arXiv subjects

Wei Niu

Publications and source records attributed to Wei Niu.

At least 19 recordsLinked to original sources

arXiv2026

RCR-Router: Efficient Role-Aware Context Routing for Multi-Agent LLM Systems with Structured Memory

Large language model (LLM) agents are increasingly organized into multi-agent systems for complex reasoning and decision-making tasks. However, scaling such systems can substantially increase inference cost, especially when multiple agents repeatedly process overlapping or role-irrelevant context. Many existing multi-agent frameworks focus on orchestration and role specialization, but still rely on full-context sharing or static context routing, leading to redundant token consumption, increased runtime, and inefficient memory utilization. We present \emph{Role-Conditioned Context Routing (RCR)}, an execution-time layer that allocates role-relevant shared-memory items to agents under token budgets. Given a structured shared memory and a set of role-specialized agents, RCR selectively routes task-relevant memory items to each agent according to its role, optional task-stage metadata, and token budget. RCR integrates role-conditioned relevance scoring, token-budgeted memory selection, and iterative memory updates to reduce redundant context exposure during multi-agent execution. Experiments on multi-hop QA, embodied reasoning (ALFWorld), and web-based decision-making (WebShop) show that RCR reduces token usage by 35\% on average and runtime by 38\% on average relative to Full-Context, while maintaining or improving task performance.

cs.CL↗

arXiv2026

BitLLM: Efficient Low-Bit, Conversion-Hiding LLM Decoding on Mobile GPUs

The rapid scaling of large language models (LLMs) has unlocked remarkable capabilities across various domains. However, deploying LLMs on mobile devices remains challenging because of their high computational and memory requirements. Low-bit quantization (e.g., 4-bit weights, 8-bit activations) can reduce model size while preserving accuracy, but its practical efficacy on mobile GPUs has been limited so far. In this paper, we identify that the performance gap stems primarily from frequent data type conversions between quantized and floating-point values, and dynamic and inefficient memory access patterns caused by the growing Key-Value (KV) cache during autoregressive decoding. We introduce BitLLM, a mobile inference engine that addresses type conversion overhead and suboptimal memory accesses in quantized LLMs. Our approach features an adaptive layout and packing strategy for the KV cache that ensures efficient memory access throughout LLM decoding. This is coupled with an integer-native execution pipeline that defers and overlaps type conversion outside the compute hot path, keeping the critical MatMul stages in the integer domain. Implemented on off-the-shelf mobile GPUs, BitLLM outperforms leading engines, i.e., LLAMA.CPP, MNN-LLM, LiteRT-LM, and MLC-LLM, achieving average speedups of 1.78$\times$, 1.73$\times$, 1.58$\times$, and 1.47$\times$, respectively, across various models and quantization levels. Under long-context generation, BitLLM reduces average energy consumption (per token) by up to 88.9%.

cs.AR↗

arXiv2026

Splaxel: Efficient Distributed Training of 3D Gaussian Splatting for Large-scale Scene Reconstruction via Pixel-level Communication

3D Gaussian Splatting (3DGS) enables high-fidelity and real-time 3D scene reconstruction, but scaling training to large-scale scenes requires optimizing hundreds of millions of Gaussians across multiple GPUs. Existing distributed approaches either partition scenes into isolated regions, causing global inconsistency, or rely on global Gaussian-level exchanges, which lead to substantial growth in inter-GPU communication and quickly dominate iteration time. We propose Splaxel, a communication-efficient distributed 3DGS training framework based on pixel-level local rendering and global composition. Instead of synchronizing Gaussians, each GPU renders its local subset and exchanges only partial pixel values, maintaining mathematical consistency while keeping communication cost stable as the scene size increases. Splaxel further reduces pixel-level redundancy through geometric and transmittance visibility prediction and improves GPU utilization via conflict-free camera-view consolidation. Evaluated on large-scale datasets with up to 120M Gaussians, Splaxel achieves up to 7.6$\times$ speedup over the state-of-the-art distributed 3DGS framework while preserving high reconstruction quality.

