arXiv subjects

Bo Li

Publications and source records attributed to Bo Li.

At least 433 records · Page 24Linked to original sources

arXiv2025

On Memorization of Large Language Models in Logical Reasoning

Large language models (LLMs) achieve good performance on challenging reasoning benchmarks, yet could also make basic reasoning mistakes. This contrasting behavior is puzzling when it comes to understanding the mechanisms behind LLMs' reasoning capabilities. One hypothesis is that the increasingly high and nearly saturated performance on common reasoning benchmarks could be due to the memorization of similar problems. In this paper, we systematically investigate this hypothesis with a quantitative measurement of memorization in reasoning tasks, using a dynamically generated logical reasoning benchmark based on Knights and Knaves (K&K) puzzles. We find that LLMs could interpolate and memorize the training puzzles (achieving near-perfect accuracy) after fine-tuning, yet they struggle with slight variations of these puzzles. On the other hand, we show that while fine-tuning leads to heavy memorization, it also consistently improves generalization performance. Through in-depth analyses with perturbation tests, cross difficulty-level transferability, probing model internals, and fine-tuning with wrong answers, we establish that LLMs develop reasoning skills on K&K puzzles alongside memorization. Finally, our analysis based on a per-sample memorization score sheds light on how LLMs switch between reasoning and memorization when solving logical puzzles. Our code and data are available at https://memkklogic.github.io.

cs.CL↗

arXiv2025

BECAUSE: Bilinear Causal Representation for Generalizable Offline Model-based Reinforcement Learning

Offline model-based reinforcement learning (MBRL) enhances data efficiency by utilizing pre-collected datasets to learn models and policies, especially in scenarios where exploration is costly or infeasible. Nevertheless, its performance often suffers from the objective mismatch between model and policy learning, resulting in inferior performance despite accurate model predictions. This paper first identifies the primary source of this mismatch comes from the underlying confounders present in offline data for MBRL. Subsequently, we introduce \textbf{B}ilin\textbf{E}ar \textbf{CAUS}al r\textbf{E}presentation~(BECAUSE), an algorithm to capture causal representation for both states and actions to reduce the influence of the distribution shift, thus mitigating the objective mismatch problem. Comprehensive evaluations on 18 tasks that vary in data quality and environment context demonstrate the superior performance of BECAUSE over existing offline RL algorithms. We show the generalizability and robustness of BECAUSE under fewer samples or larger numbers of confounders. Additionally, we offer theoretical analysis of BECAUSE to prove its error bound and sample efficiency when integrating causal representation into offline MBRL.

cs.LG↗

arXiv2025

Learning for Feasible Region on Coal Mine Virtual Power Plants with Imperfect Information

The feasible region assessment (FRA) in industrial virtual power plants (VPPs) is driven by the need to activate large-scale latent industrial loads for demand response, making it essential to aggregate these flexible resources for peak regulation. However, the large number of devices and the need for privacy preservation in coal mines pose challenges to accurately aggregating these resources into a cohesive coal mine VPP. In this paper, we propose an efficient and reliable data-driven approach for FRA in the coal mine VPP that can manage incomplete information. Our data-driven FRA algorithm approximates equipment and FRA parameters based on historical energy dispatch data, effectively addressing the challenges of imperfect information. Simulation results illustrate that our method approximates the accurate feasible operational boundaries under dynamic and imperfect information conditions.

eess.SY↗

arXiv2025

SORRY-Bench: Systematically Evaluating Large Language Model Safety Refusal

Evaluating aligned large language models' (LLMs) ability to recognize and reject unsafe user requests is crucial for safe, policy-compliant deployments. Existing evaluation efforts, however, face three limitations that we address with SORRY-Bench, our proposed benchmark. First, existing methods often use coarse-grained taxonomies of unsafe topics, and are over-representing some fine-grained topics. For example, among the ten existing datasets that we evaluated, tests for refusals of self-harm instructions are over 3x less represented than tests for fraudulent activities. SORRY-Bench improves on this by using a fine-grained taxonomy of 44 potentially unsafe topics, and 440 class-balanced unsafe instructions, compiled through human-in-the-loop methods. Second, linguistic characteristics and formatting of prompts are often overlooked, like different languages, dialects, and more -- which are only implicitly considered in many evaluations. We supplement SORRY-Bench with 20 diverse linguistic augmentations to systematically examine these effects. Third, existing evaluations rely on large LLMs (e.g., GPT-4) for evaluation, which can be computationally expensive. We investigate design choices for creating a fast, accurate automated safety evaluator. By collecting 7K+ human annotations and conducting a meta-evaluation of diverse LLM-as-a-judge designs, we show that fine-tuned 7B LLMs can achieve accuracy comparable to GPT-4 scale LLMs, with lower computational cost. Putting these together, we evaluate over 50 proprietary and open-weight LLMs on SORRY-Bench, analyzing their distinctive safety refusal behaviors. We hope our effort provides a building block for systematic evaluations of LLMs' safety refusal capabilities, in a balanced, granular, and efficient manner. Benchmark demo, data, code, and models are available through https://sorry-bench.github.io.

