Wenlong Deng

Anushree Bannadabhavi, Soojin Lee, Wenlong Deng et al.

Autism spectrum disorder(ASD) is a lifelong neurodevelopmental condition that affects social communication and behavior. Investigating functional magnetic resonance imaging (fMRI)-based brain functional connectome can aid in the understanding and diagnosis of ASD, leading to more effective treatments. The brain is modeled as a network of brain Regions of Interest (ROIs), and ROIs form communities and knowledge of these communities is crucial for ASD diagnosis. On the one hand, Transformer-based models have proven to be highly effective across several tasks, including fMRI connectome analysis to learn useful representations of ROIs. On the other hand, existing transformer-based models treat all ROIs equally and overlook the impact of community-specific associations when learning node embeddings. To fill this gap, we propose a novel method, Com-BrainTF, a hierarchical local-global transformer architecture that learns intra and inter-community aware node embeddings for ASD prediction task. Furthermore, we avoid over-parameterization by sharing the local transformer parameters for different communities but optimize unique learnable prompt tokens for each community. Our model outperforms state-of-the-art (SOTA) architecture on ABIDE dataset and has high interpretability, evident from the attention module. Our code is available at https://github.com/ubc-tea/Com-BrainTF.

Wenlong Deng, Lang Lang, Zhen Liu et al.

In light of the smoothness property brought by skip connections in ResNet, this paper proposed the Skip Logit to introduce the skip connection mechanism that fits arbitrary DNN dimensions and embraces similar properties to ResNet. Meta Tanh Normalization (MTN) is designed to learn variance information and stabilize the training process. With these delicate designs, our Skip Meta Logit (SML) brought incremental boosts to the performance of extensive SOTA ctr prediction models on two real-world datasets. In the meantime, we prove that the optimization landscape of arbitrarily deep skip logit networks has no spurious local optima. Finally, SML can be easily added to building blocks and has delivered offline accuracy and online business metrics gains on app ads learning to rank systems at TikTok.

Wenlong Deng, Jiaji Huang, Kaan Ozkara et al.

Reward hacking arises when a model improves a proxy reward by exploiting shortcuts rather than solving the intended task. We study this failure mode through the geometry of reinforcement learning updates in language models and argue that hacking emerges when optimization drifts away from a stable low-dimensional learning trajectory. We analyze this drift through dominant singular directions of parameter updates and show that reward-hacking runs exhibit substantially larger directional change than clean runs. Motivated by this observation, we introduce trusted-direction projection, which constrains gradients to remain within a clean reference subspace. Across reward-hacking experiments on mathematical reasoning, the proposed approach delays shortcut exploitation and better preserves task performance.

Wenlong Deng, Yuan Zhong, Qi Dou et al.

Mitigating the discrimination of machine learning models has gained increasing attention in medical image analysis. However, rare works focus on fair treatments for patients with multiple sensitive demographic ones, which is a crucial yet challenging problem for real-world clinical applications. In this paper, we propose a novel method for fair representation learning with respect to multi-sensitive attributes. We pursue the independence between target and multi-sensitive representations by achieving orthogonality in the representation space. Concretely, we enforce the column space orthogonality by keeping target information on the complement of a low-rank sensitive space. Furthermore, in the row space, we encourage feature dimensions between target and sensitive representations to be orthogonal. The effectiveness of the proposed method is demonstrated with extensive experiments on the CheXpert dataset. To our best knowledge, this is the first work to mitigate unfairness with respect to multiple sensitive attributes in the field of medical imaging.

Juncheng Wu, Wenlong Deng, Xingxuan Li et al.

