arXiv subjects

Zhuoran Li

Publications and source records attributed to Zhuoran Li.

At least 19 recordsLinked to original sources

arXiv2026

HBF Sucks? A Full-Stack Characterization of High-Bandwidth Flash for KV-Centric LLM Serving

A faster storage device should make serving faster. We find the opposite. High-Bandwidth Flash (HBF) stacks NAND behind a wide, package-local interface, promising flash-scale capacity with far lower read latency and higher bandwidth than an SSD. The obvious move is to keep an SSD-style Mooncake KV-offloading stack and swap in HBF underneath. We built that system and measured it: an extended TokenSim, four complete two-hour Qwen-Bailian production traces, five dense and mixture-of-experts models, and H100/B200 profiles. The upgrade backfires. Average end-to-end latency rises 2--5.5$\times$ and maximum SLO goodput falls 1.1--2.7$\times$ across H100 and B200, so the faster device yields a slower system. A cost-benefit model explains the paradox: a faster far tier pays off only when read I/O is the bottleneck, reads outweigh writes, and delivered bandwidth is sustainable. Transient KV violates all three at once. Buying flash through the package costs GPU near-tier capacity and bandwidth, while HBF's own read/write latency barely matters: scaling it 3.75$\times$ moves latency less than 1\%. Worse, the two-tier hierarchy keeps reuse in the near tier and hands HBF a relentless write-heavy stream. Writes outnumber reads on every trace, so a 3D-ICE model shows the stack hits its thermal limit well below peak bandwidth, and a TLC tier wears out sooner than the SSD pool it replaced. The device is fine; the drop-in deployment is not. HBF sucks as an SSD replacement for transient KV, but earns its place in LLM serving when used selectively with reuse-aware placement, write budgeting, and thermal coordination.

cs.AR↗

arXiv2026

The Router Within: Eliciting Native Skill Routing from a Frozen LLM

Skills extend an LLM agent beyond its parametric knowledge, and the gain they promise rests on picking the right one. Deployed harnesses route by preloading every skill's metadata into the context, which disperses the agent's attention and caps the library size. Retrieval pipelines move the selection out of the context, but also out of the agent's capability. We show that the frozen agent LLM already carries the routing signal in its own forward passes, and that two linear maps suffice to read it out with no skill text in the context. Gavel (Glance And Verdict from a frozen LLM) reads it in two steps. A glance projects the task's and each skill's mid-layer states through the two maps, the only parameters trained, and scores the full library against compact per-skill banks that one forward pass builds at installation. A verdict then resumes the shortlisted skills' forward passes and reads the model's own likelihood and yes/no judgment, fused with the glance as a product of experts. Trained once, Gavel transfers zero-shot to three public benchmarks and SkillTraj, our new benchmark of 372 simulated agent trajectories. On Qwen3-32B it outperforms progressive disclosure and retrieve-and-rerank pipelines that add 1.2B to 16B external parameters, by up to 13.4 points on written tasks and up to 21.9 when the need for a skill arises mid-rollout. Routing accuracy improves as the backbone does, and in a bash-agent harness the same 32B triggers the correct skill on Skill-Use more often than far larger frontier models running in Codex.

cs.LG↗

arXiv2026

Scoring, Reasoning, and Selecting the Best! Ensembling Large Language Models via a Peer-Review Process

We propose LLM-PeerReview, an unsupervised LLM Ensemble method that selects the most ideal response from multiple LLM-generated candidates for each query, harnessing the collective wisdom of multiple models with diverse strengths. LLM-PeerReview is built on a novel, peer-review-inspired framework that offers a transparent and interpretable mechanism, while remaining fully unsupervised for flexible adaptability and generalization. Specifically, it operates in three stages: For scoring, we use the emerging LLM-as-a-Judge technique to evaluate each response by reusing multiple LLMs at hand; For reasoning, we can apply a straightforward averaging strategy or a principled graphical model-based truth inference algorithm to aggregate multiple scores to produce a final score for each response; Finally, the highest-scoring response is selected as the best ensemble output. LLM-PeerReview is conceptually simple and empirically powerful. Our results across four datasets show that the two variants of the proposed approach outperform the advanced model Smoothie-Global by 6.9% and 7.3% points, cross diverse task types including factual recall QA, math reasoning, and instruction following. Notably, we also establish a carefully curated benchmark suite for LLM Ensemble, integrating 12 methods across four classic datasets and three task families, all evaluated under a rigorous and consistent protocol. We hope this repository will help researchers reproduce the LLM Ensemble baselines.