cs.DC↗

arXiv2026

Multiple representations in difference sets under polynomial density constraints

Motivated by the classical study of additive representation functions and the combinatorial phenomenon of sum-difference asymmetry, we investigate the subtractive representation function $d_{\mathcal{B},\mathcal{A}}(n)$, which counts the number of solutions of the equation $n = b - a$ with $a\in \mathcal{A}$ and $b\in \mathcal{B}$. Specifically, we study lower bounds and the asymptotic behavior of $d_{\mathcal{B},\mathcal{A}}(n)$ under polynomial density constraints of orders $x^α$ and $x^{1-α}$, where $0 < α< 1$. We first establish a general lower bound for the first moment of $d_{\mathcal{B},\mathcal{A}}(n)$. Furthermore, by employing a recursive construction, we prove that this lower bound is essentially optimal under certain regularity assumptions on the counting functions. Finally, utilizing probabilistic methods, we show the existence of sets for which $d_{\mathcal{B},\mathcal{A}}(n)$ has a logarithmic point-wise upper bound.

math.NT↗

arXiv2026

Mask-Aware Execution for Efficient JEPA Training

Joint Embedding Predictive Architectures (JEPAs) are becoming a core representation-learning primitive and a building block for latent world models across vision, video, audio, brain dynamics, and time series. Despite (potential of) wide deployment, current JEPA training pipelines are inefficient: each input is executed through multiple mask-specific branches, with redundant target-side work, and memory-bound token routing. These costs grow with the number of masks and limit GPU efficiency. We present M-JEPA, a mask-aware execution architecture that restructures JEPA training without changing the learning objective. M-JEPA separates mask-independent computation from mask-dependent routing, enabling shared context encoder execution, fused token routing and slicing with backward support, sparse target encoder execution over the union of target tokens, and masked patch embedding for sparse inputs. The resulting pipeline preserves training semantics while reducing computation, memory traffic, and synchronization overhead. We implement M-JEPA for five JEPA variants and evaluate it on NVIDIA A100 GPUs. Compared against the state-of-the-art baselines, M-JEPA achieves up to 1.7x end-to-end training speedup for 2-10 masks. Separately, with masked patch embedding, 4.75x patch-embedding speedup at high sparsity. These results show that execution restructuring, rather than changes to the JEPA objective, is a key lever for efficient JEPA training.

cs.DC↗

arXiv2026

Learning Multi-Step Query Rewriting via Corpus Feedback for Conversational Search

Conversational Query Rewriting (CQR) turns a context dependent user turn into a standalone query for a retriever, and most methods do this in a single step from the dialogue history before retrieving once. The rewrite is therefore fixed before any corpus evidence is available to correct its reference resolution or its vocabulary. We recast CQR as a sequential retrieval problem: an agent rewrites the current turn, retrieves, and conditions its next rewrite on the returned passages. The agent acts in a typed space of three rewriting operations, resolving conversational intent into a standalone query, generating lexical reformulations, or synthesizing pseudo-documents for document-to-document matching, together with a stop action that ends the episode. We train the policy with supervised fine-tuning followed by reinforcement learning against a single retrieval-quality reward, using no human rewrite annotations. Across TopiOCQA and QReCC, the agent outperforms several retrieval-aligned baselines, while remaining effective across retrieval backends and generalizing to the CAsT benchmarks without additional training. Further analysis shows that, through retrieval-reward optimization alone, the learned policy develops a behavior of grounding pseudo-documents in passages retrieved by earlier steps, substantially improving retrieval.

cs.IR↗

arXiv2026

BitNest: Bit-Nested Speculative Decoding for Memory-Efficient LLM Inference Acceleration