Medical tasks such as diagnosis and treatment planning require precise and complex reasoning, particularly in life-critical domains. Unlike mathematical reasoning, medical reasoning demands meticulous, verifiable thought processes to ensure reliability and accuracy. However, there is a notable lack of datasets that provide transparent, step-by-step reasoning to validate and enhance the medical reasoning ability of AI models. To bridge this gap, we introduce MedReason, a large-scale high-quality medical reasoning dataset designed to enable faithful and explainable medical problem-solving in large language models (LLMs). We utilize a structured medical knowledge graph (KG) to convert clinical QA pairs into logical chains of reasoning, or ``thinking paths'', which trace connections from question elements to answers via relevant KG entities. Each path is validated for consistency with clinical logic and evidence-based medicine. Our pipeline generates detailed reasoning for various medical questions from 7 medical datasets, resulting in a dataset of 32,682 question-answer pairs, each with detailed, step-by-step explanations. Experiments demonstrate that fine-tuning with our dataset consistently boosts medical problem-solving capabilities, achieving significant gains of up to 7.7% for DeepSeek-Ditill-8B. Our top-performing model, MedReason-8B, outperforms the Huatuo-o1-8B, a state-of-the-art medical reasoning model, by up to 4.2% on the clinical benchmark MedBullets. We also engage medical professionals from diverse specialties to assess our dataset's quality, ensuring MedReason offers accurate and coherent medical reasoning. Our data, models, and code is available at https://github.com/UCSC-VLAA/MedReason.

Wenlong Deng, Christos Thrampoulidis, Xiaoxiao Li

Vision Transformers (ViT) and Visual Prompt Tuning (VPT) achieve state-of-the-art performance with improved efficiency in various computer vision tasks. This suggests a promising paradigm shift of adapting pre-trained ViT models to Federated Learning (FL) settings. However, the challenge of data heterogeneity among FL clients presents a significant hurdle in effectively deploying ViT models. Existing Generalized FL (GFL) and Personalized FL (PFL) methods have limitations in balancing performance across both global and local data distributions. In this paper, we present a novel algorithm, SGPT, that integrates GFL and PFL approaches by employing a unique combination of both shared and group-specific prompts. This design enables SGPT to capture both common and group-specific features. A key feature of SGPT is its prompt selection module, which facilitates the training of a single global model capable of automatically adapting to diverse local client data distributions without the need for local fine-tuning. To effectively train the prompts, we utilize block coordinate descent (BCD), learning from common feature information (shared prompts), and then more specialized knowledge (group prompts) iteratively. Theoretically, we justify that learning the proposed prompts can reduce the gap between global and local performance. Empirically, we conduct experiments on both label and feature heterogeneity settings in comparison with state-of-the-art baselines, along with extensive ablation studies, to substantiate the superior performance of SGPT.

Wenlong Deng, Yize Zhao, Vala Vakilian et al.

Storing open-source fine-tuned models separately introduces redundancy and increases response times in applications utilizing multiple models. Delta-parameter pruning (DPP), particularly the random drop and rescale (DARE) method proposed by Yu et al., addresses this by pruning the majority of delta parameters--the differences between fine-tuned and pre-trained model weights--while typically maintaining minimal performance loss. However, DARE fails when either the pruning rate or the magnitude of the delta parameters is large. We highlight two key reasons for this failure: (1) an excessively large rescaling factor as pruning rates increase, and (2) high mean and variance in the delta parameters. To push DARE's limits, we introduce DAREx (DARE the eXtreme), which features two algorithmic improvements: (1) DAREx-q, a rescaling factor modification that significantly boosts performance at high pruning rates (e.g., >30 % on COLA and SST2 for encoder models, with even greater gains in decoder models), and (2) DAREx-L2, which combines DARE with AdamR, an in-training method that applies appropriate delta regularization before DPP. We also demonstrate that DAREx-q can be seamlessly combined with vanilla parameter-efficient fine-tuning techniques like LoRA and can facilitate structural DPP. Additionally, we revisit the application of importance-based pruning techniques within DPP, demonstrating that they outperform random-based methods when delta parameters are large. Through this comprehensive study, we develop a pipeline for selecting the most appropriate DPP method under various practical scenarios.

Ruinan Jin, Wenlong Deng, Minghui Chen et al.

