arXiv subjects

Pan Li

Publications and source records attributed to Pan Li.

At least 73 records · Page 4Linked to original sources

arXiv2026

Graph-KV: Breaking Sequence via Injecting Structural Biases into Large Language Models

Modern large language models (LLMs) are inherently auto-regressive, requiring input to be serialized into flat sequences regardless of their structural dependencies. This serialization hinders the model's ability to leverage structural inductive biases, especially in tasks such as retrieval-augmented generation (RAG) and reasoning on data with native graph structures, where inter-segment dependencies are crucial. We introduce Graph-KV with the potential to overcome this limitation. Graph-KV leverages the KV-cache of text segments as condensed representations and governs their interaction through structural inductive biases. In this framework, 'target' segments selectively attend only to the KV-caches of their designated 'source' segments, rather than all preceding segments in a serialized sequence. This approach induces a graph-structured block mask, sparsifying attention and enabling a message-passing-like step within the LLM. Furthermore, strategically allocated positional encodings for source and target segments reduce positional bias and context window consumption. We evaluate Graph-KV across three scenarios: (1) seven RAG benchmarks spanning direct inference, multi-hop reasoning, and long-document understanding; (2) Arxiv-QA, a novel academic paper QA task with full-text scientific papers structured as citation ego-graphs; and (3) paper topic classification within a citation network. By effectively reducing positional bias and harnessing structural inductive biases, Graph-KV substantially outperforms baselines, including standard costly sequential encoding, across various settings. Code and the Graph-KV data are publicly available.

cs.LG↗

arXiv2026

From Small to Large: Generalization Bounds for Transformers on Variable-Size Inputs

Transformers exhibit a notable property of \emph{size generalization}, demonstrating an ability to extrapolate from smaller token sets to significantly longer ones. This behavior has been documented across diverse applications, including point clouds, graphs, and natural language. Despite its empirical success, this capability still lacks some rigorous theoretical characterizations. In this paper, we develop a theoretical framework to analyze this phenomenon for geometric data, which we represent as discrete samples from a continuous source (e.g., point clouds from manifolds, graphs from graphons). Our core contribution is a bound on the error between the Transformer's output for a discrete sample and its continuous-domain equivalent. We prove that for Transformers with stable positional encodings, this bound is determined by the sampling density and the intrinsic dimensionality of the data manifold. Experiments on graphs and point clouds of various sizes confirm the tightness of our theoretical bound.

cs.LG↗

arXiv2025

Privacy Loss of Noise Perturbation via Concentration Analysis of A Product Measure

Noise perturbation is one of the most fundamental approaches for achieving $(ε,δ)$-differential privacy (DP) guarantees when releasing the result of a query or function $f(\cdot)\in\mathbb{R}^M$ evaluated on a sensitive dataset $\mathbf{x}$. In this approach, calibrated noise $\mathbf{n}\in\mathbb{R}^M$ is used to obscure the difference vector $f(\mathbf{x})-f(\mathbf{x}')$, where $\mathbf{x}'$ is known as a neighboring dataset. A DP guarantee is obtained by studying the tail probability bound of a privacy loss random variable (PLRV), defined as the Radon-Nikodym derivative between two distributions. When $\mathbf{n}$ follows a multivariate Gaussian distribution, the PLRV is characterized as a specific univariate Gaussian. In this paper, we propose a novel scheme to generate $\mathbf{n}$ by leveraging the fact that the perturbation noise is typically spherically symmetric (i.e., the distribution is rotationally invariant around the origin). The new noise generation scheme allows us to investigate the privacy loss from a geometric perspective and express the resulting PLRV using a product measure, $W\times U$; measure $W$ is related to a radius random variable controlling the magnitude of $\mathbf{n}$, while measure $U$ involves a directional random variable governing the angle between $\mathbf{n}$ and the difference $f(\mathbf{x})-f(\mathbf{x}')$. We derive a closed-form moment bound on the product measure to prove $(ε,δ)$-DP. Under the same $(ε,δ)$-DP guarantee, our mechanism yields a smaller expected noise magnitude than the classic Gaussian noise in high dimensions, thereby significantly improving the utility of the noisy result $f(\mathbf{x})+\mathbf{n}$. To validate this, we consider convex and non-convex empirical risk minimization (ERM) problems in high dimensional space and apply the proposed product noise to achieve privacy.