Speculative decoding accelerates autoregressive generation by using a lightweight draft to propose multiple tokens for parallel verification. However, existing methods often require an additional draft model or weight representation, introducing non-negligible memory overhead on resource-constrained devices. Self-speculative approaches reduce this overhead, yet still face trade-offs between draft quality, target quality, and storage efficiency. We propose BitNest, a bit-nested speculative decoding framework that embeds a low-precision draft directly into the higher-precision target representation. Instead of deriving a draft from a predefined target, BitNest first constructs a strong low-precision base and then recovers the higher-precision target through residual refinement, enabling both models to share a single physical weight representation. BitNest further extends this progressive-precision design to the KV cache for long-context inference. Across multiple 7B--8B edge-friendly LLMs and diverse workloads, BitNest achieves an average speculative acceptance rate of 95.2% while closely preserving higher-precision model quality, and delivers 1.48--1.61x end-to-end speedup over FP16 autoregressive decoding. On the LLaMA models supported by all representative self-speculative baselines, BitNest also achieves consistently competitive or higher decoding speedup.

cs.AI↗

arXiv2026

TierKV: Long-Context On-Device LLMs via Predictive Multi-Tier KV Caching

Large language models (LLMs) are moving onto mobile devices for increasingly diverse workloads over text, images, video, and audio. These applications often require long contexts, making the Key-Value (KV) cache a dominant memory bottleneck because it grows linearly with sequence length and is accessed at every decoding step. Prior work reduces KV-cache footprint through low-rank compression, token eviction, or flash offloading, but the resulting reconstruction overhead, irreversible token loss, or I/O stalls can offset the benefit of saving memory. We present TierKV, a mobile LLM inference framework built on Predictive Multi-Tier Cache Optimization (PMCO). Before decoding starts, PMCO predicts future cache demand from prefill hidden states and jointly assigns tokens to exact, low-rank, and flash-offloaded tiers under the device memory and accuracy budgets. This formulation retains access to the full context, removes the circular dependency of reactive eviction, and admits a closed-form solver that selects tier boundaries and per-layer ranks at runtime. Across eight text, vision, and audio models on three mobile SoCs, TierKV improves prefill throughput by up to 17.6x over existing mobile LLM frameworks, reduces RAM-resident KV cache by 12.5-34%, thereby enabling substantially longer contexts under the same memory budget, while incurring only minor accuracy degradation.

cs.LG↗

arXiv2026

EAServe: Encode-Aware Disaggregated Serving for Multimodal Large Language Models

Disaggregating the two stages, Prefill and Decode, onto separate GPU pools is now a standard optimization for (text-only) LLM serving. However, multimodal LLMs (MLLMs), which add a third phase, Encode, pose new challenges for resource allocation. Encode turns images, video, or audio into embeddings that the language model can consume, yielding a three-stage Encode-Prefill-Decode (EPD) pipeline. Existing frameworks offer only partial answers: text-only PD systems lack Encode, while EPD frameworks expose it as a separate service without regulating downstream request flow. The pipeline also carries a structural resource imbalance: every request enters through Encode before downstream work can begin, yet per-request execution leaves the encode GPU severely underutilized even at high loads, starving the downstream Prefill and Decode workers. Addressing this, we reposition Encode as the control point of the EPD pipeline, exposing three tightly coupled dimensions: when work enters downstream, where prefill executes, and how the GPU is shared. We instantiate this in EAServe across two co-designed layers. Its runtime manages load-adaptive micro-batching, rate-controlled partial offload to a co-resident prefill worker, and dynamic SM partitioning for predictable co-location. The configuration layer, Hybrid Auto Selection (HAS), navigates the joint space of GPU allocation, encode batch size, and offload ratio by pruning unbalanced allocations with per-stage capacity profiling and refining the remainder through TPE-based Bayesian optimization. Evaluated on three MLLM architectures spanning image, video, and audio, EAServe delivers up to 4.3x and 1.7x higher goodput than NVIDIA Dynamo and vLLM, respectively, under identical SLO constraints, sustains more balanced and higher GPU utilization across the EPD pipeline, and reaches near-optimal configurations faster than baseline search methods.

