Current frontier large-language models rely on reasoning to achieve state-of-the-art performance. Many existing interpretability are limited in this area, as standard methods have been designed to study single forward passes of a model rather than the multi-token computational steps that unfold during reasoning. We argue that analyzing reasoning traces at the sentence level is a promising approach to understanding reasoning processes. We introduce a black-box method that measures each sentence's counterfactual importance by repeatedly sampling replacement sentences from the model, filtering for semantically different ones, and continuing the chain of thought from that point onwards to quantify the sentence's impact on the distribution of final answers. We discover that certain sentences can have an outsized impact on the trajectory of the reasoning trace and final answer. We term these sentences \textit{thought anchors}. These are generally planning or uncertainty management sentences, and specialized attention heads consistently attend from subsequent sentences to thought anchors. We further show that examining sentence-sentence causal links within a reasoning trace gives insight into a model's behavior. Such information can be used to predict a problem's difficulty and the extent different question domains involve sequential or diffuse reasoning. As a proof-of-concept, we demonstrate that our techniques together provide a practical toolkit for analyzing reasoning models by conducting a detailed case study of how the model solves a difficult math problem, finding that our techniques yield a consistent picture of the reasoning trace's structure. We provide an open-source tool (thought-anchors.com) for visualizing the outputs of our methods on further problems. The convergence across our methods shows the potential of sentence-level analysis for a deeper understanding of reasoning models.

Generating high fidelity, differentially private (DP) synthetic images offers a promising route to share and analyze sensitive visual data without compromising individual privacy. However, existing DP image synthesis methods struggle to produce high resolution outputs that faithfully capture the structure of the original data. In this paper, we introduce a novel method, referred to as Synthesis via Private Textual Intermediaries (SPTI), that can generate high resolution DP images with easy adoption. The key idea is to shift the challenge of DP image synthesis from the image domain to the text domain by leveraging state of the art DP text generation methods. SPTI first summarizes each private image into a concise textual description using image to text models, then applies a modified Private Evolution algorithm to generate DP text, and finally reconstructs images using text to image models. Notably, SPTI requires no model training, only inference with off the shelf models. Given a private dataset, SPTI produces synthetic images of substantially higher quality than prior DP approaches. On the LSUN Bedroom dataset, SPTI attains an FID equal to 26.71 under epsilon equal to 1.0, improving over Private Evolution FID of 40.36. Similarly, on MM CelebA HQ, SPTI achieves an FID equal to 33.27 at epsilon equal to 1.0, compared to 57.01 from DP fine tuning baselines. Overall, our results demonstrate that Synthesis via Private Textual Intermediaries provides a resource efficient and proprietary model compatible framework for generating high resolution DP synthetic images, greatly expanding access to private visual datasets.

Large Language Models (LLMs) deployed in real-world settings increasingly face the need to unlearn sensitive, outdated, or proprietary information. Existing unlearning methods typically formulate forgetting and retention as a regularized trade-off, combining both objectives into a single scalarized loss. This often leads to unstable optimization and degraded performance on retained data, especially under aggressive forgetting. We propose a new formulation of LLM unlearning as a constrained optimization problem: forgetting is enforced via a novel logit-margin flattening loss that explicitly drives the output distribution toward uniformity on a designated forget set, while retention is preserved through a hard constraint on a separate retain set. Compared to entropy-based objectives, our loss is softmax-free, numerically stable, and maintains non-vanishing gradients, enabling more efficient and robust optimization. We solve the constrained problem using a scalable primal-dual algorithm that exposes the trade-off between forgetting and retention through the dynamics of the dual variable, all without any extra computational overhead. Evaluations on the TOFU and MUSE benchmarks across diverse LLM architectures demonstrate that our approach consistently matches or exceeds state-of-the-art baselines, effectively removing targeted information while preserving downstream utility.

We introduce $Ψ$-Sampler, an SMC-based framework incorporating pCNL-based initial particle sampling for effective inference-time reward alignment with a score-based generative model. Inference-time reward alignment with score-based generative models has recently gained significant traction, following a broader paradigm shift from pre-training to post-training optimization. At the core of this trend is the application of Sequential Monte Carlo (SMC) to the denoising process. However, existing methods typically initialize particles from the Gaussian prior, which inadequately captures reward-relevant regions and results in reduced sampling efficiency. We demonstrate that initializing from the reward-aware posterior significantly improves alignment performance. To enable posterior sampling in high-dimensional latent spaces, we introduce the preconditioned Crank-Nicolson Langevin (pCNL) algorithm, which combines dimension-robust proposals with gradient-informed dynamics. This approach enables efficient and scalable posterior sampling and consistently improves performance across various reward alignment tasks, including layout-to-image generation, quantity-aware generation, and aesthetic-preference generation, as demonstrated in our experiments. Project Webpage: https://psi-sampler.github.io/

Understanding the internal mechanisms of large language models (LLMs) remains a challenging and complex endeavor. Even fundamental questions, such as how fine-tuning affects model behavior, often require extensive empirical evaluation. In this paper, we introduce a novel perspective based on the geometric properties of contextual latent embeddings to study the effects of training and fine-tuning. To that end, we measure the local dimensions of a contextual language model's latent space and analyze their shifts during training and fine-tuning. We show that the local dimensions provide insights into the model's training dynamics and generalization ability. Specifically, the mean of the local dimensions predicts when the model's training capabilities are exhausted, as exemplified in a dialogue state tracking task, overfitting, as demonstrated in an emotion recognition task, and grokking, as illustrated with an arithmetic task. Furthermore, our experiments suggest a practical heuristic: reductions in the mean local dimension tend to accompany and predict subsequent performance gains. Through this exploration, we aim to provide practitioners with a deeper understanding of the implications of fine-tuning on embedding spaces, facilitating informed decisions when configuring models for specific applications. The results of this work contribute to the ongoing discourse on the interpretability, adaptability, and generalizability of LLMs by bridging the gap between intrinsic model mechanisms and geometric properties in the respective embeddings.