In the era of Foundation Models' (FMs) rising prominence in AI, our study addresses the challenge of biases in medical images while the model operates in black-box (e.g., using FM API), particularly spurious correlations between pixels and sensitive attributes. Traditional methods for bias mitigation face limitations due to the restricted access to web-hosted FMs and difficulties in addressing the underlying bias encoded within the FM API. We propose a D(ebiased) N(oise) E(diting) strategy, termed DNE, which generates DNE noise to mask such spurious correlation. DNE is capable of mitigating bias both within the FM API embedding and the images themselves. Furthermore, DNE is suitable for both white-box and black-box FM APIs, where we introduced G(reedy) (Z)eroth-O(rder) (GeZO) optimization for it when the gradient is inaccessible in black-box APIs. Our whole pipeline enables fairness-aware image editing that can be applied across various medical contexts without requiring direct model manipulation or significant computational resources. Our empirical results demonstrate the method's effectiveness in maintaining fairness and utility across different patient groups and diseases. In the era of AI-driven medicine, this work contributes to making healthcare diagnostics more equitable, showcasing a practical solution for bias mitigation in pre-trained image FMs. Our code is provided at https://github.com/ubc-tea/DNE-foundation-model-fairness.

Wenlong Deng, Yushu Li, Boying Gong et al.

Tool-integrated (TI) reinforcement learning (RL) enables large language models (LLMs) to perform multi-step reasoning by interacting with external tools such as search engines and retrievers. Group Relative Policy Optimization (GRPO), exemplified by the recent Search-R1, offers fast convergence and a value-free formulation that makes it appealing for this setting, yet consistently suffers from training collapse. We identify Lazy Likelihood Displacement (LLD), a systematic reduction or stagnation in the likelihood of both correct and incorrect responses, as the core mechanism driving this failure. LLD emerges early and triggers a self-reinforcing LLD Death Spiral, where declining likelihood leads to low-confidence responses, inflating gradients, and ultimately causing collapse. We empirically characterize this process across models on a Search-R1-style, search-integrated question answering task, revealing a consistent three-phase trajectory: early stagnation, steady decay, and accelerated collapse. To address this, we propose a lightweight likelihood-preserving regularization LLDS for GRPO that activates only when a trajectory's likelihood decreases, and regularizes only the tokens responsible. This fine-grained structure mitigates LLD with minimal interference to optimization. Across seven open-domain and multi-hop QA benchmarks, our method stabilizes training, prevents gradient explosion, and yields substantial performance improvements, including +37.8% gains on Qwen2.5-3B and +32.0% gains on Qwen2.5-7B. Our results establish LLD as a fundamental bottleneck in GRPO-based TIRL and provide a practical path toward stable, scalable training of tool-integrated LLM.

Beidi Zhao, Wenlong Deng, Xinting Liao et al.

While Retrieval-Augmented Generation (RAG) is one of the dominant paradigms for enhancing Large Vision-Language Models (LVLMs) on knowledge-based VQA tasks, recent work attributes RAG failures to insufficient attention towards the retrieved context, proposing to reduce the attention allocated to image tokens. In this work, we identify a distinct failure mode that previous study overlooked: Attention Distraction (AD). When the retrieved context is sufficient (highly relevant or including the correct answer), the retrieved text suppresses the visual attention globally, and the attention on image tokens shifts away from question-relevant regions. This leads to failures on questions the model could originally answer correctly without the retrieved text. To mitigate this issue, we propose MAD-RAG, a training-free intervention that decouples visual grounding from context integration through a dual-question formulation, combined with attention mixing to preserve image-conditioned evidence. Extensive experiments on OK-VQA, E-VQA, and InfoSeek demonstrate that MAD-RAG consistently outperforms existing baselines across different model families, yielding absolute gains of up to 4.76%, 9.20%, and 6.18% over the vanilla RAG baseline. Notably, MAD-RAG rectifies up to 74.68% of failure cases with negligible computational overhead.

Beidi Zhao, Wenlong Deng, Zi Han et al.