cs.CR↗

arXiv2025

Step-DeepResearch Technical Report

As LLMs shift toward autonomous agents, Deep Research has emerged as a pivotal metric. However, existing academic benchmarks like BrowseComp often fail to meet real-world demands for open-ended research, which requires robust skills in intent recognition, long-horizon decision-making, and cross-source verification. To address this, we introduce Step-DeepResearch, a cost-effective, end-to-end agent. We propose a Data Synthesis Strategy Based on Atomic Capabilities to reinforce planning and report writing, combined with a progressive training path from agentic mid-training to SFT and RL. Enhanced by a Checklist-style Judger, this approach significantly improves robustness. Furthermore, to bridge the evaluation gap in the Chinese domain, we establish ADR-Bench for realistic deep research scenarios. Experimental results show that Step-DeepResearch (32B) scores 61.4% on Scale AI Research Rubrics. On ADR-Bench, it significantly outperforms comparable models and rivals SOTA closed-source models like OpenAI and Gemini DeepResearch. These findings prove that refined training enables medium-sized models to achieve expert-level capabilities at industry-leading cost-efficiency.

cs.CL↗

arXiv2025

CoDA: A Context-Decoupled Hierarchical Agent with Reinforcement Learning

Large Language Model (LLM) agents trained with reinforcement learning (RL) show great promise for solving complex, multi-step tasks. However, their performance is often crippled by "Context Explosion", where the accumulation of long text outputs overwhelms the model's context window and leads to reasoning failures. To address this, we introduce CoDA, a Context-Decoupled hierarchical Agent, a simple but effective reinforcement learning framework that decouples high-level planning from low-level execution. It employs a single, shared LLM backbone that learns to operate in two distinct, contextually isolated roles: a high-level Planner that decomposes tasks within a concise strategic context, and a low-level Executor that handles tool interactions in an ephemeral, isolated workspace. We train this unified agent end-to-end using PECO (Planner-Executor Co-Optimization), a reinforcement learning methodology that applies a trajectory-level reward to jointly optimize both roles, fostering seamless collaboration through context-dependent policy updates. Extensive experiments demonstrate that CoDA achieves significant performance improvements over state-of-the-art baselines on complex multi-hop question-answering benchmarks, and it exhibits strong robustness in long-context scenarios, maintaining stable performance while all other baselines suffer severe degradation, thus further validating the effectiveness of our hierarchical design in mitigating context overload.

cs.CL↗

arXiv2025

RoFt-Mol: Benchmarking Robust Fine-Tuning with Molecular Graph Foundation Models

In the era of foundation models, fine-tuning pre-trained models for specific downstream tasks has become crucial. This drives the need for robust fine-tuning methods to address challenges such as model overfitting and sparse labeling. Molecular graph foundation models (MGFMs) face unique difficulties that complicate fine-tuning. These models are limited by smaller pre-training datasets and more severe data scarcity for downstream tasks, both of which require enhanced model generalization. Moreover, MGFMs must accommodate diverse objectives, including both regression and classification tasks. To better understand and improve fine-tuning techniques under these conditions, we classify eight fine-tuning methods into three mechanisms: weight-based, representation-based, and partial fine-tuning. We benchmark these methods on downstream regression and classification tasks across supervised and self-supervised pre-trained models in diverse labeling settings. This extensive evaluation provides valuable insights and informs the design of a refined robust fine-tuning method, ROFT-MOL. This approach combines the strengths of simple post-hoc weight interpolation with more complex weight ensemble fine-tuning methods, delivering improved performance across both task types while maintaining the ease of use inherent in post-hoc weight interpolation.

cs.LG↗

arXiv2025

Branching Strategies Based on Subgraph GNNs: A Study on Theoretical Promise versus Practical Reality

Graph Neural Networks (GNNs) have emerged as a promising approach for ``learning to branch'' in Mixed-Integer Linear Programming (MILP). While standard Message-Passing GNNs (MPNNs) are efficient, they theoretically lack the expressive power to fully represent MILP structures. Conversely, higher-order GNNs (like 2-FGNNs) are expressive but computationally prohibitive. In this work, we investigate Subgraph GNNs as a theoretical middle ground. Crucially, while previous work [Chen et al., 2025] demonstrated that GNNs with 3-WL expressive power can approximate Strong Branching, we prove a sharper result: node-anchored Subgraph GNNs whose expressive power is strictly lower than 3-WL [Zhang et al., 2023] are sufficient to approximate Strong Branching scores. However, our extensive empirical evaluation on four benchmark datasets reveals a stark contrast between theory and practice. While node-anchored Subgraph GNNs theoretically offer superior branching decisions, their $O(n)$ complexity overhead results in significant memory bottlenecks and slower solving times than MPNNs and heuristics. Our results indicate that for MILP branching, the computational cost of expressive GNNs currently outweighs their gains in decision quality, suggesting that future research must focus on efficiency-preserving expressivity.