cs.AI↗

arXiv2025

EAIRA: Establishing a Methodology for Evaluating AI Models as Scientific Research Assistants

Recent advancements have positioned AI, and particularly Large Language Models (LLMs), as transformative tools for scientific research, capable of addressing complex tasks that require reasoning, problem-solving, and decision-making. Their exceptional capabilities suggest their potential as scientific research assistants but also highlight the need for holistic, rigorous, and domain-specific evaluation to assess effectiveness in real-world scientific applications. This paper describes a multifaceted methodology for Evaluating AI models as scientific Research Assistants (EAIRA) developed at Argonne National Laboratory. This methodology incorporates four primary classes of evaluations. 1) Multiple Choice Questions to assess factual recall; 2) Open Response to evaluate advanced reasoning and problem-solving skills; 3) Lab-Style Experiments involving detailed analysis of capabilities as research assistants in controlled environments; and 4) Field-Style Experiments to capture researcher-LLM interactions at scale in a wide range of scientific domains and applications. These complementary methods enable a comprehensive analysis of LLM strengths and weaknesses with respect to their scientific knowledge, reasoning abilities, and adaptability. Recognizing the rapid pace of LLM advancements, we designed the methodology to evolve and adapt so as to ensure its continued relevance and applicability. This paper describes the methodology state at the end of February 2025. Although developed within a subset of scientific domains, the methodology is designed to be generalizable to a wide range of scientific domains.

cs.AI↗

arXiv2025

Towards Statistical Factuality Guarantee for Large Vision-Language Models

Advancements in Large Vision-Language Models (LVLMs) have demonstrated promising performance in a variety of vision-language tasks involving image-conditioned free-form text generation. However, growing concerns about hallucinations in LVLMs, where the generated text is inconsistent with the visual context, are becoming a major impediment to deploying these models in applications that demand guaranteed reliability. In this paper, we introduce a framework to address this challenge, ConfLVLM, which is grounded on conformal prediction to achieve finite-sample distribution-free statistical guarantees on the factuality of LVLM output. This framework treats an LVLM as a hypothesis generator, where each generated text detail (or claim) is considered an individual hypothesis. It then applies a statistical hypothesis testing procedure to verify each claim using efficient heuristic uncertainty measures to filter out unreliable claims before returning any responses to users. We conduct extensive experiments covering three representative application domains, including general scene understanding, medical radiology report generation, and document understanding. Remarkably, ConfLVLM reduces the error rate of claims generated by LLaVa-1.5 for scene descriptions from 87.8\% to 10.0\% by filtering out erroneous claims with a 95.3\% true positive rate. Our results further demonstrate that ConfLVLM is highly flexible, and can be applied to any black-box LVLMs paired with any uncertainty measure for any image-conditioned free-form text generation task while providing a rigorous guarantee on controlling the risk of hallucination.

cs.LG↗

arXiv2025

PhenoProfiler: Advancing Phenotypic Learning for Image-based Drug Discovery

In the field of image-based drug discovery, capturing the phenotypic response of cells to various drug treatments and perturbations is a crucial step. However, existing methods require computationally extensive and complex multi-step procedures, which can introduce inefficiencies, limit generalizability, and increase potential errors. To address these challenges, we present PhenoProfiler, an innovative model designed to efficiently and effectively extract morphological representations, enabling the elucidation of phenotypic changes induced by treatments. PhenoProfiler is designed as an end-to-end tool that processes whole-slide multi-channel images directly into low-dimensional quantitative representations, eliminating the extensive computational steps required by existing methods. It also includes a multi-objective learning module to enhance robustness, accuracy, and generalization in morphological representation learning. PhenoProfiler is rigorously evaluated on large-scale publicly available datasets, including over 230,000 whole-slide multi-channel images in end-to-end scenarios and more than 8.42 million single-cell images in non-end-to-end settings. Across these benchmarks, PhenoProfiler consistently outperforms state-of-the-art methods by up to 20%, demonstrating substantial improvements in both accuracy and robustness. Furthermore, PhenoProfiler uses a tailored phenotype correction strategy to emphasize relative phenotypic changes under treatments, facilitating the detection of biologically meaningful signals. UMAP visualizations of treatment profiles demonstrate PhenoProfiler ability to effectively cluster treatments with similar biological annotations, thereby enhancing interpretability. These findings establish PhenoProfiler as a scalable, generalizable, and robust tool for phenotypic learning.

