arXiv subjects

Kaixing Zhang

Publications and source records attributed to Kaixing Zhang.

3 recordsLinked to original sources

arXiv2026

UniSteer: Text-Guided Flow Matching in Activation Space for Versatile LLM Steering

Activation-based control steers large language models (LLMs) by intervening on their internal representations during inference, and has emerged as an effective paradigm for controlling behaviors such as persona and style. However, existing methods often rely on fixed steering directions or task-specific intervention modules, making them difficult to adapt to fine-grained concepts and compositional constraints. We propose UniSteer, a text-guided activation flow matching model that learns a conditional distribution over residual-stream activations from natural-language conditions. Instead of fitting a separate intervention for each target behavior, UniSteer learns a universal conditional velocity field in activation space. At inference time, UniSteer performs flow inversion by partially transporting a source activation toward a latent state and regenerating it under a target textual condition before injecting it back into the frozen LLM. The same conditional model supports activation-space classification by selecting the textual label with the lowest reconstruction energy. Experiments on three target LLMs show that UniSteer provides a unified interface across behavioral control, truthfulness steering, fine-grained concept steering, multi-constraint instruction following, and activation-space classification.

cs.CL↗

arXiv2026

Interpreting and Controlling LLM Reasoning through Integrated Policy Gradient

Large language models (LLMs) demonstrate strong reasoning abilities in solving complex real-world problems. Yet, the internal mechanisms driving these complex reasoning behaviors remain opaque. Existing interpretability approaches targeting reasoning either identify components (e.g., neurons) correlated with special textual patterns, or rely on human-annotated contrastive pairs to derive control vectors. Consequently, current methods struggle to precisely localize complex reasoning mechanisms or capture sequential influence from model internal workings to the reasoning outputs. In this paper, built on outcome-oriented and sequential-influence-aware principles, we focus on identifying components that have sequential contribution to reasoning behavior where outcomes are cumulated by long-range effects. We propose Integrated Policy Gradient (IPG), a novel framework that attributes reasoning behaviors to model's inner components by propagating compound outcome-based signals such as post reasoning accuracy backward through model inference trajectories. Empirical evaluations demonstrate that our approach achieves more precise localization and enables reliable modulation of reasoning behaviors (e.g., reasoning capability, reasoning strength) across diverse reasoning models.

cs.AI↗

arXiv2025

Algorithmic Complexity Attacks on All Learned Cardinality Estimators: A Data-centric Approach

Learned cardinality estimators show promise in query cardinality prediction, yet they universally exhibit fragility to training data drifts, posing risks for real-world deployment. This work is the first to theoretical investigate how minimal data-level drifts can maximally degrade the accuracy of learned estimators. We propose data-centric algorithmic complexity attacks against learned estimators in a black-box setting, proving that finding the optimal attack strategy is NP-Hard. To address this, we design a polynomial-time approximation algorithm with a $(1-κ)$ approximation ratio. Extensive experiments demonstrate our attack's effectiveness: on STATS-CEB and IMDB-JOB benchmarks, modifying just 0.8\% of training tuples increases the 90th percentile Qerror by three orders of magnitude and raises end-to-end processing time by up to 20$\times$. Our work not only reveals critical vulnerabilities in deployed learned estimators but also provides the first unified worst-case theoretical analysis of their fragility under data updates. Additionally, we identify two countermeasures to mitigate such black-box attacks, offering insights for developing robust learned database optimizers.

cs.DB↗