cs.LG↗

arXiv2025

Locality-Sensitive Hashing-Based Efficient Point Transformer for Charged Particle Reconstruction

Charged particle track reconstruction is a foundational task in collider experiments and the main computational bottleneck in particle reconstruction. Graph neural networks (GNNs) have shown strong performance for this problem, but costly graph construction, irregular computations, and random memory access patterns substantially limit their throughput. The recently proposed Hashing-based Efficient Point Transformer (HEPT) offers a theoretically guaranteed near-linear complexity for large point cloud processing via locality-sensitive hashing (LSH) in attention computations; however, its evaluations have largely focused on embedding quality, and the object condensation pipeline on which HEPT relies requires a post-hoc clustering step (e.g., DBScan) that can dominate runtime. In this work, we make two contributions. First, we present a unified, fair evaluation of physics tracking performance for HEPT and a representative GNN-based pipeline under the same dataset and metrics. Second, we introduce HEPTv2 by extending HEPT with a lightweight decoder that eliminates the clustering stage and directly predicts track assignments. This modification preserves HEPT's regular, hardware-friendly computations while enabling ultra-fast end-to-end inference. On the TrackML dataset, optimized HEPTv2 achieves approximately 28 ms per event on an A100 while maintaining competitive tracking efficiency. These results position HEPTv2 as a practical, scalable alternative to GNN-based pipelines for fast tracking.

hep-ex↗

arXiv2025

When Truthful Representations Flip Under Deceptive Instructions?

Large language models (LLMs) tend to follow maliciously crafted instructions to generate deceptive responses, posing safety challenges. How deceptive instructions alter the internal representations of LLM compared to truthful ones remains poorly understood beyond output analysis. To bridge this gap, we investigate when and how these representations ``flip'', such as from truthful to deceptive, under deceptive versus truthful/neutral instructions. Analyzing the internal representations of Llama-3.1-8B-Instruct and Gemma-2-9B-Instruct on a factual verification task, we find the model's instructed True/False output is predictable via linear probes across all conditions based on the internal representation. Further, we use Sparse Autoencoders (SAEs) to show that the Deceptive instructions induce significant representational shifts compared to Truthful/Neutral representations (which are similar), concentrated in early-to-mid layers and detectable even on complex datasets. We also identify specific SAE features highly sensitive to deceptive instruction and use targeted visualizations to confirm distinct truthful/deceptive representational subspaces. % Our analysis pinpoints layer-wise and feature-level correlates of instructed dishonesty, offering insights for LLM detection and control. Our findings expose feature- and layer-level signatures of deception, offering new insights for detecting and mitigating instructed dishonesty in LLMs.

cs.AI↗

arXiv2025

Do LLMs Really Forget? Evaluating Unlearning with Knowledge Correlation and Confidence Awareness

Machine unlearning techniques aim to mitigate unintended memorization in large language models (LLMs). However, existing approaches predominantly focus on the explicit removal of isolated facts, often overlooking latent inferential dependencies and the non-deterministic nature of knowledge within LLMs. Consequently, facts presumed forgotten may persist implicitly through correlated information. To address these challenges, we propose a knowledge unlearning evaluation framework that more accurately captures the implicit structure of real-world knowledge by representing relevant factual contexts as knowledge graphs with associated confidence scores. We further develop an inference-based evaluation protocol leveraging powerful LLMs as judges; these judges reason over the extracted knowledge subgraph to determine unlearning success. Our LLM judges utilize carefully designed prompts and are calibrated against human evaluations to ensure their trustworthiness and stability. Extensive experiments on our newly constructed benchmark demonstrate that our framework provides a more realistic and rigorous assessment of unlearning performance. Moreover, our findings reveal that current evaluation strategies tend to overestimate unlearning effectiveness. Our code is publicly available at https://github.com/Graph-COM/Knowledge_Unlearning.git.

cs.CL↗

arXiv2025

Haystack Engineering: Context Engineering for Heterogeneous and Agentic Long-Context Evaluation

