-cost transition from the current state, and inserts that direction into the decoder as a virtual token after each reasoning step so inference follows a single autoregressive trajectory without expanding or reranking candidates; on competition mathematics and code generation experiments it matches or exceeds the accuracy of search- and verifier-based baselines while generating substantially fewer tokens than those baselines and about as many as few-shot prompting.
HyperGuide treats reasoning as movement through a tree of states embedded in hyperbolic space (the Poincaré ball), first training a state encoder and then a guidance head that predicts the direction of a minimum-cost transition from the current state, and inserts that direction into the decoder as a virtual token after each reasoning step so inference follows a single autoregressive trajectory without expanding or reranking candidates; on competition mathematics and code generation experiments it matches or exceeds the accuracy of search- and verifier-based baselines while generating substantially fewer tokens than those baselines and about as many as few-shot prompting.
Yale University New Jersey Institute of Technology Stony Brook University
Yuyu Liu（First author）
HyperGuide Multi-step reasoning Large language models Code Generation
arXiv Oct 2, 2026
BiasReducer edits only the reward head, lifting five reward models by 8.3/18.0/6.9 points on average across three bias benchmarks
The work proposes BiasReducer, a lightweight framework that edits only the linear reward head without retraining the reward model: an SAE-style encoder with semantic supervision learns internal representations for predefined attributes such as length, confidence, and sycophancy; the framework then determines the direction and amount by which to adjust the reward head for each attribute and stores these as an edit bank; for a new dataset it ranks attributes by their influence on reward scores and applies the matching edits. Across five reward models, BiasReducer-M improves RM-Bench-Hard, Arena-StyleConflict, and JudgeBiasBench by 8.3, 18.0, and 6.
The work proposes BiasReducer, a lightweight framework that edits only the linear reward head without retraining the reward model: an SAE-style encoder with semantic supervision learns internal representations for predefined attributes such as length, confidence, and sycophancy; the framework then determines the direction and amount by which to adjust the reward head for each attribute and stores these as an edit bank; for a new dataset it ranks attributes by their influence on reward scores and applies the matching edits. Across five reward models, BiasReducer-M improves RM-Bench-Hard, Arena-StyleConflict, and JudgeBiasBench by 8.3, 18.0, and 6.
New Jersey Institute of Technology Carnegie Mellon University
Shuang Liu（First author）
BiasReducer Reward-Model Bias Mitigation Sparse Autoencoder Reward Hacking Semantic Supervision

You’re all caught up

MLCommons releases Jailbreak Benchmark v1.0: unsafe-response rate across eight open-weight models rises from 11.08% to 18.65% under jailbreak conditions, with an average Resilience Gap of 7.57%

HyperGuide guides LLM reasoning along a single trajectory via virtual tokens in hyperbolic space, matching or exceeding search and verifier baselines on competition math and code generation with far fewer tokens

BiasReducer edits only the reward head, lifting five reward models by 8.3/18.0/6.9 points on average across three bias benchmarks