cs.DC↗

arXiv2026

Giving Credit Where It's Due: Redundancy-Aware Learning for Efficient Reasoning

Large reasoning models can produce correct yet unnecessarily long reasoning traces. Existing methods improve reasoning efficiency with trajectory-level objectives or local token- and step-level signals, but rarely model inter-step semantic dependencies. This limits their ability to distinguish redundant steps from those that support later deductions, making it harder to shorten reasoning without sacrificing accuracy. We introduce RECAP (REdundancy-aware Credit Assignment via Propagation), which addresses this limitation by assigning credit where it is due based on both a step's downstream role in the reasoning structure and its contribution to solving the problem correctly. We define structural responsibility to capture the step's downstream role by measuring how strongly later reasoning depends on it, using credit propagated backward from the final-answer node through an outcome-independent, LLM-annotated semantic dependency graph. However, a step can have high structural responsibility yet steer the reasoning away from the correct solution. RECAP therefore introduces step efficacy to measure answer-directed progress through changes in gold-answer log-likelihood as each step is added. Together, these signals reshape rollout-level GRPO advantages into step-specific updates. RECAP requires neither a separately trained process reward model nor preconstructed concise trajectories. Across two 7B models and four mathematical reasoning benchmarks, RECAP improves the accuracy-efficiency trade-off. On Qwen2.5-Math-7B, it improves pass@1 by 2.0-3.7 percentage points while reducing reasoning tokens by 8%-31% relative to GRPO across all four benchmarks. Analysis suggests these savings reflect fewer reasoning operations and less dead-end reasoning, rather than more compact expression.

cs.CL↗

arXiv2026

Local Edits, Global Ripples: Replay-Informed Policy Adaptation for Workflow Synthesis

Prompt-policy editing offers a practical way to improve agents that synthesize executable workflows without updating the underlying model. However, persistent prompt editing has two coupled properties. First, edit locality does not imply effect locality: an edit confined to one policy segment can ripple through downstream execution, altering behavior beyond the edited segment. Second, edit effects are composition-sensitive: edits that work in isolation can interfere after composition, causing one or both to lose their benefit or become harmful. Persistent adaptation must therefore support two distinct decisions: identifying where the policy should change from execution feedback, and determining whether the resulting edit remains safe to persist after composition. To address these challenges, we introduce RIPPLE (Replay-Informed Persistent Policy Localization and Editing), which separates where an edit is made from whether it remains safe after composition. It diagnoses failed trajectories, maps each actionable failure to a predefined policy segment, and restricts the correction to that part of the policy. RIPPLE then evaluates candidates against the same iteration-start policy to compare their isolated gains, before replaying promising edits after previously accepted updates to expose downstream effects and interactions. Only edits that remain safe under composition are retained. We evaluate RIPPLE on Flow-HO, a synthetic held-out benchmark for executable workflow synthesis. RIPPLE improves validation success by up to 23.1% and yields positive gains on two additional frozen language-model backbones, while maintaining edit efficiency and low execution cost. Targeted interaction analysis further demonstrates both properties: a segment-local tool-use edit changes downstream resource resolution and validation, while an edit beneficial in isolation becomes harmful after composition.

cs.CL↗

arXiv2026

When Should a Robot Think? Resource-Aware Reasoning via Reinforcement Learning for Embodied Robotic Decision-Making

Embodied robotic systems increasingly rely on large language model (LLM)-based agents to support high-level reasoning, planning, and decision-making during interactions with the environment. However, invoking LLM reasoning introduces substantial computational latency and resource overhead, which can interrupt action execution and reduce system reliability. Excessive reasoning may delay actions, while insufficient reasoning often leads to incorrect decisions and task failures. This raises a fundamental question for embodied agents: when should the agent reason, and when should it act? In this work, we propose RARRL (Resource-Aware Reasoning via Reinforcement Learning), a hierarchical framework for resource-aware orchestration of embodied agents. Rather than learning low-level control policies, RARRL learns a high-level orchestration policy that operates at the agent's decision-making layer. This policy enables the agent to adaptively determine whether to invoke reasoning, which reasoning role to employ, and how much computational budget to allocate based on current observations, execution history, and remaining resources. Extensive experiments, including evaluations with empirical latency profiles derived from the ALFRED benchmark, show that RARRL consistently improves task success rates while reducing execution latency and enhancing robustness compared with fixed or heuristic reasoning strategies. These results demonstrate that adaptive reasoning control is essential for building reliable and efficient embodied robotic agents.

