论文检索

General Machine Learning · Representation Learning

Yisi Luo, Deyu Meng

Separable neural networks (SepNNs) are emerging neural architectures that significantly reduce computational costs by factorizing a multivariate function into linear combinations of univariate functions, benefiting downstream applications such as implicit neural representations (INRs) and physics-informed neural networks (PINNs). However, fundamental theoretical analysis for SepNN, including detailed representation capacity and spectral bias characterization \& alleviation, remains unexplored. This work makes three key contributions to theoretically understanding and improving SepNN. First, using Weierstrass-based approximation and universal approximation theory, we prove that SepNN can approximate any multivariate function with arbitrary precision, confirming its representation completeness. Second, we derive the neural tangent kernel (NTK) regimes for SepNN, showing that the NTK of infinite-width SepNN converges to a deterministic (or random) kernel under infinite (or fixed) decomposition rank, with corresponding convergence and spectral bias characterization. Third, we propose an efficient separable preconditioned gradient descent (SepPGD) for optimizing SepNN, which alleviates the spectral bias of SepNN by provably adjusting its NTK spectrum. The SepPGD enjoys an efficient $\mathcal{O}(nD)$ complexity for $n^D$ training samples, which is much more efficient than previous neural network PGD methods. Extensive experiments for kernel ridge regression, image and surface representation using INRs, and numerical PDEs using PINNs validate the efficiency of SepNN and the effectiveness of SepPGD for alleviating spectral bias.

Computer Vision · Vision Models & Multimodal

Xinyi Gu, Jiayuan Mao, Zhang-Wei Hong, Zhuoran Yu, Pengyuan Li, Dhiraj Joshi, Rogerio Feris, Zexue He

Pretrained multi-modal large language models (MLLMs) demonstrate strong performance on diverse multimodal tasks, but remain limited in reasoning capabilities for domains where annotations are difficult to collect. In this work, we focus on artificial image domains such as charts, rendered documents, and webpages, which are abundant in practice yet lack large-scale human annotated reasoning datasets. We introduce COGS (COmposition-Grounded data Synthesis), a data-efficient framework for equipping MLLMs with advanced reasoning abilities from a small set of seed questions. The key idea is to decompose each seed question into primitive perception and reasoning *factors*, which can then be systematically recomposed with new images to generate large collections of synthetic question-answer pairs. Each generated question is paired with subquestions and intermediate answers, enabling reinforcement learning with factor-level process rewards. Experiments on chart reasoning show that COGS substantially improves performance on unseen questions, with the largest gains on reasoning-heavy and compositional questions. Moreover, training with a factor-level mixture of different seed data yields better transfer across multiple datasets, suggesting that COGS induces generalizable capabilities rather than dataset-specific overfitting. We further demonstrate that the framework extends beyond charts to other domains such as webpages. We release the code and data at https://cogsynthesis.github.io.

Optimization · Large Scale, Parallel and Distributed

Xiaowen Jiang, Anton Rodomanov, Sebastian Stich

Different federated optimization algorithms typically employ distinct client-selection strategies: some methods communicate only with a randomly sampled subset of clients at each round, while others need to periodically communicate with all clients or use a hybrid scheme that combines both strategies. However, existing metrics for comparing optimization methods typically do not distinguish between these strategies, which often incur different communication costs in practice. To address this disparity, we introduce a simple and natural model of federated optimization that quantifies communication and local computation complexities. This new model allows for several commonly used client-selection strategies and explicitly associates each with a distinct cost. Within this setting, we propose a new algorithm that achieves the best-known communication and local complexities among existing federated optimization methods for non-convex optimization. This algorithm is based on the inexact composite gradient method with a carefully constructed gradient estimator and a special procedure for solving the auxiliary subproblem at each iteration. The gradient estimator is based on SAGA, a popular variance-reduced gradient estimator. We first derive a new variance bound for it, showing that SAGA can exploit functional similarity. We then introduce the Recursive-Gradient technique as a general way to potentially improve the error bound of a given conditionally unbiased gradient estimator, including both SAGA and SVRG. By applying this technique to SAGA, we obtain a new estimator, RG-SAGA, which has an improved error bound compared to the original one.