cs.LG↗

arXiv2025

Practical Performative Policy Learning with Strategic Agents

This paper studies the performative policy learning problem, where agents adjust their features in response to a released policy to improve their potential outcomes, inducing an endogenous distribution shift. There has been growing interest in training machine learning models in strategic environments, including strategic classification and performative prediction. However, existing approaches often rely on restrictive parametric assumptions: micro-level utility models in strategic classification and macro-level data distribution maps in performative prediction, severely limiting scalability and generalizability. We approach this problem as a complex causal inference task, relaxing parametric assumptions on both micro-level agent behavior and macro-level data distribution. Leveraging bounded rationality, we uncover a practical low-dimensional structure in distribution shifts and construct an effective mediator in the causal path from the deployed model to the shifted data. We then propose a gradient-based policy optimization algorithm with a differentiable classifier as a substitute for the high-dimensional distribution map. Our algorithm efficiently utilizes batch feedback and limited manipulation patterns. Our approach achieves high sample efficiency compared to methods reliant on bandit feedback or zero-order optimization. We also provide theoretical guarantees for algorithmic convergence. Extensive and challenging experiments on high-dimensional settings demonstrate our method's practical efficacy.

cs.LG↗

arXiv2025

The Lottery LLM Hypothesis, Rethinking What Abilities Should LLM Compression Preserve?

Motivated by reducing the computational and storage costs of LLMs, model compression and KV cache compression have attracted much attention from researchers. However, current methods predominantly emphasize maintaining the performance of compressed LLMs, as measured by perplexity or simple accuracy on tasks of common sense knowledge QA and basic arithmetic reasoning. In this blog, we present a brief review of recent advancements in LLMs related to retrieval-augmented generation, multi-step reasoning, external tools, and computational expressivity, all of which substantially enhance LLM performance. Then, we propose a lottery LLM hypothesis suggesting that for a given LLM and task, there exists a smaller lottery LLM capable of producing the same performance as the original LLM with the assistance of multi-step reasoning and external tools. Based on the review of current progress in LLMs, we discuss and summarize the essential capabilities that the lottery LLM and KV cache compression must possess, which are currently overlooked in existing methods.

cs.LG↗

arXiv2025

FetDTIAlign: A Deep Learning Framework for Affine and Deformable Registration of Fetal Brain dMRI

Diffusion MRI (dMRI) provides unique insights into fetal brain microstructure in utero. Longitudinal and cross-sectional fetal dMRI studies can reveal crucial neurodevelopmental changes but require precise spatial alignment across scans and subjects. This is challenging due to low data quality, rapid brain development, and limited anatomical landmarks. Existing registration methods, designed for high-quality adult data, struggle with these complexities. To address this, we introduce FetDTIAlign, a deep learning approach for fetal brain dMRI registration, enabling accurate affine and deformable alignment. FetDTIAlign features a dual-encoder architecture and iterative feature-based inference, reducing the impact of noise and low resolution. It optimizes network configurations and domain-specific features at each registration stage, enhancing both robustness and accuracy. We validated FetDTIAlign on data from 23 to 36 weeks gestation, covering 60 white matter tracts. It consistently outperformed two classical optimization-based methods and a deep learning pipeline, achieving superior anatomical correspondence. Further validation on external data from the Developing Human Connectome Project confirmed its generalizability across acquisition protocols. Our results demonstrate the feasibility of deep learning for fetal brain dMRI registration, providing a more accurate and reliable alternative to classical techniques. By enabling precise cross-subject and tract-specific analyses, FetDTIAlign supports new discoveries in early brain development.

eess.IV↗

arXiv2025

Giant Uncompensated Magnon Spin Currents in X-type Magnets

Magnon spin currents in insulating magnets are useful for low-power spintronics. However, in magnets stacked by antiferromagnetic (AFM) exchange coupling, which have recently aroused significant interest for potential applications in spintronics, these currents are largely counteracted by opposite magnetic sublattices, thus suppressing their net effect. Contrary to this common observation, here, we show that magnets with X-type AFM stacking, where opposite magnetic sublattices form orthogonal intersecting chains, support giant magnon spin currents with minimal compensation. Our model Hamiltonian calculations predict magnetic chain locking of magnon spin currents in these X-type magnets, significantly reducing their compensation ratio. In addition, the one-dimensional nature of the chain-like magnetic sublattices enhances magnon spin conductivities surpassing those of two-dimensional ferromagnets and canonical altermagnets. Notably, uncompensated X-type magnets, such as odd-layer antiferromagnets and ferrimagnets, can exhibit magnon spin currents polarized opposite to those expected by their net magnetization. These unprecedented properties of X-type magnets, combined with their inherent advantages resulting from AFM coupling, offer a promising new path for low-power high-performance spintronics.