cs.CL↗

arXiv2026

TokenCake: A KV-Cache-centric Serving Framework for LLM-based Multi-Agent Applications

Large Language Models (LLMs) are increasingly deployed in complex multi-agent applications that rely on external function calls. This workload creates severe performance challenges for the KV Cache: spatial contention leads to the eviction of critical agents' caches and temporal underutilization leaves the cache of agents stalled on long-running function calls idling in GPU memory. We present TokenCake, a KV-Cache-centric serving framework that bridges this gap by co-optimizing scheduling and memory management through an agent-aware design. TokenCake's Temporal Scheduler employs an event-driven, opportunistic policy to proactively offload idle KV Caches during function calls and uses predictive uploading to hide data transfer latency. TokenCake's Spatial Scheduler uses dynamic memory partitioning, guided by a hybrid priority metric combining graph structure and runtime state, to reserve GPU memory for critical-path agents. Our evaluation on representative multi-agent benchmarks shows that TokenCake reduces end-to-end latency by over 47.06% and improves effective GPU memory utilization by up to 16.9% compared to vLLM.

cs.DC↗

arXiv2026

Optimal Skill Selection for LLM Agents with Provable Bicriteria Guarantees

Loading reusable skill documents into a bounded context window is now the primary way large language model (LLM) agents acquire task-specific capabilities, which makes skill selection a first-order determinant of task performance and token cost. Yet current agents score skills independently by semantic relevance and assemble the set by top-$k$ or greedy packing, with no quality guarantee or cost awareness on the selected set. As a result, redundant or poorly chosen skills waste scarce context tokens and can even degrade performance. We give the first model of how the selected skill set shapes execution outcomes and cast skill selection as an optimization problem: choose a skill set under a hard token budget to maximize a monotone submodular benefit minus context penalty. For this problem, we develop Best Prefix Selection (BPS), a polynomial-time algorithm, and prove, to our knowledge, the first performance guarantee for skill selection: a bicriteria $(1-1/e,1)$ approximation whose benefit coefficient is optimal in polynomial time. On a contamination-controlled BigCodeBench variant, BPS outperforms all the baselines, reaching $0.73$ measured task success versus $0.20$--$0.52$ for released skill routers, text retrievers, and the executor's own selection, on $28\%$ fewer tokens than the strongest released router.

cs.AI↗

arXiv2026

MLREF: Efficient Module Reuse for Reward Design in Reinforcement Learning via Large Language Models

Reward function design remains a bottleneck in reinforcement learning. While large language models (LLMs) have enabled automated reward generation, existing methods generate and revise reward functions as monolithic programs, making it difficult to reliably preserve and reuse effective components discovered in earlier iterations, leading to unstable performance across iterations. To address this, we propose Module Level Reward Evolution Framework (MLREF). At the core of MLREF is a module pool, a persistent repository of reusable reward components. MLREF treats the module pool as the primary optimization object: the pool evolves across iterations by accumulating successful modules, refining underperforming ones, and reusing proven components; while reward functions are constructed as linear combinations of modules drawn from this pool. To drive this evolution, MLREF integrates three mechanisms: reflection-based refinement, hybrid credit assignment, and a merge strategy with rollback, which together improve the effectiveness and robustness of reward optimization. Experiments on 17 tasks show that MLREF outperforms strong baselines by 25.2% in locomotion and 6.6% in manipulation, with more stable optimization dynamics.

cs.LG↗

arXiv2026

When Personal Memory Has No Single Answer: Evaluating LLM Agents under Irreducible Conflict