Applications · Robotics

William Liang, Sam Wang, Hung-Ju Wang, Osbert Bastani, Yecheng Jason Ma, Dinesh Jayaraman

The ability to conduct and learn from interaction and experience is a central challenge in robotics, offering a scalable alternative to labor-intensive human demonstrations. However, realizing such "play" requires (1) a policy robust to diverse, potentially out-of-distribution environment states, and (2) a procedure that continuously produces useful robot experience. To address these challenges, we introduce Tether, a method for autonomous functional play involving structured, task-directed interactions. First, we design a novel open-loop policy that warps actions from a small set of source demonstrations (≤10) by anchoring them to semantic keypoint correspondences in the target scene. We show that this design is extremely data-efficient and robust even under significant spatial and semantic variations. Second, we deploy this policy for autonomous functional play in the real world via a continuous cycle of task selection, execution, evaluation, and improvement, guided by the visual understanding capabilities of vision-language models. This procedure generates diverse, high-quality datasets with minimal human intervention. In a household-like multi-object setup, our method is the first to perform many hours of autonomous multi-task play in the real world starting from only a handful of demonstrations. This produces a stream of data that consistently improves the performance of closed-loop imitation policies over time, ultimately yielding over 1000 expert-level trajectories and training policies competitive with those learned from human-collected demonstrations.

Applications · Language, Speech and Dialog

Wayne Chi, Valerie Chen, Ryan Shar, Aditya Mittal, Jenny Liang, Wei-Lin Chiang, Anastasios Angelopoulos, Ion Stoica, Graham Neubig, Ameet Talwalkar 等

Instructed code editing, where LLMs directly modify a developer's existing code based on a user instruction, is becoming a widely used interaction mode in AI coding assistants. However, few benchmarks directly evaluate this capability and current datasets often rely on artificial sources. We introduce EditBench, a benchmark for evaluating LLM code editing capabilities grounded in real-world usage, i.e.,~user instructions and code contexts collected in the wild. EditBench comprises of 545 problems, multiple natural and programming languages, and a diverse set of real-world use cases, ranging from resolving errors to adding features. EditBench introduces context-dependent problems that require the model to understand code context, highlighted code, and cursor position in addition to the user instruction. We evaluate 40 diverse LLMs and observe that EditBench is a challenging set of problems where only 3 models score over 60\%. We find that model performance varies across different categories of user instructions. Further, we find that varying levels of contextual information greatly affect task success rate, with performance varying up to 11\%, indicating the importance of evaluating with realistic context.

Computer Vision · Vision Models & Multimodal

Jingzhong Lin, Xinru Li, Yuanyuan Qi, Bohao Zhang, Wenxiang Liu, Kecheng Tang, Wenxuan Huang, Xiangfeng Xu, Bangyan Li, Changbo Wang 等

Reactive dance generation (RDG), the task of generating a dance conditioned on a lead dancer's motion, holds significant promise for enhancing human-robot interaction and immersive digital entertainment. Despite progress in duet synchronization and motion-music alignment, two key challenges remain: generating fine-grained spatial interactions and ensuring long-term temporal coherence. In this work, we introduce $\textbf{ReactDance}$, a diffusion framework that operates on a novel hierarchical latent space to address these spatiotemporal challenges in RDG. First, for fine-grained spatial control and artistic expression, we propose Hierarchical Finite Scalar Quantization ($\textbf{HFSQ}$). This multi-scale motion representation effectively disentangles coarse body posture from high-frequency dynamics, enabling independent and detailed control over both aspects through a layered guidance mechanism. Second, to efficiently generate long sequences with high temporal coherence, we propose Blockwise Local Context ($\textbf{BLC}$), a non-autoregressive sampling strategy. Departing from slow, frame-by-frame generation, BLC partitions the sequence into blocks and synthesizes them in parallel via periodic causal masking and positional encodings. Coherence across these blocks is ensured by a dense sliding-window training approach that enriches the representation with local temporal context. Extensive experiments show that ReactDance substantially outperforms state-of-the-art methods in motion quality, long-term coherence, and sampling efficiency.