Fine-tuning large language models (LLMs) on telecom datasets is a common practice to adapt general-purpose models to the telecom domain. However, little attention has been paid to how this process may compromise model safety. Recent research has shown that even benign fine-tuning can degrade the safety alignment of LLMs, causing them to respond to harmful or unethical user queries. In this paper, we investigate this issue by fine-tuning LLMs on three representative telecom datasets and show that safety degrades even for light telecom domain adaptation. To this end, we introduce TeleHarm, the first telecom-specific red-teaming benchmark, which we use alongside established Direct-Harm and HexPhi datasets to systematically assess harmful behavior. We further extend our analysis to publicly available TeleLLMs that were continually pre-trained on large telecom corpora, revealing that safety alignment is severely lacking, primarily due to the omission of safety-focused instruction tuning. To address these issues, we evaluate three realignment defenses: SafeInstruct, SafeLoRA, SafeMERGE. We show that, across all settings, the proposed defenses can effectively restore safety without compromising telecom task performance, leading to Safe teleCOMMunication (SafeCOMM) models. Our work serves as both a diagnostic study and practical guide for safety realignment in telecom-tuned LLMs, underscoring the need for safety-aware instruction and fine-tuning in the telecom domain.

Open-domain question answering (OpenQA) represents a cornerstone in natural language processing (NLP), primarily focused on extracting answers from unstructured textual data. With the rapid advancements in Large Language Models (LLMs), LLM-based OpenQA methods have reaped the benefits of emergent understanding and answering capabilities enabled by massive parameters compared to traditional methods. However, most of these methods encounter two critical challenges: how to integrate knowledge into LLMs effectively and how to adaptively generate results with specific answer formats for various task situations. To address these challenges, we propose a novel framework named GenKI, which aims to improve the OpenQA performance by exploring Knowledge Integration and controllable Generation on LLMs simultaneously. Specifically, we first train a dense passage retrieval model to retrieve associated knowledge from a given knowledge base. Subsequently, we introduce a novel knowledge integration model that incorporates the retrieval knowledge into instructions during fine-tuning to intensify the model. Furthermore, to enable controllable generation in LLMs, we leverage a certain fine-tuned LLM and an ensemble based on text consistency incorporating all coherence, fluency, and answer format assurance. Finally, extensive experiments conducted on the TriviaQA, MSMARCO, and CMRC2018 datasets, featuring diverse answer formats, have demonstrated the effectiveness of GenKI with comparison of state-of-the-art baselines. Moreover, ablation studies have disclosed a linear relationship between the frequency of retrieved knowledge and the model's ability to recall knowledge accurately against the ground truth. Our code of GenKI is available at https://github.com/USTC-StarTeam/GenKI

Large language models are able to exploit in-context learning to access external knowledge beyond their training data through retrieval-augmentation. While promising, its inner workings remain unclear. In this work, we shed light on the mechanism of in-context retrieval augmentation for question answering by viewing a prompt as a composition of informational components. We propose an attribution-based method to identify specialized attention heads, revealing in-context heads that comprehend instructions and retrieve relevant contextual information, and parametric heads that store entities' relational knowledge. To better understand their roles, we extract function vectors and modify their attention weights to show how they can influence the answer generation process. Finally, we leverage the gained insights to trace the sources of knowledge used during inference, paving the way towards more safe and transparent language models.

E-commerce sellers are recommended keyphrases based on their inventory on which they advertise to increase buyer engagement (clicks/sales). The relevance of advertiser keyphrases plays an important role in preventing the inundation of search systems with numerous irrelevant items that compete for attention in auctions, in addition to maintaining a healthy seller perception. In this work, we describe the shortcomings of training Advertiser keyphrase relevance filter models on click/sales/search relevance signals and the importance of aligning with human judgment, as sellers have the power to adopt or reject said keyphrase recommendations. In this study, we frame Advertiser keyphrase relevance as a complex interaction between 3 dynamical systems -- seller judgment, which influences seller adoption of our product, Advertising, which provides the keyphrases to bid on, and Search, who holds the auctions for the same keyphrases. This study discusses the practicalities of using human judgment via a case study at eBay Advertising and demonstrate that using LLM-as-a-judge en-masse as a scalable proxy for seller judgment to train our relevance models achieves a better harmony across the three systems -- provided that they are bound by a meticulous evaluation framework grounded in business metrics.

We present Agentic Retrieval-Augmented Code Synthesis (ARCS), a system that improves LLM-based code generation without fine-tuning. ARCS operates through a budgeted synthesize-execute-repair loop over a frozen model: it retrieves relevant code context before generation, proposes candidates, executes them against tests, and repairs based on execution feedback. This retrieval-before-generation design reduces hallucination and accelerates convergence. We formalize ARCS as a state-action process with provable guarantees on termination, monotonic improvement, and bounded cost. A tiered controller (Small/Medium/Large) trades latency for accuracy predictably. On HumanEval, ARCS achieves up to 87.2% pass@1 with Llama-3.1-405B, surpassing CodeAgent (82.3%) while using simpler control than tree-search methods. On TransCoder, it achieves >= 90% accuracy on most translation pairs. On a LANL scientific corpus, it improves CodeBLEU by +0.115 over baseline RAG. ARCS provides a practical, reproducible approach to reliable code synthesis using existing LLM checkpoints.