LLM agents increasingly maintain personal memory across sessions, but it can conflict. Preferences depend on context, behavior evolves, and sources can conflict. When a query lacks context, time, or source authority to interpret conflict, treating one memory as definitive converts unresolved conflict into an unjustified, overconfident action. Existing benchmarks recover one answer from conflicting evidence, overlooking whether agents recognize underdetermination, preserve alternatives, seek missing information, and choose appropriate actions. We introduce \underline{T}esting \underline{A}gents' \underline{N}avigation of \underline{G}enuine, \underline{L}atent, and \underline{E}ntangled Memory Conflicts (\textsc{TANGLE}), a benchmark for genuinely unresolvable memory conflicts. It comprises 541 instances across 40 personas and three types: Context-Partitioned Conflict (CPC), Behavior-Oscillation Conflict (BOC), and Source-Contradiction Conflict (SCC). We evaluate two tracks---an oracle track with curated memory and a pipeline track that extracts memory from multi-session dialogues---on five dimensions: conflict perception, causal reasoning, confidence calibration, clarification seeking, and memory faithfulness. Experiments reveal pipeline challenges. With curated memory, models recognize conflicts more reliably than they calibrate actions or seek targeted clarification. With end-to-end pipeline memory, extraction fails to preserve conflict-bearing relations needed for downstream reasoning. Policy comparisons show fixed rules are insufficient when actions must reflect conflict. These findings motivate Conflict-Aware Action Policy (CAAP), which adapts actions to each conflict using available evidence. \textsc{TANGLE} frames conflict handling as recognizing underdetermination, retaining conflicting evidence, and acting without forcing a definitive answer.

cs.AI↗

arXiv2026

RoboHarness: A Memory-Augmented Policy Harness for Vision-Language-Action Model Robustness via In-Context Adaptation

Despite the promise of Vision-Language-Action (VLA) models as generalist robotic controllers, their robustness against perceptual noise and environmental variations in out-of-distribution (OOD) tasks remains fundamentally limited by the absence of long-term memory, causal failure attribution, and dynamic intervention capability. To address this, we propose RoboHarness, a memory-augmented policy harness that upgrades frozen VLA policies for robust in-context adaptation without parameter fine-tuning. Specifically, RoboHarness operates through an online pipeline of contrastive Dual-Memory Retrieval-Augmented Generation (RAG), an attribution-driven vision-language orchestrator implemented with a multimodal large language model, and extensible Model Context Protocol (MCP) interventions, while an offline Memory Consolidation module continuously distills the execution traces into reliable priors. Experimental evaluations across three backbone models ($π_0$, $π_{0.5}$, and SmolVLA) on LIBERO-PRO and our proposed LIBERO-RoboHarness benchmarks demonstrate that RoboHarness achieves an average absolute success rate gain of 56.6%. This includes a significant absolute improvement of 89.1% in long-horizon task chaining. The project page and source code are available at https://github.com/LZY-1021/RoboHarness.

cs.RO↗

arXiv2026

TokenStack: A Heterogeneous HBM-PIM Architecture and Runtime for Efficient LLM Inference

Large language model (LLM) serving is now limited by the key-value (KV) cache. During decode, each new token rereads prior KV state, so attention becomes a bandwidth- and capacity-heavy memory task. HBM-PIM helps by moving attention closer to memory, but current stack organizations still waste resources. In practice, only hot KV blocks benefit from near-memory compute. Weights, activations, and cold KV mainly need dense storage and GPU-visible bandwidth. A uniform HBM-PIM stack makes all layers pay for PIM logic, while a dedicated-PIM design such as AttAcc recovers capacity but shrinks the HBM bandwidth left for GPU-side work. We propose TokenStack, a vertically heterogeneous HBM-PIM architecture for KV-centric LLM serving that leverages HBM4's logic-die substrate. TokenStack separates each stack into dense capacity layers and PIM-enabled compute layers, then uses the logic base die as a stack-local control point that manages cross-layer movement without host-side overhead. The base-die controller handles cross-layer DMA, layered address translation, attention-side gather/broadcast coordination, and inline quantization during migration. On top of this hardware, TokenStack uses topology-aware KV placement, workload-aware eviction, and bounded replication to keep hot KV near PIM compute while moving colder state to dense layers. Using production-derived traces across four models, completed multi-QPS runs show that TokenStack increases geometric-mean token throughput by 1.62x and SLO-compliant serving capacity by 1.70x over AttAcc, and reduces per-token energy by 30-47%.