Theory · Interpretability and Visualization

Xingshu Chen, Yifeng Guo, Jin-Hong Du

Quantifying feature importance with valid statistical uncertainty is central to interpretable machine learning, yet classical model-agnostic methods often fail under feature correlation, producing unreliable attributions and compromising inference. Statistical approaches that address correlation through feature decorrelation have shown promise but remain restricted to $\ell_2$ loss, limiting their applicability across diverse machine learning tasks. We introduce Flow-Disentangled Feature Importance (FDFI), a model-agnostic framework that resolves these limitations by combining principled statistical inference with computational flexibility. FDFI leverages flow matching to learn flexible disentanglement maps that not only handle arbitrary feature distributions but also provide an interpretable pathway for understanding how importance is attributed through the data's correlation structure. The framework generalizes the decorrelation-based attribution to general differentiable loss functions, enabling statistically valid importance assessment for black-box predictors across regression and classification. We establish statistical inference theory, deriving semiparametric efficiency of FDFI estimators, which enables valid confidence intervals and hypothesis testing with Type I error control. Experiments demonstrate that FDFI achieves substantially higher statistical power than removal-based and conditional permutation approaches, while maintaining robust and interpretable attributions even under severe interdependence. These findings hold across synthetic benchmarks and a broad collection of real datasets spanning diverse domains.

Social Aspects · Fairness, Equity, Justice and Safety

Zichen TANG, Zhenheng Tang, Gaoning Pan, Buhua Liu, Xin He, Kunfeng Lai, Xiaowen Chu, Bo Li

Geo-distributed training and Federated Learning (FL) provide viable solutions to address the substantial data and computational resource needs associated with training large language models (LLMs). However, we empirically demonstrate that a single adversarial participant can significantly compromise the safety alignment of LLMs through malicious training, exposing serious security risks. We identify two existing server-side defense strategies that effectively counter naive jailbreak attacks—Task Performance Check (TPC), which filters out model updates with low downstream performance, and Malicious Output Scrutiny (MOS), which detects harmful outputs by prompting uploaded models with malicious queries. To evade both defenses, we design a trigger-based jailbreak variant that preserves downstream performance using a novel regularization method to limit the excessive model updates on jailbreak datasets. We further conceal malicious triggers by mixing the malicious dataset with pseudo-contrastive safety-aligned answers to maintain the original safety alignment. Experiments on three widely-used safety-aligned LLMs show that a single adversarial participant can implant triggers into the global model without degrading downstream performance, achieving an 80\% attack success rate (ASR) with a 7\% low detection true rate (DTR).

Deep Learning · Generative Models and Autoencoders

Aditya Shankar, Yuandou Wang, Rihan Hai, Lydia Chen

Generating tabular data under conditions is critical to applications requiring precise control over the generative process. Existing methods rely on training-time strategies that do not generalise to unseen constraints during inference, and struggle to handle conditional tasks beyond tabular imputation. While manifold theory offers a principled way to guide generation, current formulations are tied to specific inference-time objectives and are limited to continuous domains. We extend manifold theory to tabular data and expand its scope to handle diverse inference-time objectives. On this foundation, we introduce Harpoon, a tabular diffusion method that guides unconstrained samples along the manifold geometry to satisfy diverse tabular conditions at inference. We validate our theoretical contributions empirically on tasks such as imputation and enforcing inequality constraints, demonstrating Harpoon's strong performance across diverse datasets and the practical benefits of manifold-aware guidance for tabular data. Code URL: https://github.com/adis98/Harpoon

Computer Vision · Vision Models & Multimodal

Kun Cheng, Qibing Qin, Wenfeng Zhang, Lei Huang, Jie Nie

Hard negative generation (HNG) provides valuable signals for deep learning, but existing methods mostly rely on local correlations while neglecting the global geometry of the embedding space. This limitation often leads to weak discrimination, particularly in cross-modal hashing, which obtains compact binary codes. We propose Deep Global-sense Hard-negative Discriminative Generation Hashing (DGHDGH), a framework that constructs a structured graph with dual-iterative message propagation to capture global correlations, and then performs difficulty-adaptive, channel-wise interpolation to synthesize semantically consistent hard negatives aligned with global Hamming geometry. Our approach yields more informative negatives, sharpens semantic boundaries in the Hamming co-space, and substantially enhances cross-modal retrieval. Experiments on multiple benchmarks consistently demonstrate improvements in retrieval accuracy, verifying the discriminative advantages brought by global-sense HNG in cross-modal hashing.

