arXiv subjects

Yu Li

Publications and source records attributed to Yu Li.

At least 19 recordsLinked to original sources

arXiv2026

LRMoE: Demand-Side Load Re-Calibration for Efficient MoE Training on Memory-Constrained Nodes

Sparse Mixture-of-Experts (MoE) models offer scalable capacity for Web applications, but uneven routing creates computation and communication bottlenecks. Existing placement and replication methods redistribute this workload, yet additional expert copies compete with token-related buffers for limited GPU memory. We present LRMoE, a memory-aware MoE training system that jointly recalibrates backward demand and balances the retained workload. During forward execution, it plans retention quotas and expert placement under resource constraints while preserving all routed computations. During backpropagation, a load recalibration metric estimates each token-expert pair's contribution to the expert update and selects pairs within these quotas. A capacity-aware routing algorithm distributes retained pairs among eligible instances of their original experts. Physical compaction then forms smaller communication payloads and variable-size grouped GEMMs, reducing backward traffic, computation, and execution-buffer demand. Experiments on Qwen3-30B-A3B and GLM4.5-106B-A12B show that LRMoE reduces cumulative MoE-layer training time to fixed-loss targets by 21.2-44.8% relative to SYMI.

cs.AI↗

arXiv2026

UniRec: Cross-stage Multi-Task Fusion with Preference Optimization for Cascaded Recommender Systems

Industrial recommender systems use cascaded stages with different candidate spaces, data distributions, and latency constraints. Each stage include multi-objective prediction and multi-task fusion (MTF) modules. Existing MTF methods typically design and optimize the fusion module independently for each stage, which can lead to inconsistent preferences across stages. Although cross-stage coordination methods have been proposed, they mainly transfer supervision to upstream scoring models rather than jointly optimizing the fusion modules. We propose UniRec, a Unified Cross-stage Recommendation Fusion framework that jointly optimizes pre-ranking and ranking fusion agents. UniRec employs a coupled dual-agent architecture with partially shared embeddings and stage-specific fusion functions. It further introduces Dual-Axis Preference Optimization to combine stage-local preferences with downstream-to-upstream preference transfer. To enable stable and efficient cross-stage training, we address two practical challenges: reward-induced distribution drift and the computational cost of pairwise preference learning. Extensive offline and online experiments on a production short-video system demonstrate consistent gains over single-stage fusion and cross-stage coordination baselines. In a nine-week online A/B test, UniRec improves total watch time by $0.675\%$ and app usage time by $0.616\%$, and active users by $0.189\%$, and has been fully deployed in production.

cs.IR↗

arXiv2026

Beyond Perturbation Magnitude: Direction-Dependent Responses in Multimodal Geometric Representations

Geometric alignment scores based on Gram determinants provide a compact way to model higher-order consistency among modalities, yet how such scores respond to modality degradation is poorly understood. This paper asks whether the response of a multimodal geometric score is determined primarily by the magnitude of the perturbation-induced displacement. Using frozen cohorts from MSR-VTT (N=878) and DiDeMo (N=980), we apply controlled video blur and audio noise and analyze the response in the relational geometry on which the score is defined. Displacement magnitude explains at most 15% of the out-of-sample variance in the absolute response, and magnitude-matched pairs respond systematically differently, so scalar magnitude does not organize the response. The closed-form first-order expansion of the Gramian volume yields the Directional Geometric Response (DGR): the projection of the displacement onto the local volume gradient, which jointly captures the clean operating point, displacement magnitude, and displacement direction. The absolute first-order DGR term explains the observed response with out-of-sample R^2 of 0.838-0.969, matched-magnitude ranking accuracies of 0.864-0.963, and response-sign accuracies of 0.909-0.989, whereas the tested direction-free alternatives remain weak or unstable under the corresponding evaluation protocols. A pre-specified gain-normalization candidate, V/(g_V+eps), fails its predictability and clean-order gates. DGR uses the observed degraded-state displacement and is therefore an explanatory quantity, not a deployment-time predictor: geometric response depends on where the representation operates, how far degradation moves the relational geometry, and in which direction it moves.