In computational pathology, multiple instance learning (MIL) is widely used to circumvent the computational impasse in giga-pixel whole slide image (WSI) analysis. It usually consists of two stages: patch-level feature extraction and slide-level aggregation. Recently, pretrained models or self-supervised learning have been used to extract patch features, but they suffer from low effectiveness or inefficiency due to overlooking the task-specific supervision provided by slide labels. Here we propose a weakly-supervised Label-Efficient WSI Screening method, dubbed LESS, for cytological WSI analysis with only slide-level labels, which can be effectively applied to small datasets. First, we suggest using variational positive-unlabeled (VPU) learning to uncover hidden labels of both benign and malignant patches. We provide appropriate supervision by using slide-level labels to improve the learning of patch-level features. Next, we take into account the sparse and random arrangement of cells in cytological WSIs. To address this, we propose a strategy to crop patches at multiple scales and utilize a cross-attention vision transformer (CrossViT) to combine information from different scales for WSI classification. The combination of our two steps achieves task-alignment, improving effectiveness and efficiency. We validate the proposed label-efficient method on a urine cytology WSI dataset encompassing 130 samples (13,000 patches) and FNAC 2019 dataset with 212 samples (21,200 patches). The experiment shows that the proposed LESS reaches 84.79%, 85.43%, 91.79% and 78.30% on a urine cytology WSI dataset, and 96.88%, 96.86%, 98.95%, 97.06% on FNAC 2019 dataset in terms of accuracy, AUC, sensitivity and specificity. It outperforms state-of-the-art MIL methods on pathology WSIs and realizes automatic cytological WSI cancer screening.

Wenlong Deng, Yi Ren, Muchen Li et al.

Reinforcement learning (RL) has become popular in enhancing the reasoning capabilities of large language models (LLMs), with Group Relative Policy Optimization (GRPO) emerging as a widely used algorithm in recent systems. Despite GRPO's widespread adoption, we identify a previously unrecognized phenomenon we term Lazy Likelihood Displacement (LLD), wherein the likelihood of correct responses marginally increases or even decreases during training. This behavior mirrors a recently discovered misalignment issue in Direct Preference Optimization (DPO), attributed to the influence of negative gradients. We provide a theoretical analysis of GRPO's learning dynamic, identifying the source of LLD as the naive penalization of all tokens in incorrect responses with the same strength. To address this, we develop a method called NTHR, which downweights penalties on tokens contributing to the LLD. Unlike prior DPO-based approaches, NTHR takes advantage of GRPO's group-based structure, using correct responses as anchors to identify influential tokens. Experiments on math reasoning benchmarks demonstrate that NTHR effectively mitigates LLD, yielding consistent performance gains across models ranging from 0.5B to 3B parameters.

Minghui Chen, Ruinan Jin, Wenlong Deng et al.

Recent studies highlight the promise of LLM-based prompt optimization, especially with TextGrad, which automates differentiation'' via texts and backpropagates textual feedback. This approach facilitates training in various real-world applications that do not support numerical gradient propagation or loss calculation. In this paper, we systematically explore the potential and challenges of incorporating textual gradient into Federated Learning (FL). Our contributions are fourfold. Firstly, we introduce a novel FL paradigm, Federated Textual Gradient (FedTextGrad), that allows clients to upload locally optimized prompts derived from textual gradients, while the server aggregates the received prompts. Unlike traditional FL frameworks, which are designed for numerical aggregation, FedTextGrad is specifically tailored for handling textual data, expanding the applicability of FL to a broader range of problems that lack well-defined numerical loss functions. Secondly, building on this design, we conduct extensive experiments to explore the feasibility of FedTextGrad. Our findings highlight the importance of properly tuning key factors (e.g., local steps) in FL training. Thirdly, we highlight a major challenge in FedTextGrad aggregation: retaining essential information from distributed prompt updates. Last but not least, in response to this issue, we improve the vanilla variant of FedTextGrad by providing actionable guidance to the LLM when summarizing client prompts by leveraging the Uniform Information Density principle. Through this principled study, we enable the adoption of textual gradients in FL for optimizing LLMs, identify important issues, and pinpoint future directions, thereby opening up a new research area that warrants further investigation.

