arXiv subjects

Fangzheng Wu

Publications and source records attributed to Fangzheng Wu.

7 recordsLinked to original sources

arXiv2026

Risk-Conditioned Fine-Tuning of Large Language Models

Large Language Models (LLMs) are increasingly deployed in settings where rare but severe harmful generations can have significant consequences. Existing Risk-Averse RLHF addresses this issue by optimizing Conditional Value-at-Risk (CVaR), but it trains policies for fixed risk levels and therefore cannot adjust the desired degree of risk aversion at inference time. In this paper, we propose risk-conditioned RLHF, a framework that trains a single policy that provides a continuous risk-control interface, enabling users to select different degrees of risk aversion without retraining or deploying multiple risk-specific models. Experiments across multiple benchmarks demonstrate that a single risk-conditioned policy can adapt to different risk levels at inference time, enabling more flexible and risk-aware LLM deployment.

cs.LG↗

arXiv2026

Adaptive Multilevel Twisted Sequential Monte Carlo for Rare Events Estimation in Language Models

Rare unsafe behaviors in large language models can remain practically significant even when their probability is extremely small, particularly at deployment scales involving millions or billions of interactions. Twisted Sequential Monte Carlo (SMC) provides a principled framework for rare-event probability estimation by learning twist functions that guide generation toward a target event. However, the standard twist learning framework relies on positive samples from the rare-event target distribution, which may be nearly absent before an informative twist has been learned, resulting in unreliable rare-event estimation. We propose Adaptive Multilevel Twisted SMC, which learns the rare-event twist through a sequence of progressively rarer intermediate events. At each level, the learned twist provides more informative positive examples for learning the next twist, ultimately leading to a more accurate final twist for the target rare event. Experiments across diverse tasks and model scales show that the proposed method produces more accurate rare-event probability estimates. By enabling more reliable discovery of hard-to-observe unsafe behaviors, our method provides a practical tool for strengthening the evaluation and safety alignment of deployed language models.

cs.LG↗

arXiv2026

Robust General Utility for Reinforcement Learning

Reinforcement learning (RL) with general utility extends classic RL by optimizing an arbitrary utility functional of the policy-induced occupancy measure, thereby enabling a broader range of applications. However, previous work on general utility RL typically assumes the evaluation utility is fixed and correctly specified. In practice, the utility used at deployment can deviate from the training one, creating a robustness gap that prior work does not address. Motivated by this, we propose robust general-utility RL, a minimax learning framework that trains policies against utility misspecification within a prescribed uncertainty set. Our framework strictly generalizes standard general-utility RL while also providing a unified view of many existing RL frameworks, including reward-robust RL and constrained RL, through appropriate choices of the utility uncertainty set. We further develop provably convergent stochastic algorithms for two regimes. For concave utilities, we develop a projected stochastic gradient descent-ascent method and establish stationarity guarantees. For the more challenging nonconcave regime, we propose a stochastic prox-extragradient algorithm that mitigates ill-posed behavior induced by nonconcavity, with convergence guarantees to approximate first-order stationarity. Experiments on LLM safety alignment and exploration maximization tasks further corroborate the convergence behavior consistent with our theory.

cs.LG↗

arXiv2026

Attention Sinks in Diffusion Transformers: A Causal Analysis

Attention sinks -- tokens that receive disproportionate attention mass -- are assumed to be functionally important in autoregressive language models, but their role in diffusion transformers remains unclear. We present a causal analysis in text-to-image diffusion, dynamically identifying dominant attention recipients per timestep and suppressing them via paired, training-free interventions on the score and value paths. Across 553 GenEval prompts on Stable Diffusion~3 (with SDXL corroboration), removing these sinks does not degrade text-image alignment (CLIP-T) or preference proxies (ImageReward, HPS-v2) at $k{=}1$; only under stronger interventions ($k\!\geq\!10$) does HPS-v2 exhibit a metric-dependent boundary, while CLIP-T remains robust throughout. The perceptual shifts induced by suppression are nonetheless \emph{sink-specific} -- $\sim\!6\times$ larger than equal-budget random masking -- revealing an empirical dissociation between trajectory-level perturbation and \emph{semantic alignment} in diffusion transformers. \footnote{Code available at https://github.com/wfz666/ICML26-attention-sink.}

cs.CV↗

arXiv2026

SteeringDiffusion: A Bottlenecked Activation Control Interface for Diffusion Models

We introduce SteeringDiffusion, a bottlenecked activation-level control interface for diffusion models that exposes a smooth, monotonic, and runtime-adjustable control surface over the content--style trade-off. Our method keeps the U-Net backbone frozen and learns a small, prompt-conditioned latent code projected to FiLM/AdaGN-style modulation parameters. A zero-initialized design guarantees exact equivalence to the base model at zero scale, while timestep-aware gating restricts modulation to later denoising stages. A single scalar at inference continuously traverses the control surface without retraining. Across experiments on Stable Diffusion~1.5 and SDXL covering multiple artistic styles, we show that SteeringDiffusion produces smooth and monotonic content--style trade-offs. Under matched parameter budgets, it outperforms LoRA in controllability and stability, while ControlNet and rank-1 adapters do not expose a comparable control surface. We further introduce an inversion-stability diagnostic based on DDIM inversion, used as a post-hoc trajectory probe, which reveals strong correlations with intervention magnitude. These results position \emph{Steering Bottlenecked Explicit Control (S-BEC)} as a practical, general-purpose control interface for frozen diffusion backbones.

cs.CV↗

arXiv2026

CPT: Controllable and Editable Design Variations with Language Models

Designing visually diverse and high-quality designs remains a manual, time-consuming process, limiting scalability and personalization in creative workflows. We present a system for generating editable design variations using a decoder-only language model, the Creative Pre-trained Transformer (CPT), trained to predict visual style attributes in design templates. At the core of our approach is a new representation called Creative Markup Language (CML), a compact, machine-learning-friendly format that captures canvas-level structure, page layout, and element-level details (text, images, and vector graphics), including both content and style. We fine-tune CPT on a large corpus of design templates authored by professional designers, enabling it to learn meaningful, context-aware predictions for attributes such as color schemes and font choices. The model produces semantically structured and stylistically coherent outputs, preserving internal consistency across elements. Unlike generative image models, our system yields fully editable design documents rather than pixel-only images, allowing users to iterate and personalize within a design editor. In experiments, our approach generates contextual color and font variations for existing templates and shows promise in adjusting layouts while maintaining design principles.

cs.LG↗

arXiv2026

Model-Centric Diagnostics: A Framework for Internal State Readouts

We present a model-centric diagnostic framework that treats training state as a latent variable and unifies a family of internal readouts -- head-gradient norms, confidence, entropy, margin, and related signals -- as anchor-relative projections of that state. A preliminary version of this work introduced a head-gradient probe for checkpoint selection. In this version, we focus on the unifying perspective and structural diagnostics; full algorithmic details, theoretical analysis, and experimental validation will appear in a forthcoming paper. We outline the conceptual scaffold: any prediction head induces a local loss landscape whose geometry (gradient magnitude, curvature, sharpness) reflects how well the upstream features are aligned with the task. Different readout choices -- gradient norms, softmax entropy, predictive margin -- correspond to different projections of this geometry, each with complementary strengths. The framework suggests that checkpoint selection, early stopping, and lightweight architecture pre-screening can all be viewed as querying the same underlying state through different lenses. Illustrative experiments on ImageNet classification and COCO detection/segmentation hint at the practical potential; rigorous benchmarks and ablations are deferred to the full paper.

cs.CV↗