cs.CV↗

arXiv2026

MTA-Zero: Self-Improving Multimodal Deep Search via Online QA-Solver Co-training

Multimodal large language models (MLLMs) have become strong visual reasoners, yet remain limited in open-world, multi-step search that requires grounding images, selecting tools, and integrating web evidence. Existing approaches typically construct and filter fixed offline corpora whose question distributions cannot adapt as the target deep-search agent improves. Building on MTA-Agent's image-grounded QA construction, we present MTA-Zero, an online multimodal QA-Solver co-training framework for self-improving multimodal deep-search agents that requires no precomputed multi-hop corpus. Starting from image-grounded VQA seed contexts, a trainable QA Generator retrieves evidence and constructs verified multi-hop questions online. A separate deep-search Solver then selects tools, gathers evidence, and answers the generated questions. To focus generation on the Solver's learning frontier, we derive a Goldilocks reward for the QA Generator from the Solver's group-wise rollout success rate: questions solved by all rollouts or by none receive zero reward, whereas questions near the current frontier are favored. We optimize the two policies independently with GRPO, without sharing parameters or gradients. Across six multimodal deep-search benchmarks, MTA-Zero improves a Qwen3-VL-8B Solver from 36.91% to 51.58% mean accuracy and achieves performance comparable to GPT-5 under the same agent workflow. The same framework also improves a newer Qwen3.6-27B Solver by 10.82 points. These results show that verified online question generation provides a policy-specific curriculum for continually improving multimodal deep-search agents.

cs.CL↗

arXiv2026

Recurrent Self-Improvement: Dynamic Cross-Loop On-Policy Distillation for Looped Language Models

Looped Language Models (LoopLMs) offer a parameter efficient approach to scaling reasoning by reusing shared parameters across recurrent computation steps. Despite their promise, effective post-training of LoopLMs remains challenging. Existing approaches either provide reward based supervision that is sparse or costly to extend across loops, or rely on external teachers or privileged information, leading to limited teacher availability or teacher-student context mismatch. To address these limitations, we introduce LoopOPD, a cross-loop on-policy distillation framework that uses additional recurrent computation within a LoopLM as its own source of supervision. LoopOPD uses a frozen terminal loop policy as a compute privileged teacher for an intermediate loop student on student generated rollouts, providing dense supervision without an external teacher or privileged information. We further propose Dynamic LoopOPD (D-LoopOPD), which continually refreshes the terminal loop teacher as the shared model parameters are updated, enabling recurrent self-improvement. We characterize how distillation updates propagate across loop depths and derive sufficient conditions under which a single update yields simultaneous local improvement at both loop depths. Experiments on Ouro-Thinking models show that LoopOPD improves mathematical reasoning, while D-LoopOPD yields further gains through dynamic teacher updates. Despite being trained only on mathematical data, the resulting models also improve on general reasoning and code generation benchmarks, demonstrating that recurrent computation can serve as an effective source of supervision for LoopLMs. Our code and model checkpoints will be released upon acceptance.

cs.LG↗

arXiv2026

Towards Effective and Transferable Physical Camouflage against Multi-View BEV-based 3D Perception in Autonomous Driving

Modern autonomous driving systems widely adopt multi-view Bird's-Eye-View (BEV) based 3D perception models due to their superior performance. Despite their success, the robustness of these models against adversarial camouflage attacks remains largely unexplored. Existing camouflage attacks primarily target single-view 2D detectors, limiting their effectiveness in multi-view 3D perception settings. To address this gap, we propose BEVCA, a novel adversarial camouflage generation framework tailored for multi-view BEV-based 3D perception. Our framework integrates a BEV-feature-based adversarial loss with a multi-view neural rendering module, enabling effective and transferable camouflage attacks across different models and tasks. Extensive experiments in both digital and physical settings demonstrate that BEVCA outperforms state-of-the-art baselines and exhibits strong robustness under real-world conditions. Our code is available at https://github.com/WAVE-AT-IDEA/BEVCA.

cs.CV↗

arXiv2026

Traceable Quantum Secret Sharing

Quantum secret sharing protects distributed quantum information against unauthorized reconstruction, yet existing schemes provide no intrinsic accountability when shares are leaked or forged. This missing capability is particularly relevant in realistic distributed settings, where compromised participants may either expose valid shares or maliciously interfere with reconstruction. Here we establish a framework for traceable quantum secret sharing under these two threat models. We introduce leakage-verifiable and traceable quantum secret sharing (LVT-QSS), which attributes leaked shares through an explicit leakage verification mechanism, and malicious-traceable and accountable quantum secret sharing (MTA-QSS), which identifies participants who submit forged shares during reconstruction. We formulate security definitions for both settings and prove that the proposed protocols satisfy the corresponding guarantees against the relevant adversaries. Numerical simulations under per-share depolarizing noise further quantify the robustness of the proposed schemes and identify verification as the dominant noise sensitive stage. These results incorporate accountability into quantum secret sharing and provide a foundation for traceable distributed quantum cryptography.

