Yushu Li

Yushu Li, Xun Xu, Yongyi Su et al.

Generalizing deep learning models to unknown target domain distribution with low latency has motivated research into test-time training/adaptation (TTT/TTA). Existing approaches often focus on improving test-time training performance under well-curated target domain data. As figured out in this work, many state-of-the-art methods fail to maintain the performance when the target domain is contaminated with strong out-of-distribution (OOD) data, a.k.a. open-world test-time training (OWTTT). The failure is mainly due to the inability to distinguish strong OOD samples from regular weak OOD samples. To improve the robustness of OWTTT we first develop an adaptive strong OOD pruning which improves the efficacy of the self-training TTT method. We further propose a way to dynamically expand the prototypes to represent strong OOD samples for an improved weak/strong OOD data separation. Finally, we regularize self-training with distribution alignment and the combination yields the state-of-the-art performance on 5 OWTTT benchmarks. The code is available at https://github.com/Yushu-Li/OWTTT.

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.

Kaijun Gong, Zhuowen Yin, Yushu Li et al.

The redundancy of Convolutional neural networks not only depends on weights but also depends on inputs. Shuffling is an efficient operation for mixing channel information but the shuffle order is usually pre-defined. To reduce the data-dependent redundancy, we devise a dynamic shuffle module to generate data-dependent permutation matrices for shuffling. Since the dimension of permutation matrix is proportional to the square of the number of input channels, to make the generation process efficiently, we divide the channels into groups and generate two shared small permutation matrices for each group, and utilize Kronecker product and cross group shuffle to obtain the final permutation matrices. To make the generation process learnable, based on theoretical analysis, softmax, orthogonal regularization, and binarization are employed to asymptotically approximate the permutation matrix. Dynamic shuffle adaptively mixes channel information with negligible extra computation and memory occupancy. Experiment results on image classification benchmark datasets CIFAR-10, CIFAR-100, Tiny ImageNet and ImageNet have shown that our method significantly increases ShuffleNets' performance. Adding dynamic generated matrix with learnable static matrix, we further propose static-dynamic-shuffle and show that it can serve as a lightweight replacement of ordinary pointwise convolution.

Yushu Li, Yongyi Su, Adam Goodge et al.

Vision-language models (VLMs) have revolutionized machine learning by leveraging large pre-trained models to tackle various downstream tasks. Although label, training, and data efficiency have improved, many state-of-the-art VLMs still require task-specific hyperparameter tuning and fail to fully exploit test samples. To overcome these challenges, we propose a graph-based approach for label-efficient adaptation and inference. Our method dynamically constructs a graph over text prompts, few-shot examples, and test samples, using label propagation for inference without task-specific tuning. Unlike existing zero-shot label propagation techniques, our approach requires no additional unlabeled support set and effectively leverages the test sample manifold through dynamic graph expansion. We further introduce a context-aware feature re-weighting mechanism to improve task adaptation accuracy. Additionally, our method supports efficient graph expansion, enabling real-time inductive inference. Extensive evaluations on downstream tasks, such as fine-grained categorization and out-of-distribution generalization, demonstrate the effectiveness of our approach. The source code is available at https://github.com/Yushu-Li/ECALP.

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.

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.

Tiankai Chen, Yushu Li, Adam Goodge et al.

Out-of-distribution (OOD) detection in 3D point cloud data remains a challenge, particularly in applications where safe and robust perception is critical. While existing OOD detection methods have shown progress for 2D image data, extending these to 3D environments involves unique obstacles. This paper introduces a training-free framework that leverages Vision-Language Models (VLMs) for effective OOD detection in 3D point clouds. By constructing a graph based on class prototypes and testing data, we exploit the data manifold structure to enhancing the effectiveness of VLMs for 3D OOD detection. We propose a novel Graph Score Propagation (GSP) method that incorporates prompt clustering and self-training negative prompting to improve OOD scoring with VLM. Our method is also adaptable to few-shot scenarios, providing options for practical applications. We demonstrate that GSP consistently outperforms state-of-the-art methods across synthetic and real-world datasets 3D point cloud OOD detection.

Ingvild M. Helgøy, Yushu Li

The Bayesian Lasso is constructed in the linear regression framework and applies the Gibbs sampling to estimate the regression parameters. This paper develops a new sparse learning model, named the Bayesian Lasso Sparse (BLS) model, that takes the hierarchical model formulation of the Bayesian Lasso. The main difference from the original Bayesian Lasso lies in the estimation procedure; the BLS method uses a learning algorithm based on the type-II maximum likelihood procedure. Opposed to the Bayesian Lasso, the BLS provides sparse estimates of the regression parameters. The BLS method is also derived for nonlinear supervised learning problems by introducing kernel functions. We compare the BLS model to the well known Relevance Vector Machine, the Fast Laplace method, the Byesian Lasso, and the Lasso, on both simulated and real data. The numerical results show that the BLS is sparse and precise, especially when dealing with noisy and irregular dataset.