cs.AR↗

arXiv2026

When Context Returns: Toward Robust Internalization in On-Policy Distillation

Recent work has shown that on-policy distillation can internalize privileged context, such as system prompts or task hints, into a student model so that the context is no longer needed at inference time. However, we identify a counterintuitive and previously unstudied phenomenon: reintroducing the original privileged context to the distilled student often degrades its performance, even on instances it already solves correctly without context. We term this phenomenon context-induced degradation and argue that robust internalization requires not only matching the teacher's context-conditioned behavior, but also remaining stable when the privileged context is reintroduced, a desirable property we call context invariance. To promote this property, we formulate a novel view-robust internalization risk and propose No-Context Anchoring (NCA), a lightweight yet effective consistency regularizer that uses the student's stop-gradient no-context output as an anchor and aligns its context-conditioned output via forward KL divergence. Across 14 configurations spanning diverse domains and model families, NCA improves context-conditioned accuracy in most settings and reduces context harm in 12 out of 14, while preserving or improving no-context performance, demonstrating greater robustness to context reintroduction.

cs.LG↗

arXiv2026

Light-Loco-Parkour: Versatile Perceptive Whole-Body Locomotion via Multi-Skill Distillation

Existing humanoid whole-body control systems still fall short of the way humans move through cluttered terrain: they either track expressive whole-body references without terrain generalization, or react to terrain online while leaving the arms, torso, and knees largely unused. We present \texttt{Light-Loco-Parkour} (LLP), an end-to-end perceptive whole-body locomotion system that closes this gap with a single deployable policy. Conditioned only on onboard depth and a velocity command, the policy decides when to walk, balance, climb, step down, or vault, with no reference input, skill label, hand-coded gate, or runtime motion graph. Compared with prior humanoid systems, LLP makes three contributions. First, it introduces a whole-body perceptive-control pipeline that extends an RL-trained, velocity-tracking locomotion policy with parkour skills learned from object-interacting motions, so the same policy tracks velocity in open terrain, executes whole-body traversal at obstacles, and resumes locomotion afterward. Second, it acquires terrain-conditioned skills from sparse seeds by expanding a single motion into dynamically feasible, terrain-paired references across obstacle geometry, rather than relying on a large motion corpus. Third, it learns autonomous skill transitions from reward, letting the policy decide when and which whole-body skill to invoke from depth and command alone, with no one-hot skill label, hand-coded state machine, or runtime motion generator. Simulation and real-world experiments show high success across both benchmarked terrains and unseen obstacle variations, and the same policy transfers zero-shot to indoor and outdoor hardware experiments. These results demonstrate autonomous perceptive whole-body locomotion on a humanoid in outdoor settings, using only onboard sensing and a single deployable policy.

cs.RO↗

arXiv2026

CDEP Agent: Connecting Meteorologically Detected Temporal Compound Events to Real-World Documentary Evidence

Compound drought-to-extreme-precipitation (CDEP) events are recognized in climate science as a growing driver of extreme impact, but whether this recognition carries over into real-world early warning and post-event documentation is unknown, so a meteorologically real CDEP event may pass with neither advance warning nor any later record. Here we present CDEP Agent, an auditable LLM-agent framework that tests this mismatch directly by linking CDEP candidates detected from meteorological reanalysis to real-world hazard and impact evidence across sources with different spatial scales, temporal resolutions, and reporting conventions. Using California as a case study, we identify 408 candidate CDEP events from ERA5 observations during 2021-2025 and evaluate each against the U.S. Drought Monitor, NOAA Storm Events, and public webpages along five dimensions: antecedent drought, extreme rainfall, local impact, hazard-impact attribution, and explicit drought-to-rainfall linkage. Only 34.3% of candidates are corroborated on both hazard components, and just 1.5% are ever explicitly linked to their antecedent drought, indicating that most meteorologically detected CDEP events go undocumented and their compound nature almost never enters the record at all. Our framework gives climate scientists a way to test physical event definitions against what actually gets documented, and gives social scientists, economists, and disaster-response agencies a provenance-linked evidence base for compound events that current warning and reporting systems largely fail to capture.