Yushu Li, Wenlong Deng, Jiajin Li et al.

Test-time scaling has become a dominant paradigm for improving LLM agent reliability, yet current approaches treat compute as an abundant resource, allowing agents to exhaust token and tool budgets on redundant steps or dead-end trajectories. Existing budget-aware methods either require expensive fine-tuning or rely on coarse, trajectory-level heuristics that cannot intervene mid-execution. We propose the Budget-Aware Value Tree (BAVT), a training-free inference-time framework that models multi-hop reasoning as a dynamic search tree guided by step-level value estimation within a single LLM backbone. Another key innovation is a budget-conditioned node selection mechanism that uses the remaining resource ratio as a natural scaling exponent over node values, providing a principled, parameter-free transition from broad exploration to greedy exploitation as the budget depletes. To combat the well-known overconfidence of LLM self-evaluation, BAVT employs a residual value predictor that scores relative progress rather than absolute state quality, enabling reliable pruning of uninformative or redundant tool calls. We further provide a theoretical convergence guarantee, proving that BAVT reaches a terminal answer with probability at least $1-ε$ under an explicit finite budget bound. Extensive evaluations on four multi-hop QA benchmarks across two model families demonstrate that BAVT consistently outperforms parallel sampling baselines. Most notably, BAVT under strict low-budget constraints surpasses baseline performance at $4\times$ the resource allocation, establishing that intelligent budget management fundamentally outperforms brute-force compute scaling.

Wenlong Deng, Yi Ren, Yushu Li et al.

Reinforcement learning with verifiable rewards has significantly advanced the reasoning capabilities of large language models, yet how to explicitly steer training toward exploration or exploitation remains an open problem. We introduce Token Hidden Reward (THR), a token-level metric that quantifies each token's influence on the likelihood of correct responses under Group Relative Policy Optimization (GRPO). We find that training dynamics are dominated by a small subset of tokens with high absolute THR values. Most interestingly, tokens with positive THR strengthen confidence in correct outputs, thus favoring exploitation, while tokens with negative THR preserve probability mass for alternative outputs, enabling exploration. This insight suggests a natural intervention: a THR-guided reweighting algorithm that modulates GRPO's learning signals to explicitly bias training toward exploitation or exploration. We validate the efficacy of this algorithm on diverse math reasoning benchmarks. By amplifying tokens with positive THR value and weakening negative ones, our algorithm improves greedy-decoding accuracy, favoring exploitation. The reverse strategy yields consistent gains in Pass@K accuracy, favoring exploration. We further demonstrate that our algorithm integrates seamlessly with other RL objectives such as GSPO and generalizes across architectures including Llama. These findings establish THR as a principled and fine-grained mechanism for dynamically controlling exploration and exploitation in RL-tuned LLMs, providing new tools for targeted fine-tuning in reasoning-intensive applications.

Running Yang, Wenlong Deng, Minghui Chen et al.

Clinical tasks such as diagnosis and treatment require strong decision-making abilities, highlighting the importance of rigorous evaluation benchmarks to assess the reliability of large language models (LLMs). In this work, we introduce a knowledge-guided data augmentation framework that enhances the difficulty of clinical multiple-choice question (MCQ) datasets by generating distractors (i.e., incorrect choices that are similar to the correct one and may confuse existing LLMs). Using our KG-based pipeline, the generated choices are both clinically plausible and deliberately misleading. Our approach involves multi-step, semantically informed walks on a medical knowledge graph to identify distractor paths-associations that are medically relevant but factually incorrect-which then guide the LLM in crafting more deceptive distractors. We apply the designed knowledge graph guided distractor generation (KGGDG) pipline, to six widely used medical QA benchmarks and show that it consistently reduces the accuracy of state-of-the-art LLMs. These findings establish KGGDG as a powerful tool for enabling more robust and diagnostic evaluations of medical LLMs.

Minghui Chen, Wenlong Deng, James Zou et al.

