This analysis gives sufficient conditions under which suppression at supervised single-turn contexts yields a bound on multi-turn trajectory risk, and characterizes contraction relative to a base-policy risk budget evaluated on the trained policy's contexts.
Abstract
Safety-aligned large language models (LLMs) often refuse a harmful request but comply once the same goal is spread over several turns. Preference objectives score whole responses to single prompts, so their training loss alone cannot control risk on unseen histories. Our analysis gives sufficient conditions under which suppression at supervised single-turn contexts yields a bound on multi-turn trajectory risk. The bound accounts for coverage, transfer slack, and leakage, and characterizes contraction relative to a base-policy risk budget evaluated on the trained policy's contexts. TRACE (Trajectory Return Attribution and Contrastive Erasure) turns this principle into a token-level objective. On the safe response, each token is weighted by the discounted return of a refusal-attributable advantage. The advantage compares a frozen reference model with its refusal-ablated copy, allowing earlier response tokens to receive credit from later refusal-related evidence. At high-gap positions on rejected responses, TRACE combines the observed token with policy-selected alternatives in the erasure target. A gradient-norm penalty replaces the retain set. Across five open-weight models and seven multi-turn attacks, TRACE gives the lowest attack success rate (ASR) in all 35 model and attack pairs, while the model utility evaluated on MMLU and HellaSwag drop by at most 1\.23 points. Source code can be found in the supplemental material.
It is found that Llama 3 delivers comparable quality to leading language models such as GPT-4 on a plethora of tasks, and performs competitively with the state-of-the-art on image, video, and speech recognition tasks.
This paper investigates developing real-world perception for active agents, proposes Gibson Environment for this purpose, and showcases a set of perceptual tasks learned therein.
F. Xia, Amir Zamir, Zhi-Yang He et al.· 2018 IEEE/CVF Conference on...· 1,069 citations· ⚡112
GAOKAO-Bench is introduced, an intuitive benchmark that employs questions from the Chinese GAOKAO examination as test samples, including both subjective and objective questions that contribute a robust evaluation benchmark for future large language models and offers valuable insights into the advantages and limitations...
Xiaotian Zhang, Chun-yan Li, Yi Zong et al.· arXiv.org· 216 citations· ⚡17
An algorithm is proposed which leverages disentangled semantic factors to generate adversarial perturbation by altering controlled semantic attributes to fool the learner towards various "adversarial" targets.
Haonan Qiu, Chaowei Xiao, Lei Yang et al.· arXiv.org· 208 citations· ⚡17
This work proposes a HALO that directly maximizes the utility of generations instead of maximizing the log-likelihood of preferences, and suggests that the best loss depends on the inductive biases most appropriate for a given setting, an oft-overlooked consideration.
This work investigates the possibilities of using LLMs in a resume screening setting via a document retrieval framework that simulates job candidate selection and finds that the MTEs are biased, significantly favoring White-associated names in 85% of cases and female-associated names in only 11.1% of cases.