arXiv subjects

Claire Zhang

Publications and source records attributed to Claire Zhang.

5 recordsLinked to original sources

arXiv2026

Inference-Time Consensus for Mitigating Hidden Behaviors from LLM Fine-Tuning

Recent work shows that fine-tuning language models on even a small amount of poisoned data can install targeted misbehavior, and ostensibly benign data can transmit hidden preferences that generalize broadly. Standard defenses, such as data filtering, mixing in harmless data, and regularization, attenuate these effects but do not eliminate them. We instead pursue robustness through redundancy: collecting multiple datasets from different sources and only learning what is common between them. Thus, if only a subset of sources are malicious, the misbehavior will be blocked. In order to implement this defense strategy, we fine-tune a separate reference model on each source's dataset and aggregate their next-token distributions at decoding time. We introduce two consensus decoders: a token-wise minimum, which caps each token at the lowest probability any source assigns, and a base-relative variant, which reverts to the base probability on any token the sources move in opposing directions. We further relax exact agreement to tolerate partial support across sources and different surface expressions of the same intention. Across controlled poisoning tasks, subliminal learning, and emergent misalignment, consensus decoding suppresses source-specific misbehavior while preserving shared desirable behavior, including cases where union training and weight averaging retain the unwanted behavior.

cs.AI↗

arXiv2026

Scalable Energy-Based Models via Adversarial Training: Unifying Discrimination and Generation

Simultaneously achieving robust classification and high-fidelity generative modeling within a single framework presents a significant challenge. Hybrid approaches, such as Joint Energy-Based Models (JEM), interpret classifiers as EBMs but are often limited by the instability and poor sample quality inherent in training based on Stochastic Gradient Langevin Dynamics (SGLD). We address these limitations by proposing a novel training framework that integrates adversarial training (AT) principles for both discriminative robustness and stable generative learning. The proposed method introduces three key innovations: (1) the replacement of SGLD-based JEM learning with a stable, AT-based approach that optimizes the energy function through a Binary Cross-Entropy (BCE) loss that discriminates between real data and contrastive samples generated via Projected Gradient Descent (PGD); (2) adversarial training for the discriminative component that enhances classification robustness while implicitly providing the gradient regularization needed for stable EBM training; and (3) a two-stage training strategy that addresses normalization-related instabilities and enables leveraging pretrained robust classifiers, generalizing effectively across architectures. Experiments on CIFAR-10/100 and ImageNet demonstrate that our approach: (1) is the first EBM-based hybrid to scale to high-resolution datasets with high training stability, simultaneously achieving state-of-the-art discriminative and generative performance on ImageNet 256x256; (2) uniquely combines generative quality with adversarial robustness, enabling faithful counterfactual explanations; and (3) functions as a competitive standalone generative model, matching autoregressive models and surpassing diffusion models while offering additional versatility.

cs.LG↗

arXiv2026

CoRPO: Adding a Correctness Bias to GRPO Improves Generalization

Group-Relative Policy Optimization (GRPO) has emerged as the standard for training reasoning capabilities in large language models through reinforcement learning. By estimating advantages using group-mean rewards rather than a learned critic, GRPO has enabled efficient scaling of reinforcement learning from verifiable rewards (RLVR). However, we identify a fundamental limitation: GRPO's mean baseline can assign positive advantages to incorrect solutions simply because they outperform a poorly-performing group average. It leads to overestimation of advantages and reinforcement of incorrect behaviours. To address this, we propose Correctness-Relative Policy Optimization (CoRPO), a simple modification to the GRPO objective that clips the minimum baseline to a fixed correctness threshold. We show that baseline clipping introduces a protective bias to advantage estimation that mitigates overfitting while preserving effective exploration. Empirically, CoRPO-trained models improve cross-domain reasoning, generalizing more consistently to out-of-domain (OOD) tasks. When trained on coding tasks, CoRPO outperforms GRPO on math, and vice-versa, indicating that CoRPO learns robust, transferable reasoning patterns rather than task-specific solutions.

cs.AI↗

arXiv2026

Semantic Search At LinkedIn

Semantic search with large language models (LLMs) enables retrieval by meaning rather than keyword overlap, but scaling it requires major inference efficiency advances. We present LinkedIn's LLM-based semantic search framework for AI Job Search and AI People Search, combining an LLM relevance judge, embedding-based retrieval, and a compact Small Language Model trained via multi-teacher distillation to jointly optimize relevance and engagement. A prefill-oriented inference architecture co-designed with model pruning, context compression, and text-embedding hybrid interactions boosts ranking throughput by over 75x under a fixed latency constraint while preserving near-teacher-level NDCG, enabling one of the first production LLM-based ranking systems with efficiency comparable to traditional approaches and delivering significant gains in quality and user engagement.

cs.IR↗

arXiv2025

Shotgun assembly of random regular graphs

Mossel and Ross (2019) introduce the shotgun assembly problem for random graphs: what radius $R$ ensures that the random graph $G$ can be uniquely recovered from its list of rooted $R$-neighborhoods, with high probability? Here we consider this question for random regular graphs of fixed degree $d\ge3$. A result of Bollobás (1982) implies efficient recovery at $R = (1 + ε) \frac12 \log_{d-1}n$ with high probability -- moreover, this recovery algorithm uses only a summary of the distances in each neighborhood. We show that using the full neighborhood structure gives a sharper bound \[ R = \frac{\log n + \log\log n}{2\log(d-1)} + O(1)\,, \] which we prove is tight up to the $O(1)$ term. One consequence of our proof is that if $G,H$ are independent graphs where $G$ follows the random regular law, then with high probability the graphs are non-isomorphic; furthermore, this can be efficiently certified by testing the $R$-neighborhood list of $H$ against the $R$-neighborhood of a single adversarially chosen vertex of $G$.

math.PR↗