cs.RO↗

arXiv2026

Optimizing What Policies Learn From: Recoverability-aware Rollout Intervention Learning

Critic-free group-based reinforcement learning has become a scalable approach for post-training large language models. However, most existing methods allocate the same number of rollouts to every task and trajectory state, even though some rollouts provide much more useful learning signals than others. Recent work has started to treat rollout generation as an adaptive decision, but two important limitations remain. First, intervention strategies are often based on fixed heuristics and therefore cannot adjust as the policy changes during training. Second, these methods usually decide only how many rollouts to generate, without explicitly controlling where and how to intervene. To address these limitations, we propose Recoverability-Aware Intervention Learning (RAIL), a training-time framework that learns how to generate rollouts based on the improvement produced by each intervention. RAIL models intervention selection as an online contextual-bandit problem and trains a recoverability controller using intervention traces collected through a shadow-to-live procedure. This allows the controller to keep learning while the underlying policy evolves. We evaluate RAIL in terms of effectiveness, adaptivity, expressiveness, and efficiency. Across multiple settings, RAIL consistently improves performance under limited rollout budgets. These results show that recoverability-aware intervention provides a principled way to generate more informative and less redundant rollouts, leading to stronger learning signals during post-training.

cs.LG↗

arXiv2026

An asymptotic Sidon basis of order $3-η$

Pilatte recently proved that there exists an infinite Sidon set of positive integers which is an asymptotic basis of order $3$, answering a problem posed by Erdős, Sárkőzy and Sós in 1994. In this paper, we strengthen this result by proving that for any $0<η<0.0527$, there exists an infinite Sidon set $\mathcal{S}\subset \mathbb{N}$ which is an asymptotic basis of order $3-η$; that is, every sufficiently large integer $m$ can be represented as \[ m=s_1+s_2+s_3 \] for some $s_1,s_2,s_3\in \mathcal{S}$ satisfying \[ \min\{s_1,s_2,s_3\}\leq m^{1-η}. \] To prove this, we develop a truncated version of Pilatte's construction and use a deep result of Sawin on sums of Dirichlet convolutions of the von Mangoldt function over function fields.

math.CO↗

arXiv2026

FBLayout: Optimizing Memory Layout for Efficient LLM Finetuning on Mobile GPUs

Transformer-based models have enabled unprecedented capabilities across language, vision, and multimodal tasks. On-device fine-tuning of transformer models offers a privacy-preserving path to personalized AI, yet remains inefficient on mobile GPUs due to severe memory constraints and frequent layout transformations in attention mechanism during training. Existing mobile training frameworks either use unified layouts for forward and backward passes -- leading to fragmented memory access and poor GPU utilization during backpropagation -- or rely on explicit layout conversions, which introduce significant transformation overhead. To overcome this, we propose FBLayout, a layout-aware framework that co-designs tensor organization with mobile GPU platforms. FBLayout introduces: (1) a unified R-Tile layout for multi-dimensional reductions across forward/backward passes; (2) tile-based index transformation to eliminate physical data movement; and (3) activation-guided layout selection to propagate efficient layouts globally. Evaluations on seven transformer models across different mobile phones (including ARM Mali and Qualcomm Adreno GPUs) show that FBLayout achieves 2.2-5.7x speedup over MNN, TFLite, and TVM, while significantly improving cache efficiency and reducing memory footprint, enabling practical on-device large model fine-tuning.