Large language models (LLMs) are increasingly deployed as part of compound AI systems that coordinate multiple modules (e.g., retrievers, tools, verifiers) over long-horizon workflows. Recent approaches that propagate textual feedback globally (e.g., TextGrad) make it feasible to optimize such pipelines, but we find that performance degrades as system depth grows. In particular, long-horizon agentic workflows exhibit two depth-scaling failure modes: 1) exploding textual gradient, where textual feedback grows exponentially with depth, leading to prohibitively long message and amplifies evaluation biases; and 2) vanishing textual gradient, where limited long-context ability causes models overemphasize partial feedback and compression of lengthy feedback causes downstream messages to lose specificity gradually as they propagate many hops upstream. To mitigate these issues, we introduce Textual Equilibrium Propagation (TEP), a local learning principle inspired by Equilibrium Propagation in energy-based models. TEP includes two phases: 1) a free phase where a local LLM critics iteratively refine prompts until reaching equilibrium (no further improvements are suggested); and 2) a nudged phase which applies proximal prompt edits with bounded modification intensity, using task-level objectives that propagate via forward signaling rather than backward feedback chains. This design supports local prompt optimization followed by controlled adaptation toward global goals without the computational burden and signal degradation of global textual backpropagation. Across long-horizon QA benchmarks and multi-agent tool-use dataset, TEP consistently improves accuracy and efficiency over global propagation methods such as TextGrad. The gains grows with depth, while preserving the practicality of black-box LLM components in deep compound AI system.

Christos Thrampoulidis, Sadegh Mahdavi, Wenlong Deng

This note reconciles two seemingly distinct approaches to policy gradient optimization for the Pass@K objective in reinforcement learning with verifiable rewards: (1) direct REINFORCE-style methods, and (2) advantage-shaping techniques that directly modify GRPO. We show that these are two sides of the same coin. By reverse-engineering existing advantage-shaping algorithms, we reveal that they implicitly optimize surrogate rewards. We specifically interpret practical "hard-example up-weighting" modifications to GRPO as reward-level regularization. Conversely, starting from surrogate reward objectives, we provide a simple recipe for deriving both existing and new advantage-shaping methods. This perspective provides a lens for RLVR policy gradient optimization beyond our original motivation of Pass@K.

Wenlong Deng, Jiaming Zhang, Qi Zeng et al.

Quantifying the influence of individual training samples is essential for enhancing the transparency and accountability of large language models (LLMs) and vision-language models (VLMs). However, existing data valuation methods often rely on Hessian information or model retraining, making them computationally prohibitive for billion-parameter models. In this work, we introduce For-Value, a forward-only data valuation framework that enables scalable and efficient influence estimation for both LLMs and VLMs. By leveraging the rich representations of modern foundation models, For-Value computes influence scores using a simple closed-form expression based solely on a single forward pass, thereby eliminating the need for costly gradient computations. Our theoretical analysis demonstrates that For-Value accurately estimates per-sample influence by capturing alignment in hidden representations and prediction errors between training and validation samples. Extensive experiments show that For-Value matches or outperforms gradient-based baselines in identifying impactful fine-tuning examples and effectively detecting mislabeled data.

Wenlong Deng, Blair Chen, Beidi Zhao et al.

Mitigating biases in machine learning models has become an increasing concern in Natural Language Processing (NLP), particularly in developing fair text embeddings, which are crucial yet challenging for real-world applications like search engines. In response, this paper proposes a novel method for learning fair text embeddings. First, we define a novel content-conditional equal distance (CCED) fairness for text embeddings, ensuring content-conditional independence between sensitive attributes and text embeddings. Building on CCED, we introduce a content-conditional debiasing (CCD) loss to ensure that embeddings of texts with different sensitive attributes but identical content maintain the same distance from the embedding of their corresponding neutral text. Additionally, we tackle the issue of insufficient training data by using Large Language Models (LLMs) with instructions to fairly augment texts into different sensitive groups. Our extensive evaluations show that our approach effectively enhances fairness while maintaining the utility of embeddings. Furthermore, our augmented dataset, combined with the CCED metric, serves as an new benchmark for evaluating fairness.