cs.AI↗

arXiv2026

SLOTH: Lightweight Detection and Localization of On-Chip Fail-Slow Failures for DNN Accelerators

Spatial DNN accelerators are essential for high-performance inference, but their performance is undermined by widespread fail-slow failures. Detecting such failures on-chip is challenging, as prior methods from distributed systems are unsuitable due to strict memory limits and their inability to track failures across the hardware topology. We present SLOTH, a lightweight, hardware-aware framework for practical on-chip fail-slow detection in DNN accelerators. SLOTH combines workload-aware instrumentation for operator-level monitoring with minimal overhead, on-the-fly trace compression to operate within kilobytes of memory, and a novel topology-aware ranking algorithm to pinpoint a failure's root cause. We evaluate SLOTH on a wide range of representative DNN workloads. The results demonstrate that SLOTH reduces the storage overhead by an average of 115.9$\times$, while achieving an average fail-slow detection accuracy from 69.68\% to 86.69\%.

cs.AR↗

arXiv2026

Same Evidence, Different Target: Decoding How Diagnostic Evidence Bears on Causal Questions from Language-Model States

The same diagnostic result can support or challenge one causal claim yet fail to address another when the claims concern different populations, outcomes, estimands, pathways, or identifying assumptions. When the evidence and target vary together, a correct answer may reflect favorable or adverse wording, lexical overlap, or a familiar diagnostic pattern rather than matching the evidence to the causal question. We introduce paired prompts that repeat the same diagnostic evidence verbatim while changing the causal target. Each prompt is labeled Favors, Challenges, Unresolved, or Wrong Target according to how the evidence bears on the causal question. A pair is recovered only when both prompts are classified correctly. Using linear readouts trained on a separate development set, we analyze the final-token hidden state from the penultimate transformer block of Qwen2.5-7B-Instruct, Qwen3-8B, and Llama-3.1-8B-Instruct. On the 49-pair primary benchmark spanning nine diagnostic families, balanced accuracy ranges from 0.654 to 0.659 and 18-21 pairs are recovered. Two independent human reviewers assigned the same label to 95 of the 98 prompts (96.9%). Across checkpoints, balanced accuracy and complete-pair recovery exceed permutation nulls that preserve development scenario groups. In Qwen2.5, full-prompt balanced accuracy exceeds both restricted inputs, with paired-bootstrap intervals for both differences above zero. Readouts trained without development examples from the evaluated diagnostic family recover 21 pairs, including at least one in each of the nine families. The hidden-state readout exceeds a linear classifier on answer-option logits and text baselines in balanced accuracy and recovered pairs. These results show that the hidden state contains linearly decodable information about whether diagnostic evidence favors, challenges, or fails to address the causal target.

cs.CL↗

arXiv2026

Beyond Squared Error: Exploring Loss Design for Enhanced Training of Generative Flow Networks