cond-mat.mtrl-sci↗

arXiv2025

On the Subsidy of Envy-Free Orientations in Graphs

We study a fair division problem in (multi)graphs where $n$ agents (vertices) are pairwise connected by items (edges), and each agent is only interested in its incident items. We consider how to allocate items to incident agents in an envy-free manner, i.e., envy-free orientations, while minimizing the overall payment, i.e., subsidy. We first prove that computing an envy-free orientation with the minimum subsidy is NP-hard, even when the graph is simple and the agents have bi-valued additive valuations. We then bound the worst-case subsidy. We prove that for any multigraph (i.e., allowing parallel edges) and monotone valuations where the marginal value of each good is at most \$1 for each agent, \$1 each (a total subsidy of $n-1$, where $n$ is the number of agents) is sufficient. This is one of the few cases where linear subsidy $Θ(n)$ is known to be necessary and sufficient to guarantee envy-freeness when agents have monotone valuations. When the valuations are additive (while the graph may contain parallel edges) and when the graph is simple (while the valuations may be monotone), we improve the bound to $n/2$ and $n-2$, respectively. Moreover, these two bounds are tight.

cs.GT↗

arXiv2025

Step-Audio: Unified Understanding and Generation in Intelligent Speech Interaction

Real-time speech interaction, serving as a fundamental interface for human-machine collaboration, holds immense potential. However, current open-source models face limitations such as high costs in voice data collection, weakness in dynamic control, and limited intelligence. To address these challenges, this paper introduces Step-Audio, the first production-ready open-source solution. Key contributions include: 1) a 130B-parameter unified speech-text multi-modal model that achieves unified understanding and generation, with the Step-Audio-Chat version open-sourced; 2) a generative speech data engine that establishes an affordable voice cloning framework and produces the open-sourced lightweight Step-Audio-TTS-3B model through distillation; 3) an instruction-driven fine control system enabling dynamic adjustments across dialects, emotions, singing, and RAP; 4) an enhanced cognitive architecture augmented with tool calling and role-playing abilities to manage complex tasks effectively. Based on our new StepEval-Audio-360 evaluation benchmark, Step-Audio achieves state-of-the-art performance in human evaluations, especially in terms of instruction following. On open-source benchmarks like LLaMA Question, shows 9.3% average performance improvement, demonstrating our commitment to advancing the development of open-source multi-modal language technologies. Our code and models are available at https://github.com/stepfun-ai/Step-Audio.

cs.CL↗

arXiv2025

SafeChain: Safety of Language Models with Long Chain-of-Thought Reasoning Capabilities

Emerging large reasoning models (LRMs), such as DeepSeek-R1 models, leverage long chain-of-thought (CoT) reasoning to generate structured intermediate steps, enhancing their reasoning capabilities. However, long CoT does not inherently guarantee safe outputs, potentially leading to harmful consequences such as the introduction of security vulnerabilities in code or the spread of misinformation. Current research on large language model (LLM) safety usually focuses on short-answer responses, overlooking the long CoT style outputs of LRMs. To bridge this gap, we conduct a systematic study of LRM safety. First, we investigate safety evaluators calibrated against human annotations. Using our newly developed metrics, we thoroughly assess the safety of 12 state-of-the-art LRMs on StrongReject and WildJailbreak datasets. Our results show that LRMs are not safe compared to their reasoning advance. Further, we perform a fine-grained analysis of the reasoning trace and final answer. We find that three decoding strategies-ZeroThink, LessThink, and MoreThink-can improve model safety without additional training. However, these strategies either use constrained reasoning traces or incur high inference costs. To better strengthen LRM safety, we introduce SafeChain, the first-of-its-kind safety training dataset in CoT style. We fine-tune two LRMs with SafeChain, showing that it not only enhances model safety but also preserves performance across 6 reasoning benchmarks.

cs.AI↗

arXiv2025

DreamDDP: Accelerating Data Parallel Distributed LLM Training with Layer-wise Scheduled Partial Synchronization