Optimization · Optimization and Learning under Uncertainty

Rafael Oliveira, Dan Steinberg, Edwin Bonilla

We present a general strategy for turning generative models into candidate solution samplers for batch Bayesian optimization (BO). The use of generative models for BO enables large batch scaling as generative sampling, optimization of non-continuous design spaces, and high-dimensional and combinatorial design. Inspired by the success of direct preference optimization (DPO), we show that one can train a generative model with noisy, simple utility values directly computed from observations to then form proposal distributions whose densities are proportional to the expected utility, i.e., BO's acquisition function values. Furthermore, this approach is generalizable beyond preference-based feedback to general types of reward signals and loss functions. This perspective avoids the construction of surrogate (regression or classification) models, common in previous methods that have used generative models for black-box optimization. Theoretically, we show that the generative models within the BO process follow a sequence of distributions which asymptotically approximate an optimal target under certain conditions. We also evaluate the performance through experiments on challenging optimization problems involving large batches in high dimensions.

Reinforcement Learning · Multi-agent

Huining Yuan, Zelai Xu, Zheyue Tan, Xiangmin Yi, Mo Guang, Kaiwen Long, Haojia Hui, BOXUN LI, Xinlei Chen, Bo Zhao 等

Developing Large Language Models (LLMs) to cooperate and compete effectively within multi-agent systems (MASs) is a critical step towards more advanced intelligence. While reinforcement learning (RL) has proven effective for enhancing reasoning in single-agent tasks, its extension to multi-turn, multi-agent scenarios remains underexplored due to the challenges of long-horizon credit assignment and agent-specific advantage estimation. To address these challenges, we introduce **MARSHAL**, an end-to-end RL framework that incentivizes **M**ulti-**A**gent **R**easoning through **S**elf-play wit**H** str**A**tegic **L**LMs in both cooperative and competitive games. MARSHAL features a turn-level advantage estimator that aligns learning signals with each interaction for credit assignment, and an agent-specific advantage normalization to stabilize multi-agent training. By learning with self-play across cooperative and competitive games, MARSHAL agents trained from Qwen3-4B develop strong strategic abilities, with up to $28.7$\% performance improvements in held-out games. More importantly, the capability acquired through self-play generalizes beyond games, yielding consistent performance gains of MASs in reasoning benchmarks. When integrated into leading MASs, our MARSHAL agent achieves significant zero-shot performance gains of up to $10.0$\% on AIME, $7.6$\% on GPQA-Diamond, and $3.5$\% on average across all benchmarks. These results establish self-play in strategic games as a powerful approach for developing generalizable multi-agent reasoning capabilities in LLMs.

Computer Vision · Everything Else

Hugo Baudchon, Arthur Ouaknine, Martin Weiss, Mélisande Teng, Thomas Walla, Antoine Caron-Guay, Christopher Pal, Etienne Laliberté

Detecting individual tree crowns in tropical forests is essential to study these complex and crucial ecosystems impacted by human interventions and climate change. However, tropical crowns vary widely in size, structure, and pattern and are largely overlapping and intertwined, requiring advanced remote sensing methods applied to high-resolution imagery. Despite growing interest in tropical tree crown detection, annotated datasets remain scarce, hindering robust model development. We introduce SelvaBox, the largest open‑access dataset for tropical tree crown detection in high-resolution drone imagery. It spans three countries and contains more than $83\,000$ manually labeled crowns -- an order of magnitude larger than all previous tropical forest datasets combined. Extensive benchmarks on SelvaBox reveal two key findings: 1) higher-resolution inputs consistently boost detection accuracy; and 2) models trained exclusively on SelvaBox achieve competitive zero-shot detection performance on unseen tropical tree crown datasets, matching or exceeding competing methods. Furthermore, jointly training on SelvaBox and three other datasets at resolutions from 3 to 10 cm per pixel within a unified multi-resolution pipeline yields a detector ranking first or second across all evaluated datasets. Our dataset, code, and pre-trained weights are made public.

General Machine Learning · Representation Learning

Yixuan Tang, Yi Yang

Domain-specific embedding models have shown promise for applications that require specialized semantic understanding, such as coding agents and financial retrieval systems, often achieving higher performance gains than general models. However, state-of-the-art embedding models are typically based on LLMs, which contain billions of parameters, making deployment challenging in resource-constrained environments. Model compression through pruning offers a promising solution, but existing pruning methods treat all parameters uniformly, failing to distinguish between general semantic representations and domain-specific patterns, leading to suboptimal pruning decisions. Thus, we propose GAPrune, a pruning framework that addresses this challenge by considering both domain importance and preserving general linguistic foundation. Our method uses Fisher Information to measure importance and general-domain gradient alignment to assess parameter behavior, then combines these signals using our Domain Alignment Importance (DAI) scoring. Lower DAI scores indicate that the parameter is either less important for the domain task or creates conflicts between domain and general objectives. Experiments on two domain benchmarks, FinMTEB and ChemTEB, show that GAPrune maintains performance within 2.5\% of dense models in one-shot pruning at 50\% sparsity, while outperforming all baselines. With retraining in 100 steps, GAPrune achieves +4.51\% improvement on FinMTEB and +1.73\% on ChemTEB, demonstrating that our pruning strategy not only preserves but enhances domain-specific capabilities. Our findings demonstrate that principled pruning strategies can achieve model compression and enhanced domain specialization, providing the research community with a new approach for development.

