arXiv subjects

Bruce Changlong Xu

Publications and source records attributed to Bruce Changlong Xu.

8 recordsLinked to original sources

arXiv2026

Alignment Collapse Under KV Cache Quantization: Diagnosis and Mitigation

Key-value (KV) cache quantization is widely used to reduce Large Language Model (LLM) inference memory, yet existing evaluations solely focus on measuring perplexity and accuracy without assessing the safety impact. In this study, we explore alignment preservation under KV cache quantization. Across eleven instruction-tuned models (3.8B-72B) and five benchmarks (1,894 prompts), we find that low-bit quantization can silently destroy safety alignment: Mistral-7B loses 15.2% of its refusals at only 1.03x perplexity, and no universal safe bit-width exists, with sharp model-specific phase transitions invisible to standard metrics. We identify that the root cause is geometric: safety features occupy a low-dimensional activation subspace 10^2-10^3x more vulnerable to quantization noise than the full representation space perplexity averages over. Inspired by this observation, we propose Per-Channel Reduction (PCR), a diagnostic that classifies each model into one of three mechanistic failure modes: outlier-crushes-safety, where safety lives in non-outlier channels collaterally damaged by outlier-driven scale factors; outlier-as-safety, where safety overlaps outlier channels and finer granularity cannot rescue it; and multi-layer dilution, where safety is distributed across many layers and per-layer fixes fail. PCR predicts the correct mitigation direction on all nine primary models and one held-out model from an independent family using 20 calibration prompts. PCR generalizes across unseen prompts, models, and production quantizers, including KIVI with up to 97.2% recovery, succeeding where attention-based allocation methods fail. The resulting training-free protocol, requiring approximately 35 GPU-minutes, recovers up to 97% of lost alignment at minimal memory overhead, addressing vulnerabilities confirmed in production vLLM serving with FP8 KV cache on NVIDIA GPUs.

cs.LG↗

arXiv2026

A Controlled Audit of Pretraining Contamination in Public Medical Vision-Language Benchmarks

Medical vision-language models (VLMs) are evaluated on public benchmarks whose images and question-answer pairs have been freely downloadable for years, yet reported accuracy assumes these examples were absent from pretraining. We audit open VLMs on SLAKE-En, PathVQA, VQA-RAD, and an auxiliary public OmniMedVQA mirror using four detector families: image-side near-neighbour overlap against PMC-OA-beta, canonical-order exchangeability, cohort-relative Min-K%++ tail enrichment, and cross-model top-K overlap. We find measurable image-side source overlap on SLAKE-En: 19.8% of images are flagged under SigLIP-B-16 and 4.2% under SigLIP-SO400M, while out-of-domain controls produce 0/2000 flags. Manual adjudication shows same-modality, same-projection matches to different patients rather than verified pixel-level duplicates, so we interpret this as source or distributional overlap rather than confirmed per-image memorization. On the text side, Qwen2.5-VL on SLAKE-En shows a canonical-order exchangeability signal that survives ordering ablation and external non-medical baselines. On the OmniMedVQA mirror, exchangeability fires for five medical and general VLMs while BLIP-2 remains clean. In contrast, cohort-relative Min-K%++ tail enrichment and cross-model top-K overlap collapse under an external pre-domain baseline: BLIP-2 reproduces the apparent positive signals despite lacking plausible medical-VQA exposure. We conclude that these cohort-relative detectors are unreliable as standalone membership-inference signals on small medical-VLM cohorts.

cs.CV↗

arXiv2026

Typhoon: Towards an Effective Task-Specific Masking Strategy for Pre-trained Language Models

The choice of \emph{which} tokens to mask is a central, under-examined design decision in masked language modeling (MLM). Standard pretraining masks tokens uniformly at random, but several studies show that more informative masking targets can improve downstream performance. We study masking as a \emph{task-adaptive} component of the fine-tuning pipeline and introduce \textbf{Typhoon}, a masking strategy that uses the gradient of the task loss with respect to one-hot token inputs to estimate, online, how much each token type contributes to the objective. Typhoon maintains an exponential moving average of per-token-type saliency and calibrates these scores into a masking distribution whose expected masking rate matches a target budget, under a token-independence approximation. We formalize the method and evaluate it against random masking and whole-word masking on two GLUE tasks, MRPC and CoLA, across three BERT-family backbones (TinyBERT, DistilBERT, and BERT-base) and five random seeds per configuration ($90$ training runs in total). Our main finding is that, once seed variance is accounted for, no masking strategy is reliably better than the others on these tasks: on MRPC the gap between Typhoon and the best baseline stays within $0.004$ $F_1$, across all twelve Typhoon comparisons no paired test reaches significance, and every $95\%$ confidence interval contains zero. Typhoon's apparent advantage in single-run experiments does not survive this more careful evaluation. We read this as a cautionary, reproducibility-focused result -- gradient-based task-adaptive masking is competitive but not clearly better than resource-free random masking at this scale -- and we describe a clean modern reimplementation to support follow-up work.

cs.CL↗

arXiv2026

Training-Inference Kernel Contracts: Bounding Divergence in Post-Training and Deployment