The growth of large language models (LLMs) increases challenges of accelerating distributed training across multiple GPUs in different data centers. Moreover, concerns about data privacy and data exhaustion have heightened interest in geo-distributed data centers. Communication in geo-distributed data parallel training (DDP) with stochastic gradient descent (S-SGD) is the main bottleneck in low-bandwidth environments. Local SGD mitigates communication overhead by reducing synchronization frequency, and recent studies have successfully applied it to geo-distributedly pre-train LLMs. However, we identify that its model synchronization mechanism prevents overlapping communication and computation, which makes the system lose opportunities to overlap communication and computation. To overcome this limitation, we expand the design space of local SGD by layer-wisely decoupling model synchronization. In each iteration, only some layers are synchronized instead of the entire model after a specific number of iterations. Leveraging this methodology, we introduce DreamDDP, a training framework to accelerate low-bandwidth distributed training with three key innovations: (1) partial local SGD with theoretical assurances of convergence rates comparable to S-SGD; (2) overlapping parameter synchronization with computation without extra GPU memory occupation; (3) identifying and exploiting three properties to schedule the communication and computation to reduce the training time based on fine-grained profiling of layer-wise communication and computation time. Empirical evaluations conducted on 32 GPUs using prominent deep learning models, including ResNet-18, ResNet-50, GPT-2, and Llama-2, demonstrate that DreamDDP enhances the convergence properties of Local SGD (and Adam) and achieves speedups ranging from $1.49\times$ to $3.91\times$ over leading baseline methods.

cs.DC↗

arXiv2025

ControllableGPT: A Ground-Up Designed Controllable GPT for Molecule Optimization

Large Language Models (LLMs) employ three popular training approaches: Masked Language Models (MLM), Causal Language Models (CLM), and Sequence-to-Sequence Models (seq2seq). However, each approach has its strengths and limitations, and faces challenges in addressing specific tasks that require controllable and bidirectional generation, such as drug optimization. To address this challenge, inspired by the biological processes of growth and evolution, which involve the expansion, shrinking, and mutation of sequences, we introduce ControllableGPT. This initiative represents the first effort to combine the advantages of MLM, CLM, and seq2seq into a single unified, controllable GPT framework. It enables the precise management of specific locations and ranges within a sequence, allowing for expansion, reduction, or mutation over chosen or random lengths, while maintaining the integrity of any specified positions or subsequences. In this work, we designed ControllableGPT for drug optimization from the ground up, which included proposing the Causally Masked Seq2seq (CMS) objective, developing the training corpus, introducing a novel pre-training approach, and devising a unique generation process. We demonstrate the effectiveness and controllability of ControllableGPT by conducting experiments on drug optimization tasks for both viral and cancer benchmarks, surpassing competing baselines.

cs.LG↗

arXiv2025

Janus: Collaborative Vision Transformer Under Dynamic Network Environment

Vision Transformers (ViTs) have outperformed traditional Convolutional Neural Network architectures and achieved state-of-the-art results in various computer vision tasks. Since ViTs are computationally expensive, the models either have to be pruned to run on resource-limited edge devices only or have to be executed on remote cloud servers after receiving the raw data transmitted over fluctuating networks. The resulting degraded performance or high latency all hinder their widespread applications. In this paper, we present Janus, the first framework for low-latency cloud-device collaborative Vision Transformer inference over dynamic networks. Janus overcomes the intrinsic model limitations of ViTs and realizes collaboratively executing ViT models on both cloud and edge devices, achieving low latency, high accuracy, and low communication overhead. Specifically, Janus judiciously combines token pruning techniques with a carefully designed fine-to-coarse model splitting policy and non-static mixed pruning policy. It attains a balance between accuracy and latency by dynamically selecting the optimal pruning level and split point. Experimental results across various tasks demonstrate that Janus enhances throughput by up to 5.15 times and reduces latency violation ratios by up to 98.7% when compared with baseline approaches under various network environments.

cs.DC↗

arXiv2025

Tamper-Resistant Safeguards for Open-Weight LLMs

Rapid advances in the capabilities of large language models (LLMs) have raised widespread concerns regarding their potential for malicious use. Open-weight LLMs present unique challenges, as existing safeguards lack robustness to tampering attacks that modify model weights. For example, recent works have demonstrated that refusal and unlearning safeguards can be trivially removed with a few steps of fine-tuning. These vulnerabilities necessitate new approaches for enabling the safe release of open-weight LLMs. We develop a method, called TAR, for building tamper-resistant safeguards into open-weight LLMs such that adversaries cannot remove the safeguards even after hundreds of steps of fine-tuning. In extensive evaluations and red teaming analyses, we find that our method greatly improves tamper-resistance while preserving benign capabilities. Our results demonstrate that progress on tamper-resistance is possible, opening up a promising new avenue to improve the safety and security of open-weight LLMs.

cs.LG↗