Deep Learning · Algorithms

Wenbing Li, Zikai Song, Hang Zhou, Junqing Yu, Yunyao Zhang, Wei Yang

Recent attempts to combine low-rank adaptation (LoRA) with mixture-of-experts (MoE) for multi-task adaptation of Large Language Models (LLMs) often replace whole attention/FFN layers with switch experts or append parallel expert branches, undermining parameter efficiency and limiting task specialization. We introduce **LoRA-Mixer**, a modular MoE framework that routes task-specific LoRA experts into the core projection matrices of the attention module (input/output linear layers), rather than primarily targeting FFN blocks. The design delivers fine-grained token-level specialization by fully exploiting the attention mechanism, while remaining drop-in compatible with Transformers and state-space models (SSMs) as the linear projection layers are ubiquitous. To train robust routers from limited data while promoting stable, selective decisions and high expert reuse, **LoRA-Mixer** employs an adaptive **Routing Specialization Loss (RSL)** that jointly enforces global load balance and input-aware specialization via an entropy-shaping objective. The framework supports two regimes: (i) joint optimization of adapters and router with a differentiable hard–soft top-k routing scheme, and (ii) plug-and-play routing over frozen, pre-trained LoRA modules sourced from public repositories. Across 15 benchmarks—including MedQA, GSM8K, HumanEval, and GLUE—RSL-optimized LoRA-Mixer outperforms state-of-the-art routing and LoRA-MoE baselines while using 48% of their trainable parameters, with gains of +3.79%, +2.90%, and +3.95% on GSM8K, CoLA, and ARC-C, respectively. Cross-model transfer and adapter reuse experiments further demonstrate the approach’s versatility and data efficiency.

Deep Learning · Algorithms

Sumanth Varambally, Thomas Voice, Yanchao Sun, Zhifeng Chen, Rose Yu, Ke Ye

Large Language Models (LLMs) demonstrate impressive mathematical reasoning abilities, but their solutions frequently contain errors that cannot be automatically checked. Formal theorem proving systems such as Lean 4 offer automated verification with complete accuracy, motivating recent efforts to build specialized prover LLMs that generate verifiable proofs in formal languages. However, a significant gap remains: current prover LLMs solve substantially fewer problems than general-purpose LLMs operating in natural language. We introduce Hilbert, an agentic framework that bridges this gap by combining the complementary strengths of informal reasoning and formal verification. Our system orchestrates four components: an informal LLM that excels at mathematical reasoning, a specialized prover LLM optimized for Lean 4 tactics, a formal verifier, and a semantic theorem retriever. Given a problem that the prover is unable to solve, Hilbert employs recursive decomposition to split the problem into subgoals that it solves with the prover or reasoner LLM. It leverages verifier feedback to refine incorrect proofs as necessary. Experimental results demonstrate that Hilbert, substantially outperforms existing approaches on key benchmarks, achieving 99.2\% on miniF2F, 6.6\% points above the best publicly available method. Hilbert achieves the \textbf{strongest known result} from a publicly available model on PutnamBench. It solves 462/660 problems (70.0\%), outperforming proprietary approaches like SeedProver (50.4\%) and achieving a 422\% improvement over the best publicly available baseline. Thus, Hilbert effectively narrows the gap between informal reasoning and formal proof generation. Code is available at ~\url{https://github.com/Rose-STL-Lab/ml-hilbert}.