A modern post-training pipeline often writes one symbol for its policy, pi_theta, while evaluating it through two different programs: a training kernel optimized for autograd and an inference kernel optimized for low-precision, fused, dynamically batched serving. In finite precision, these kernels can induce different distributions at identical weights, with the gap concentrated on slices that aggregate benchmarks under-represent. This paper proposes kernel contracts: a contract-first framework for specifying acceptable divergence between K_train and K_inf. A contract C = (N, S, R, O, Pi) combines numerical, statistical, runtime, and observability clauses with an escalation policy from violations to routing actions. We derive a chain of bounds from logit drift to total-variation distance to bounded reward drift, and specialize it to RL post-training, where per-token importance-ratio drift yields a bound on policy-gradient bias under explicit support and norm assumptions. We also describe a four-stage promotion pipeline, online routing loop, and minimal YAML DSL for contract artifacts. This is a framework and vocabulary paper; we do not report production-scale empirical validation.

cs.LG↗

arXiv2026

From Theory to Decision Rule: Calibrating the Noisy-Label Crossover for Vision-Language Model Weak Supervision Across Three Medical-Imaging Benchmarks

Classical noisy-label theory predicts that downstream performance under weak supervision is bounded above by the labeler's accuracy, implying a sharp crossover: once a gold-trained classifier matches the labeler, weak labels stop helping and start hurting. The prediction is theoretical; what is missing is a benchmark calibration that turns it into an instance-level statement for modern foundation-model labelers. We provide such a calibration for BiomedCLIP-generated weak labels on three medical-imaging benchmarks (PCAM, ISIC, NIH-CXR) and six downstream architectures spanning an 11x parameter range. The crossover predicted by theory appears at ng~100 on PCAM, 20-50 on ISIC, and 250-500 on NIH-CXR; weak labels above the crossover degrade AUC by up to -0.10. The location is architecture-invariant for four of five pretrained architectures, and a within-family DenseNet sweep (2.5x parameters, identical pretraining) supports the view that the labeler, not the student, is the dominant constraint. The calibration in turn produces a decision rule operable from 10-20 gold labels: compare gold-only AUC to VLM accuracy on the user's gold set. A structured-vs-random noise sign flip on NIH-CXR shows that the rate-only formulation of the bound is incomplete and identifies a concrete refinement (label-space projection) that future benchmarks can be designed to test.

cs.CV↗

arXiv2026

Activation Sensitivity as a Unifying Principle for Post-Training Quantization

Post-training quantization (PTQ) methods for large language models rely on heuristics that implicitly estimate which weight channels most strongly influence model behavior. Two dominant paradigms have emerged: activation-aware methods such as AWQ prioritize channels with large activation magnitudes, while second-order methods such as GPTQ allocate quantization error according to input covariance structure. Despite strong empirical performance, these approaches remain conceptually fragmented, and it is unclear what underlying quantity they are approximating. In this work, we present a unified theoretical framework for PTQ by formalizing activation sensitivity, defined as the expected impact of channel-wise perturbations on the loss. Using a first-order Taylor expansion, we show that sensitivity naturally arises as the squared norm of gradient-weighted activations, yielding a principled measure of channel importance that captures both activation magnitude and downstream error propagation. Within this framework, AWQ and GPTQ can be interpreted as complementary approximations that recover sensitivity under distinct simplifying assumptions. We analyze the design space of sensitivity metrics, connect gradient-based saliency, Fisher information, and Hessian-based criteria, and clarify their relationships to classical pruning methods such as Optimal Brain Damage and Optimal Brain Surgeon. Rather than proposing a new quantization algorithm, this work provides a conceptual foundation for understanding and comparing post-training quantization methods through the lens of sensitivity.

cs.LG↗

arXiv2021

Matching Markets

Matching markets are of particular interest in computer science and economics literature as they are often used to model real-world phenomena where we aim to equitably distribute a limited amount of resources to multiple agents and determine these distributions efficiently. Although it has been shown that finding market clearing prices for Fisher markets with indivisible goods is NP-hard, there exist polynomial-time algorithms able to compute these prices and allocations when the goods are divisible and the utility functions are linear. We provide a promising research direction toward the development of a market that simulates buyers' preferences that vary according to the bundles of goods allocated to other buyers. Our research aims to elucidate unique ways in which the theory of matching markets can be extended to account for more complex and often counterintuitive microeconomic phenomena.

cs.GT↗

arXiv2021

Separating Circuits : Switching Lemmas and Random Restrictions

This was submitted as a final project for CS254B, taught by Li Yang Tan and Tom Knowles. The field of Circuit Complexity utilises careful analysis of Boolean Circuit Functions in order to extract meaningful information about a range of complexity classes. In particular, the complexity class $P / \text{Poly}$ has played a central role in much of the historical attempts to tackle the problem of whether solution and verification are equivalent i.e. the central $P$ versus $NP$ problem. Whilst circuits can potentially be easier to analyse than Turing Machines due to their non-uniform nature of computation (program size is allowed to depend on the input size), it is notoriously hard to establish lower bounds for them. In this report, we will touch upon several results published by Hastad, Sipser and Razborov that will highlight a dynamic interplay between circuit complexity and many of the central ideas of modern-day complexity theory, and in particular the central importance of Hastad's Switching Lemma.

cs.CC↗