quant-ph↗

arXiv2026

Parametrization of ancient Ricci flows with smooth compact tangent shrinkers

We study ancient Ricci flows whose tangent flow at backward infinity is a smooth compact Ricci shrinker. After normalization, every nontrivial solution either converges exponentially or has polynomial decay governed by neutral directions. We recover the finite-dimensional parametrization of the exponentially convergent solutions and establish finite-order asymptotic expansions for all slowly convergent solutions. For principal slow solutions, we obtain a locally bijective finite-dimensional parametrization compatible with diffeomorphisms, parabolic rescaling, and time translation.

math.DG↗

arXiv2026

Pushing Biomolecular Utility-Diversity Frontiers with Supergroup Relative Policy Optimization

Biomolecular generators are often adapted with reward feedback to improve task-specific utility, but pushing utility alone can concentrate generation on a narrow family of candidates. Maintaining diversity is difficult because sample diversity is a set-level property. We introduce Supergroup Relative Policy Optimization (SGRPO), a flexible GRPO-style framework that directly constructs rewards from set-level diversity. For each condition, SGRPO samples a supergroup of candidate sets, compares their diversity under the same condition, and redistributes the group diversity reward to individual rollouts through leave-one-out diversity contributions before combining it with rollout-level utility. This design decouples SGRPO from a particular generator, utility reward, or diversity metric, and allows instantiation with different GRPO-style approaches. We evaluate SGRPO on three main settings: de novo small-molecule design, pocket-based small-molecule design, and de novo protein design, instantiating it with both GRPO and Coupled-GRPO across autoregressive and discrete diffusion generators. Across decoding sweeps, SGRPO expands the utility-diversity Pareto frontier and achieves the best frontier-level metrics relative to baselines. Our analyses further show that direct set-level diversity rewards remain effective with small groups and help preserve broader generation-distribution coverage during post-training. The code is available at https://github.com/IDEA-XL/SGRPO.

cs.CE↗

arXiv2026

From Prediction to Action: Population Health World Model for Digital Medicine

Digital medicine now captures data that traces individual health trajectories and reveals population health dynamics through electronic health records, claims, remote monitoring, wearable devices, mobile devices, social sensing data, medical images, etc. But much of its modeling still centers on task-specific, outcome-specific, and prediction-specific estimation. The next frontier is not simply better prediction, but action in an evolving health world shaped by multiple decision-makers, heterogeneous populations, resource constraints, and feedback in care delivery. This is why digital medicine now needs world models: not only to estimate what will happen, but to represent how actions change what can happen next -- whom to prioritize, how to allocate scarce resources, how to adapt care across subgroups, and how interventions reshape future health trajectories and population dynamics. We call such a model population health world model (PHWM): a more foundational representation of an evolving health world in which health states, individual trajectories, population health dynamics, interventions, constraints, and feedback are intrinsically coupled. Building a PHWM endows three unmet structural gaps in current digital medicine: dynamic health-state modeling, population-level intervention intelligence, and validation, equity, and governance for action-oriented use. We argue that addressing these gaps requires three minimal capabilities: representation learning under action-dependent observation, multi-scale intervention intelligence, and validation with explicit decision boundaries, while highlighting the importance of social, contextual, and public-health sensing data for action-oriented digital medicine.

cs.AI↗

arXiv2026

Counterexamples to $R$-superlinear convergence of cyclic steepest descent

The cyclic steepest descent method computes the exact steepest descent stepsize at the start of each cycle and reuses it for a fixed number $m$ of iterations. For strictly convex quadratics in $n$ variables, it has been suggested that the method converges $R$-superlinearly whenever $m\ge(n+1)/2$. We show that this cannot hold for every initial point, already at the threshold cycle length. For $n=m=2$ we exhibit an initial point from which the method follows the classical zigzag trajectory of steepest descent: the stepsize never changes and the error is halved at every iteration. The same mechanism yields linearly convergent trajectories for every Hessian with at least two distinct eigenvalues and every cycle length. For $n=3$ and $m=2$ we construct a second, genuinely three-dimensional counterexample: all components of the gradient remain nonzero, consecutive cycles use two different stepsizes, and every four iterations reduce the gradient exactly by the factor $1/49$. Both trajectories are given in closed form. Consequently, any $R$-superlinear convergence result for this method at the threshold must exclude a set of initial points, as recent almost-everywhere results do.