Deep Learning · Generative Models and Autoencoders

Yuxiao Lu, Lin Xu, yang sun, Wenjun Li, Jie Shi

Large language models (LLMs) aligned for safety often suffer from over-refusal—the tendency to reject seemingly toxic or benign prompts by misclassifying them as toxic. This behavior undermines models' helpfulness and restricts usability in sensitive or nuanced contexts. While prior work has proposed mitigation strategies such as data augmentation and activation steering, these approaches often face a trade-off: reducing over-refusal typically degrades the model’s ability to reject genuinely harmful content. We argue that this issue arises from the ambiguous influence of toxic and seemingly toxic prompts on the model’s learning dynamics. To address it, we introduce a preceding alignment stage, DCR: $\textbf{D}$iscernment via $\textbf{C}$ontrastive $\textbf{R}$efinement. Both theoretically and empirically, we demonstrate that contrastive refinement improves an LLM’s capacity to distinguish truly toxic prompts from superficially toxic ones. Evaluation across diverse benchmarks shows that our method effectively reduces over-refusal while preserving the safety benefits of alignment. Importantly, it achieves this with minimal degradation of general capabilities, offering a more principled and robust direction for safety alignment.

Theory · Learning Theory

Stanley Wei, Alex Damian, Jason Lee

Significant recent work has studied the ability of gradient descent to recover a hidden planted direction $\theta^\star \in S^{d-1}$ in different high-dimensional settings, including tensor PCA and single-index models. The key quantity that governs the ability of gradient descent to traverse these landscapes is the information exponent $k^\star$ (Ben Arous et al., (2021)), which corresponds to the order of the saddle at initialization in the population landscape. Ben Arous et al., (2021) showed that $n \gtrsim d^{\max(1, k^\star-1)}$ samples were necessary and sufficient for online SGD to recover $\theta^\star$, and Ben Arous et al., (2020) proved a similar lower bound for Langevin dynamics. More recently, Damian et al., (2023) showed it was possible to circumvent these lower bounds by running gradient descent on a smoothed landscape, and that this algorithm succeeds with $n \gtrsim d^{\max(1, k^\star/2)}$ samples, which is optimal in the worst case. This raises the question of whether it is possible to achieve the same rate without explicit smoothing. In this paper, we show that Langevin dynamics can succeed with $n \gtrsim d^{ k^\star/2 }$ samples if one considers the average iterate, rather than the last iterate. The key idea is that the combination of noise-injection and iterate averaging is able to emulate the effect of landscape smoothing. We apply this result to both the tensor PCA and single-index model settings. Finally, we conjecture that minibatch SGD can also achieve the same rate without adding any additional noise.

Khurram Yamin, Shantanu Gupta, Gaurav Ghosal, Zachary Lipton, Bryan Wilder

The ability to robustly identify causal relationships is essential for autonomous decision-making and adaptation to novel scenarios. However, accurately inferring causal structure requires integrating both world knowledge and abstract logical reasoning. In this work, we investigate the interaction between these two capabilities through the representative task of causal reasoning over narratives. Through controlled synthetic, semi-synthetic and real-world experiments, we find that state-of-the-art large language models (LLMs) often rely on superficial heuristics—for example, inferring causality from event order or recalling memorized world knowledge without attending to context. Furthermore, we show that simple reformulations of the task can elicit more robust reasoning behavior. Our evaluation spans a range of causal structures, from linear chains to complex graphs involving colliders and forks. These findings uncover systematic patterns in how LLMs perform causal reasoning and lay the groundwork for developing methods that better align LLM behavior with principled causal inference.

Deep Learning · Attention Mechanisms

Xu Liu, Guikun Chen, Wenguan Wang

Large language models (LLMs) suffer from hallucination and context forgetting. Prior studies suggest that attention drift is a primary cause of these problems, where LLMs' focus shifts towards newly generated tokens and away from the initial input context. To address this, we make use of a related, intrinsic characteristic of LLMs: attention sink – the tendency to consistently allocate high attention to the very first token (i.e., ⟨BOS⟩) of a sequence. Concretely, we propose an advanced context anchoring method, SINKTRACK, which treats ⟨BOS⟩ as an information anchor and injects key contextual features (such as those derived from the input image or instruction) into its representation. As such, LLM remains anchored to the initial input context throughout the entire generation process. SINKTRACK is training-free, plug-and-play, and introduces negligible inference overhead. Experiments demonstrate that SINKTRACK mitigates hallucination and context forgetting across both textual (e.g., +18.9% on QuAC with Llama3.1-8B-Instruct) and multi-modal (e.g., +23.0% on M3CoT with Qwen2.5-VL-7B-Instruct) tasks. Its consistent gains across different architectures and scales underscore the robustness and generalizability. We also analyze its underlying working mechanism from the perspective of information delivery. Our source code is available at anonymous GitHub.