Generative Flow Networks (GFlowNets) are a novel class of generative models designed to sample from unnormalized distributions and have found applications in various important tasks, attracting great research interest in their training algorithms. In general, GFlowNets are trained by fitting the forward flow to the backward flow on sampled training objects. Prior work focused on the choice of training objects, parameterizations, sampling and resampling strategies, and backward policies, aiming to enhance credit assignment, exploration, or exploitation of the training process. However, the choice of regression loss, which can highly influence the exploration and exploitation behavior of the under-training policy, has been overlooked. Due to the lack of theoretical understanding for choosing an appropriate regression loss, most existing algorithms train the flow network by minimizing the squared error of the forward and backward flows in log-space, i.e., using the quadratic regression loss. In this work, we rigorously prove that distinct regression losses correspond to specific divergence measures, enabling us to design and analyze regression losses according to the desired properties of the corresponding divergence measures. Specifically, we examine two key properties: zero-forcing and zero-avoiding, where the former promotes exploitation and higher rewards, and the latter encourages exploration and enhances diversity. Based on our theoretical framework, we propose three novel regression losses, namely, Shifted-Cosh, Linex(1/2), and Linex(1). We evaluate them across three benchmarks: hyper-grid, bit-sequence generation, and molecule generation. Our proposed losses are compatible with most existing training algorithms, and significantly improve the performances of the algorithms concerning convergence speed, sample diversity, and robustness.

cs.LG↗

arXiv2026

Enhancing Multilingual Reasoning via Steerable Model Merging

Model merging is an effective technique for composing the capabilities of a multilingual model and a reasoning model. It has achieved promising generalization in multilingual reasoning tasks by aligning feature spaces of different models. However, the merged single model often fails to address the conflicts between source models, leading to suboptimal performance. In other words, the one-size-fits-all merging strategy may not align with the characteristics of different inputs which may require prioritizing certain models over others. To this end, we propose a Steerable Model Merging (ST-Merge) framework to modulate the contribution of each source model. To realize this idea, we introduce a gated cross-attention mechanism to weight or filter the two attended source models in an adaptive manner. Extensive experiments demonstrate that ST-Merge consistently outperforms multiple strong baselines on four multilingual reasoning benchmarks across 21 different languages.

cs.CL↗

arXiv2026

Improving Generalization and Data Efficiency with Diffusion in Offline Multi-agent RL

We present a novel Diffusion Offline Multi-agent Model (DOM2) for offline Multi-Agent Reinforcement Learning (MARL). Different from existing algorithms that rely mainly on conservatism in policy design, DOM2 enhances policy expressiveness and diversity based on diffusion model. Specifically, we incorporate a diffusion model into the policy network and propose a trajectory-based data-reweighting scheme in training. These key ingredients significantly improve algorithm robustness against environment changes and achieve significant improvements in performance, generalization and data-efficiency. Our extensive experimental results demonstrate that DOM2 outperforms existing state-of-the-art methods in all multi-agent particle and multi-agent MuJoCo environments, and generalizes significantly better to shifted environments {(in $28$ out of $30$ settings evaluated)} thanks to its high expressiveness and diversity. Moreover, DOM2 is ultra data efficient and requires no more than $5\%$ data for achieving the same performance compared to existing algorithms (a $20\times$ improvement in data efficiency).

cs.AI↗

arXiv2026

Offline Diffusion Policy for Multi-User Delay-Constrained Scheduling

Effective multi-user delay-constrained scheduling is crucial in various real-world applications, including embodied AI, instant messaging, live streaming, and data center management, where efficient resource allocation is required among users with diverse delay sensitivities. In these scenarios, schedulers must make real-time decisions to satisfy both delay and resource constraints without prior knowledge of system dynamics, which are often time-varying and challenging to estimate. {Current learning-based methods typically require online interactions with actual systems during the training stage. Therefore, these approaches are often difficult or impractical, as they can significantly degrade system performance and incur substantial service costs.} To address these challenges, we propose a novel offline reinforcement learning-based algorithm, named \underline{S}cheduling By \underline{O}ffline Learning with \underline{C}ritic Guidance and \underline{D}iffusion Model (SOCD), to learn efficient scheduling policies purely from pre-collected \emph{offline data}. SOCD innovatively employs a diffusion policy, complemented by a sampling-free critic network for policy guidance. By integrating the Lagrangian multiplier optimization into the offline reinforcement learning, SOCD efficiently trains high-quality constraint-aware policies exclusively from available datasets, eliminating the need for online interactions with the system. Experimental results demonstrate that SOCD is resilient to various system dynamics, including partially observable and large-scale environments, and delivers superior performance compared to existing methods.

cs.AI↗