Modern long-context large language models (LLMs) perform well on synthetic "needle-in-a-haystack" (NIAH) benchmarks, but such tests overlook how noisy contexts arise from biased retrieval and agentic workflows. We argue that haystack engineering is necessary to construct noisy long contexts that faithfully capture key real-world factors -- distraction from heterogeneous biased retrievers and cascading errors in agentic workflows -- to test models' long-context robustness. We instantiate it through HaystackCraft, a new NIAH benchmark built on the full English Wikipedia hyperlink network with multi-hop questions. HaystackCraft evaluates how heterogeneous retrieval strategies (e.g., sparse, dense, hybrid, and graph-based) affect distractor composition, haystack ordering, and downstream LLM performance. HaystackCraft further extends NIAH to dynamic, LLM-dependent settings that simulate agentic operations, where models refine queries, reflect on their past reasonings, and decide when to stop. Experiments with 15 long-context models show that (1) while stronger dense retrievers can introduce more challenging distractors, graph-based reranking simultaneously improves retrieval effectiveness and mitigates more harmful distractors; (2) in agentic tests, even advanced models like Gemini 2.5 Pro and GPT-5 suffer cascading failures from self-generated distractors or struggle to perform early stops. These results highlight persistent challenges in agentic long-context reasoning and establish HaystackCraft as a valuable testbed for future progress.

cs.CL↗

arXiv2025

Struc-EMB: The Potential of Structure-Aware Encoding in Language Embeddings

Text embeddings from Large Language Models (LLMs) have become foundational for numerous applications. However, these models typically operate on raw text, overlooking the rich structural information, such as hyperlinks or citations, that provides crucial context in many real-world datasets. This paper introduces and systematically evaluates a new paradigm for generating structure-aware text embeddings by integrating these structural relations directly into the LLM's internal encoding process, rather than relying on traditional post-hoc aggregation. We investigate two primary in-process methods: sequential concatenation and parallel caching. Through extensive zero-shot experiments across retrieval, clustering, classification, and recommendation tasks, we demonstrate that our structure-aware approaches consistently outperform both text-only and post-hoc baselines. Our analysis reveals critical trade-offs: sequential concatenation excels with noisy, moderate-length contexts, while parallel caching scales more effectively to long, high-signal contexts but is more susceptible to distractors. To address the challenge of noisy structural data, we also introduce and validate two effective techniques: Context Distillation and Semantic Balancing. This work provides the first comprehensive analysis of in-process structure-aware encoding, offering a blueprint for building more powerful and contextually aware embedding models.

cs.LG↗

arXiv2025

Langevin Unlearning: A New Perspective of Noisy Gradient Descent for Machine Unlearning

Machine unlearning has raised significant interest with the adoption of laws ensuring the ``right to be forgotten''. Researchers have provided a probabilistic notion of approximate unlearning under a similar definition of Differential Privacy (DP), where privacy is defined as statistical indistinguishability to retraining from scratch. We propose Langevin unlearning, an unlearning framework based on noisy gradient descent with privacy guarantees for approximate unlearning problems. Langevin unlearning unifies the DP learning process and the privacy-certified unlearning process with many algorithmic benefits. These include approximate certified unlearning for non-convex problems, complexity saving compared to retraining, sequential and batch unlearning for multiple unlearning requests.

cs.LG↗

arXiv2025

Towards Universal Debiasing for Language Models-based Tabular Data Generation

Large language models (LLMs) have achieved promising results in tabular data generation. However, inherent historical biases in tabular datasets often cause LLMs to exacerbate fairness issues, particularly when multiple advantaged and protected features are involved. In this work, we introduce a universal debiasing framework that minimizes group-level dependencies by simultaneously reducing the mutual information between advantaged and protected attributes. By leveraging the autoregressive structure and analytic sampling distributions of LLM-based tabular data generators, our approach efficiently computes mutual information, reducing the need for cumbersome numerical estimations. Building on this foundation, we propose two complementary methods: a direct preference optimization (DPO)-based strategy, namely UDF-DPO, that integrates seamlessly with existing models, and a targeted debiasing technique, namely UDF-MIX, that achieves debiasing without tuning the parameters of LLMs. Extensive experiments demonstrate that our framework effectively balances fairness and utility, offering a scalable and practical solution for debiasing in high-stakes applications.

cs.LG↗

arXiv2025

Privately Learning from Graphs with Applications in Fine-tuning Large Language Models