math.OC↗

arXiv2026

Convergence of cyclic steepest descent for quadratics with two distinct eigenvalues

The cyclic steepest descent method computes the exact steepest descent stepsize at the start of each cycle and reuses it for a fixed number $m\ge2$ of iterations. We study this method for strictly convex quadratics whose Hessian has exactly two distinct eigenvalues, of arbitrary multiplicities, and determine its behavior from every initial point. At the minimizer no update is needed. For nonoptimal initializations, a gradient in one eigenspace gives one-step termination. If both eigenspace components are nonzero and have equal norm, the stepsize never changes and the error contracts by the same constant factor at every iteration. In the remaining case, both components are nonzero and unequal in norm, and the method converges doubly exponentially and hence $R$-superlinearly. Thus $R$-superlinear convergence fails exactly on an explicit cone of initial points of measure zero. The proof reduces the method to an exact scalar recurrence for the ratio of the two eigenspace components at the start of each cycle. We also determine the leading term of the logarithm of the error at every iteration within a cycle, which identifies the exact exponent of the doubly exponential decay and shows that this convergence is not $Q$-superlinear.

math.OC↗

arXiv2026

Training Reasoning Models to Generate Diverse Scientific Ideas

One of the first steps of a research pipeline is generating ideas. Large Language Models (LLMs) are increasingly used to automate this step, but most of this work optimizes the quality of individual ideas, while overlooking diversity. In our work, we study document-grounded ideation, focusing on increasing the diversity of LLM generated ideas and ensuring they are high quality. Given prior papers, a model must propose a set of ideas that are each high-quality and clearly distinct from one another, while also being grounded in the input documents. To optimize for this, we propose training with a set-level reinforcement learning objective. In our setup, in a single generation, the model writes K structured idea cards grounded in the source papers, and the set is scored jointly, where each idea's quality is weighted by its distance from the ideas written before it. This teaches the model to steer each new idea away from its predecessors. We show that our trained 8B model achieves higher creative utility (a measure combining quality and diversity) than Claude Opus 4.8 and GPT-6 Astra on our contamination-free benchmark. In a pilot human study with four experts, our sets are judged as promising as Claude Opus 4.8's and contain marginally more unique ideas. When a coding agent runs proof-of-concept experiments on the ideas, our model's success@k is comparable to Claude Opus 4.8's, so we see that the added diversity does not cost any executability.

cs.CL↗

arXiv2026

EgoOmniWorld: Unifying Egocentric Understanding via Ego-Action-World Interactions

Egocentric video understanding spans diverse capabilities, modalities, and evaluation formats, yet these are often studied in isolation despite their tight coupling in real-world interaction. We propose the Ego-Action-World interaction loop as a unifying principle: the wearer perceives the world, forms goals, acts, changes the world, and uses feedback to guide subsequent behavior. Based on this view, we introduce EgoOmniWorld, a benchmark organized around four complementary interaction structures: Node captures ego, action, and world states; Edge captures directed relations among them; Chain captures linked dependencies; and Loop captures feedback-closed interactions. We develop an interaction-first, evidence-grounded data pipeline that extracts structured interactions from egocentric videos, grounds them in visual and auditory evidence, and generates questions according to their interaction structure. Required modalities and single- or multi-turn formats follow naturally from the supporting evidence and dependencies. EgoOmniWorld contains 24,166 QA groups, including 20,284 training and 3,882 test samples. Experiments on representative open- and closed-source multimodal large language models reveal substantial and uneven limitations across interaction structures. Models trained on EgoOmniWorld improve in-domain interaction understanding and generalize to existing egocentric benchmarks. EgoOmniWorld provides a unified foundation for evaluating and advancing egocentric video understanding. Data, code, and models will be released.

cs.CV↗

arXiv2026

Covering Radius and Deep Holes of Hyperderivative Reed-Solomon Codes under the NRT Metric