cs.AI↗

arXiv2026

On Sárközy-Sós Theorem related to representation functions

Let $\mathbb{N}_0$ be the set of all nonnegative integers. For a nonempty set $\mathcal{A}\subseteq \mathbb{N}_0$ and integers $n,h\ge 2$, let $r_{h}(\mathcal{A},n)$ be the number of representations of $n$ as $a_1+\cdots+a_h$, where $a_1\le \cdots\le a_h$ and $a_i\in \mathcal{A}$ for $i=1,\cdots,h$. In 2016, Chen and Tang showed that, for any given distinct positive integers $u_1,\cdots,u_k$ and positive rational numbers $α_1,\cdots,α_k$ with $α_1+\cdots+α_k=1$, there are infinitely many sets $\mathcal{A}\subseteq \mathbb{N}_0$ such that $r_{h}(\mathcal{A},n)\ge 1$ for all nonnegative integers $n$ and the set of $n$ with $r_{h}(\mathcal{A},n)=u_i$ has density $α_i$ for all integer $i=1,\cdots,k$. In this paper, we consider the irrational numbers $α_i$ as well. As a main result, we prove that, for any nonnegative numbers $α_0,\cdots,α_m$ with $α_0+\cdots+α_m=1$, there are infinitely many sets $\mathcal{A}\subseteq \mathbb{N}_0$ such that the set of $n$ with $r_{2}(\mathcal{A},n)=i$ has density $α_i$ for all integer $i=0,\cdots,m$. Other related results are also contained.

math.NT↗

arXiv2026

MetaBreak: Jailbreaking Online LLM Services via Special Token Manipulation

Unlike regular tokens derived from existing text corpora, special tokens are artificially created to annotate structured conversations during the fine-tuning process of Large Language Models (LLMs). Serving as metadata of training data, these tokens play a crucial role in instructing LLMs to generate coherent and context-aware responses. We demonstrate that special tokens can be exploited to construct four attack primitives, with which malicious users can reliably bypass the internal safety alignment of online LLM services and circumvent state-of-the-art (SOTA) external content moderation systems simultaneously. Moreover, we found that addressing this threat is challenging, as aggressive defense mechanisms-such as input sanitization by removing special tokens entirely, as suggested in academia-are less effective than anticipated. This is because such defense can be evaded when the special tokens are replaced by regular ones with high semantic similarity within the tokenizer's embedding space. We systemically evaluated our method, named MetaBreak, on both lab environment and commercial LLM platforms. Our approach achieves jailbreak rates comparable to SOTA prompt-engineering-based solutions when no content moderation is deployed. However, when there is content moderation, MetaBreak outperforms SOTA solutions PAP and GPTFuzzer by 11.6% and 34.8%, respectively. Finally, since MetaBreak employs a fundamentally different strategy from prompt engineering, the two approaches can work synergistically. Notably, empowering MetaBreak on PAP and GPTFuzzer boosts jailbreak rates by 24.3% and 20.2%, respectively.

cs.CR↗

arXiv2026

FragFuse: Bypassing Access Control of Large Language Model Agents via Memory-Based Query Fragmentation and Fusion

Large language model (LLM) agents increasingly rely on long-term memory to support complex task execution, user personalization, and domain adaptation. Meanwhile, emerging access-control mechanisms for LLM agents are being explored to block policy-violating requests and prevent misuse. We reveal a novel attack surface arising from agent memory operations: prohibited content that would trigger access control can be fragmented across interactions, stored in long-term memory in benign-appearing form, and later reconstructed through memory retrieval without appearing explicitly in the final user query. We propose FragFuse, the first attack that enables unprivileged users to bypass agent access control by exploiting this temporal channel introduced by long-term memory. FragFuse operates in three stages: (1) identifying rejection-responsive fragments via black-box adaptive querying with fragment masking; (2) injecting these fragments into memory using marker carrier queries; and (3) retrieving and fusing the stored fragments through a follow-up attack query. Although FragFuse can be instantiated manually for individual agents, we further develop a surrogate-based optimization scheme that tunes fusion instructions and marker designs, enabling automated attack generation without violating the attacker's threat-model assumptions. We evaluate FragFuse across four representative agent settings and task domains, covering three state-of-the-art agent access-control mechanisms. FragFuse achieves an average bypass success rate of 86.3% and an average end-to-end harmful task success rate of 41.1% across all settings, with only 4.4% average task-success degradation compared with configurations without access control. We also show that alternative defenses, including state-of-the-art prompt-injection detectors and perplexity detectors, do not effectively address this attack.

cs.CR↗