Graphs offer unique insights into relationships between entities, complementing data modalities like text and images and enabling AI models to extend their capabilities beyond traditional tasks. However, learning from graphs often involves handling sensitive relationships in the data, raising significant privacy concerns. Existing privacy-preserving methods, such as DP-SGD, rely on gradient decoupling assumptions and are incompatible with relational learning due to the inherent dependencies between training samples. To address this challenge, we propose a privacy-preserving pipeline for relational learning that decouples dependencies in sampled relations for training, ensuring differential privacy through a tailored application of DP-SGD. We apply this approach to fine-tune large language models (LLMs), such as Llama2, on sensitive graph data while addressing the associated computational complexities. Our method is evaluated on four real-world text-attributed graphs, demonstrating significant improvements in relational learning tasks while maintaining robust privacy guarantees. Additionally, we analyze the trade-offs between privacy, utility, and computational efficiency, offering insights into the practical deployment of our approach for privacy-preserving relational learning. Code is available at https://github.com/Graph-COM/PvGaLM.

cs.LG↗

arXiv2025

Pruning Weights but Not Truth: Safeguarding Truthfulness While Pruning LLMs

Neural network pruning has emerged as a promising approach for deploying LLMs in low-resource scenarios while preserving downstream task performance. However, for the first time, we reveal that such pruning disrupts LLMs' internal activation features crucial for lie detection, where probing classifiers (typically small logistic regression models) trained on these features assess the truthfulness of LLM-generated statements. This discovery raises a crucial open question: how can we prune LLMs without sacrificing these critical lie detection capabilities? Our investigation further reveals that naively adjusting layer-wise pruning sparsity based on importance inadvertently removes crucial weights, failing to improve lie detection performance despite its reliance on the most crucial LLM layer. To address this issue, we propose Truthful Pruning aligned by Layer-wise Outliers (TPLO), which places greater emphasis on layers with more activation outliers and stronger discriminative features simultaneously. This preserves LLMs' original performance while retaining critical features of inner states needed for robust lie detection. Moreover, we introduce a prompting rule to enrich the TruthfulQA benchmark for better calibrating LLM pruning. Empirical results show that our approach improves the hallucination detection for pruned LLMs (achieving 88% accuracy at 50% sparsity) and enhances their performance on TruthfulQA.

cs.LG↗

arXiv2025

Quantized but Deceptive? A Multi-Dimensional Truthfulness Evaluation of Quantized LLMs

Quantization enables efficient deployment of large language models (LLMs) in resource-constrained environments by significantly reducing memory and computation costs. While quantized LLMs often maintain performance on perplexity and zero-shot tasks, their impact on truthfulness-whether generating truthful or deceptive responses-remains largely unexplored. In this work, we introduce TruthfulnessEval, a comprehensive evaluation framework for assessing the truthfulness of quantized LLMs across three dimensions: (1) Truthfulness on Logical Reasoning; (2) Truthfulness on Common Sense; and (3) Truthfulness on Imitative Falsehoods. Using this framework, we examine mainstream quantization techniques (ranging from 4-bit to extreme 2-bit) across several open-source LLMs. Surprisingly, we find that while quantized models retain internally truthful representations, they are more susceptible to producing false outputs under misleading prompts. To probe this vulnerability, we test 15 rephrased variants of "honest", "neutral" and "deceptive" prompts and observe that "deceptive" prompts can override truth-consistent behavior, whereas "honest" and "neutral" prompts maintain stable outputs. Further, we reveal that quantized models "know" the truth internally yet still produce false outputs when guided by "deceptive" prompts via layer-wise probing and PCA visualizations. Our findings provide insights into future designs of quantization-aware alignment and truthfulness interventions.

cs.AI↗

arXiv2025

Rethinking Addressing in Language Models via Contexualized Equivariant Positional Encoding

Transformers rely on both content-based and position-based addressing mechanisms to make predictions, but existing positional encoding techniques often diminish the effectiveness of position-based addressing. Many current methods enforce rigid patterns in attention maps, limiting the ability to model long-range dependencies and adapt to diverse tasks. Additionally, most positional encodings are learned as general biases, lacking the specialization required for different instances within a dataset. To address this, we propose con\textbf{T}extualized equivari\textbf{A}nt \textbf{P}osition \textbf{E}ncoding (\textbf{TAPE}), a novel framework that enhances positional embeddings by incorporating sequence content across layers. TAPE introduces dynamic, context-aware positional encodings, overcoming the constraints of traditional fixed patterns. We show that TAPE can provably facilitate LLM reasoning ability by emulating a broader class of algorithms. By enforcing permutation and orthogonal equivariance, TAPE ensures the stability of positional encodings during updates, improving long-context ability. Our method can be easily integrated into pre-trained transformers, offering parameter-efficient fine-tuning with minimal overhead. Extensive experiments show that TAPE achieves superior performance in language modeling, arithmetic reasoning, and long-context retrieval tasks compared to existing positional embedding techniques. Code is available at https://github.com/VITA-Group/TAPE.

cs.CL↗