The deep hole problem is a fundamental problem in coding theory, and it has many important applications in code constructions and cryptography. The deep hole problem of Reed-Solomon codes has gained a lot of attention. As a generalization of Reed-Solomon codes under Hamming metric, we study the covering radius and deep holes of Hyperderivative Reed-Solomon (HRS) codes under the Niederreiter-Rosenbloom-Tsfasman (NRT) metric. We determine the NRT covering radius of HRS code and characterize its NRT deep holes via Maximum Distance Separable (MDS) extensions and polynomial remainder conditions. We further relate the enumeration of NRT deep holes to the characteristic polynomial of an associated hyperplane arrangement. We also provide three classes of explicit NRT deep-hole constructions, propose a Cheng-Murray type conjecture for HRS codes under the NRT metric and obtain a conditional low-layer description. Using these characterizations, we derive exact enumeration formulas in several special cases.

cs.IT↗

arXiv2026

LLM-Augmented Bayesian Optimization with Adaptive Policies for Chiplet Design Space Exploration

Chiplet technology is a key approach for extending Moore's Law. However, its design process is constrained by two primary challenges: the scarcity of early-stage platforms for rapidly evaluating chiplet-based heterogeneous systems, and the complex relationships between design parameters and optimization metrics. To address these challenges, we propose a synergistic framework that integrates a Large Language Model (LLM) with Bayesian Optimization (BO), enabling architectural-level simulation and automated optimization for the 2.5D/3D CPU-GPU heterogeneous chiplets. We extend the open-source chiplet simulator to evaluate 3D designs by supporting graph topologies. Further, to enhance the exploration efficiency of BO, we introduce an LLM that generates high-quality policies to guide the optimization process, enabled by our proposed adaptive directional policy reasoning (ADPR) and an exploration gain-aware fine-tuning method. Finally, we use the policy-guided BO to perform an automated and efficient exploration of the complex chiplet design space. Experiments on 2.5D and 3D chiplet reveal that our framework improves design space exploration efficiency.

cs.LG↗

arXiv2026

Learning from All-Failed Groups via First-Visit Coverage in Long-Horizon Agentic Tasks

Group-based policy optimization has been increasingly used to train large language model (LLM) agents from sparse outcome rewards by comparing trajectories or steps within a group. However, on difficult long-horizon agentic tasks, we found this comparison suffers from a sampling imbalance: repeated or low-effect actions dominate the high-probability region of the policy, leaving useful state-changing actions under-sampled. This imbalance produces many all-failed rollout groups, where the relative advantages within such a group provide no direction for updating the policy. Consequently, the sampling imbalance and the all-failed groups collapse into a self-perpetuating credit trap: failure-dominated sampling yields no learning signal, allowing repeated low-effect actions to persist. To break this trap, we propose Coverage-conditioned Group Policy Optimization (CovGPO), which uses first-visit observation coverage only when all samples in a group receive zero outcome reward. Specifically, within such groups, CovGPO assigns higher relative advantages to state-exploring actions without corrupting the original reward signal, as reaching new observations is a prerequisite for task success. Experiments on two challenging agentic benchmarks, ALFWorld and WebShop, show that CovGPO consistently improves over group-based baselines, with particularly large gains on hard tasks.

cs.LG↗

arXiv2026

Scaling RNA Foundation Models through Iterative Refinement

Biological function in RNA is encoded across scales, from local sequence motifs to long-range structural dependencies. Conventional RNA foundation models couple parameter capacity and inference depth within a fixed architecture, prescribing a fixed computation graph for each inference. Here we present Eco-RNA, an 80M-parameter RNA foundation model that makes inference depth adjustable through learned iterative refinement. Its shared, input-conditioned Transformer core refines sequence representations through repeated application. The iteration count is selected after training and controls effective depth at fixed model parameters. We train across recurrent depths with masked reconstruction, state regularization, and a one-sided loss that anchors deeper reconstruction to a shallow reference. On the 70-assay RNAGym fitness benchmark, increasing refinement from one to four iterations raises aggregate absolute Spearman correlation from 0.242 to 0.286, with the gain retained through eight iterations. At four iterations, Eco-RNA surpasses the current SOTA Evo 2 40B on all three aggregate fitness metrics despite 500x fewer parameters. Its iteration configurations advance the compute-performance Pareto frontier among the evaluated models. It also outperforms the evaluated baselines in non-coding RNA family classification, RNA secondary-structure prediction, and chemical probing profile prediction. These results establish internal inference computation as an effective scaling dimension for RNA foundation models.

cs.LG↗