Manjiang Yu

Manjiang Yu, Hongji Li, Junwei Chen et al.

Representation intervention has emerged as a promising paradigm for aligning large language models toward desired behaviors without modifying model weights. Existing methods typically apply a fixed intervention uniformly across all inputs. However, we find that the appropriate intervention direction and strength vary substantially across samples, and such indiscriminate intervention leads to degradation of general capabilities on benign inputs. To address these challenges, we propose Multi-Adapter Representation Interventions via Energy Calibration (MARI). Specifically, we introduce a competitive multi-adapter mechanism in which specialized experts capture non-linear correction patterns and adaptively determine the appropriate intervention direction and strength for different samples. Furthermore, we design an energy-based gating module that leverages internal propagation dynamics to distinguish inputs that are applicable for intervention. Extensive experiments across diverse model families and parameter scales demonstrate that MARI achieves state-of-the-art alignment performance. Our method significantly improves performance on TruthfulQA, BBQ, and safety benchmarks, while maintaining and even improving general capabilities on tasks such as MMLU and ARC. Our code is available at https://github.com/V1centNevwake/MARI.

Manjiang Yu, Hongji Li, Priyanka Singh et al.

Reliable behavior control is central to deploying large language models (LLMs) on the web. Activation steering offers a tuning-free route to align attributes (e.g., truthfulness) that ensure trustworthy generation. Prevailing approaches rely on coarse heuristics and lack a principled account of where to steer and how strongly to intervene. To this end, we propose Position-wise Injection with eXact Estimated Levels (PIXEL), a position-wise activation steering framework that, in contrast to prior work, learns a property-aligned subspace from dual views (tail-averaged and end-token) and selects intervention strength via a constrained geometric objective with a closed-form solution, thereby adapting to token-level sensitivity without global hyperparameter tuning. PIXEL further performs sample-level orthogonal residual calibration to refine the global attribute direction and employs a lightweight position-scanning routine to identify receptive injection sites. We additionally provide representation-level guarantees for the minimal-intervention rule, supporting reliable alignment. Across diverse models and evaluation paradigms, PIXEL consistently improves attribute alignment while preserving model general capabilities, offering a practical and principled method for LLMs' controllable generation. Our code is available at https://github.com/V1centNevwake/PIXEL-Adaptive-Steering

Hongji Li, Junchi yao, Manjiang Yu et al.

Machine unlearning aims to erase requested data from trained models without full retraining. For Reasoning Multimodal Large Language Models (RMLLMs), this is uniquely challenging: intermediate chain-of-thought steps can still leak sensitive information even when final answers are forgotten, and overly aggressive interventions easily damage general reasoning ability. Yet no benchmark jointly evaluates how well unlearning methods suppress reasoning-level leakage while preserving reasoning competence. We address this gap with RMLLMU-Bench, the first benchmark for RMLLM unlearning that extends standard forgetting metrics with dedicated measures of reasoning leakage and reasoning retention. A systematic evaluation on RMLLMU-Bench reveals that existing unlearning methods for MLLMs and Large (Language) Reasoning Models (LRMs) either leave substantial leakage in the reasoning process or severely degrade reasoning performance. To address these gaps, we propose R-MUSE (Reasoning-preserving MLLM Unlearning via Subspace guidance and Adaptive Steering), a training-free and inference-time intervention framework that steers internal representations to forget both answers and reasoning traces while explicitly preserving general reasoning. Experiments on RMLLMU-Bench demonstrate that R-MUSE achieves a substantially better balance between effective forgetting and reasoning retention.

Manjiang Yu, Priyanka Singh, Xue Li et al.

Large language models (LLMs) frequently memorize sensitive or personal information, raising significant privacy concerns. Existing variants of differential privacy stochastic gradient descent (DPSGD) inject uniform noise into every gradient step, significantly extending training time and reducing model accuracy. We propose that concentrating noise primarily on gradients associated with sensitive tokens can substantially decrease DP training time, strengthen the protection of sensitive information, and simultaneously preserve the model's performance on non-sensitive data. We operationalize this insight through Adaptive Token-Weighted Differential Privacy (ATDP), a modification of vanilla DP-SGD that adaptively assigns different gradient weights to sensitive and non-sensitive tokens. By employing a larger noise scale at the early stage of training, ATDP rapidly disrupts memorization of sensitive content. As a result, ATDP only requires a few additional epochs of lightweight post-processing following standard fine-tuning, injecting targeted noise primarily on parameters corresponding to sensitive tokens, thus minimally affecting the model's general capabilities. ATDP can be seamlessly integrated into any existing DP-based fine-tuning pipeline or directly applied to non-private models as a fast privacy-enhancing measure. Additionally, combined with an initial redacted fine-tuning phase, ATDP forms a streamlined DP pipeline that achieves comparable canary protection to state-of-the-art DP-SGD methods, significantly reduces the computational overhead of DP fine-tuning, shortening training time by approximately 90 percent, while achieving comparable or superior privacy protection and minimal accuracy degradation.

Manjiang Yu, Xue Li

In time-critical decisions, human decision-makers can interact with AI-enabled situation-aware software to evaluate many imminent and possible scenarios, retrieve billions of facts, and estimate different outcomes based on trillions of parameters in a fraction of a second. In high-order reasoning, "what-if" questions can be used to challenge the assumptions or pre-conditions of the reasoning, "why-not" questions can be used to challenge on the method applied in the reasoning, "so-what" questions can be used to challenge the purpose of the decision, and "how-about" questions can be used to challenge the applicability of the method. When above high-order reasoning questions are applied to assist human decision-making, it can help humans to make time-critical decisions and avoid false-negative or false-positive types of errors. In this paper, we present a model of high-order reasoning to offer recommendations in evidence-based medicine in a time-critical fashion for the applications in ICU. The Large Language Model (LLM) is used in our system. The experiments demonstrated the LLM exhibited optimal performance in the "What-if" scenario, achieving a similarity of 88.52% with the treatment plans of human doctors. In the "Why-not" scenario, the best-performing model tended to opt for alternative treatment plans in 70% of cases for patients who died after being discharged from the ICU. In the "So-what" scenario, the optimal model provided a detailed analysis of the motivation and significance of treatment plans for ICU patients, with its reasoning achieving a similarity of 55.6% with actual diagnostic information. In the "How-about" scenario, the top-performing LLM demonstrated a content similarity of 66.5% in designing treatment plans transferring for similar diseases. Meanwhile, LLMs managed to predict the life status of patients after their discharge from the ICU with an accuracy of 70%.