arXiv subjects

Xiaoyu Shen

Publications and source records attributed to Xiaoyu Shen.

At least 19 recordsLinked to original sources

arXiv2026

Rescuing Your Pruning: Rethinking Visual Token Pruning and Post-Training Recovery in MLLMs

Visual token pruning substantially reduces the inference cost of multimodal large language models (MLLMs), yet existing methods are mostly training-free and plug-and-play. They treat the resulting performance drop as an inevitable cost of compression and pursue increasingly complex pruning rules. We rethink this view through two questions: how much of the loss can post-training rescue, and, building on this recovery, how can joint compression push pruning to its limit? We curate a 282K-sample post-training mixture covering general VQA, reasoning, OCR, and grounding, and post-train 36 configurations that drop or merge tokens under a unified protocol. Post-training rescues most of the loss, even for OCR and grounding, and its adaptation gain in closing the gap between the pruned and the dense model grows as the budget tightens. For joint compression, we find that the vision encoder, not the LLM, dominates prefill time in native-resolution MLLMs, so where tokens are removed matters more than how many: keeping 40% of tokens at the encoder runs faster than keeping 10% at the projector. Our best configuration prunes the encoder lightly and the projector aggressively, leaves the LLM intact, and keeps general understanding and reasoning at the dense level. It accelerates the LLM stage by 3.49x but the end-to-end time-to-first-token by only 1.54x, showing that LLM-stage speedups overstate end-to-end gains. We hope these findings offer insights for future research on joint visual token compression. Our data, protocol, and code will be released.

cs.CV↗

arXiv2026

WIDE: Boosting Adaptive LLM Inference via Token-level Dynamic Width Pruning

Pruning is a promising approach for improving the efficiency of large language models (LLMs). Static structured pruning is hardware-friendly, but its input-agnostic computation allocation often degrades model quality at high sparsity. Dynamic pruning adapts computation to individual tokens, but coarse-grained layer-level routing limits allocation flexibility, and finer-grained decisions introduce irregular execution that can offset computational savings. To address these challenges, we present WIDE, the first end-to-end differentiable framework that combines token-wise structured dynamic width pruning with GPU kernel co-design for both prefill and decoding. WIDE uses differentiable routers to select GQA-aligned query-head groups and GPU-tile-aligned FFN-channel groups for each token. To execute these decisions efficiently, it reorders routing masks to cluster active tokens within each group, enabling hardware-agnostic block-level skipping and architecture-aware skipping of inactive loads and tensor-core operations. Experiments on the Llama3 and Qwen3 families demonstrate improved quality retention over representative static and dynamic pruning methods. At 50\% target routing sparsity, WIDE improves Llama3.1-8B average zero-shot accuracy by up to 20.26 points over the strongest dynamic-depth baseline under calibration-only settings, while retaining 97.6\%/88.2\% of dense Qwen3-14B likelihood/generation performance after LoRA tuning. Separate latency benchmarks demonstrate end-to-end speedups of up to $1.68\times$ for prefill and $1.55\times$ for decoding over the dense baseline. These results demonstrate that pruning--kernel co-design pushes the frontier of token-wise dynamic structured pruning, combining improved quality retention with practical acceleration across both prefill and decoding. Our code is avaliable at \url{https://github.com/EIT-NLP/LLM-Pruning/tree/main/WIDE}.

cs.AI↗

arXiv2026

ReInsVLN: Revisiting Historical Evidence for Training-Free Vision-and-Language Navigation

Vision-and-language navigation (VLN) relies on retaining and using relevant observations. Yet growing histories introduce redundancy, and retained information may remain underused. We introduce ReInsBench, a diagnostic benchmark that examines history utilization through controlled continuation trials. Its results show that history-review reinstruction can recover navigation failures using the existing task context. Motivated by this finding, we propose ReInsVLN, a training-free framework supporting memory construction and utilization in frozen navigators. Coverage- and Instruction-Aware Compression constructs compact visual context through temporal redundancy reduction, coverage-aware protection, and instruction-guided token selection. Forgetting-Inspired History Reminders guide navigators to revisit retained observations according to a predefined schedule, without additional model calls. On R2R Val-Unseen, ReInsVLN improves success rates by 8.48 and 3.37 percentage points over the native StreamVLN and InternVLA-N1, reaching 64.27% and 64.44%, respectively. Visual compression largely preserves navigation performance while reducing prefill latency and KV cache memory by up to 21.77% and 32.25%, respectively, on matched observation workloads. Real-world trials on humanoid and quadruped robots further support the applicability of ReInsVLN to sequential instruction following and ordinal navigation.

cs.RO↗

arXiv2026

ATTUNER: Recomputation-Free Agentic Context KV Reuse

Large language model (LLM) agents repeatedly load reusable content, such as skills, documents, and memory entries, into the current context. Re-encoding this content for every request wastes computation. Position-independent caching (PIC) alleviates this by encoding each artifact independently and reusing its key-value (KV) states at arbitrary positions, but it incurs a quality loss relative to full-context prefill. Existing methods repair this loss by restoring global position IDs or recomputing selected tokens. In this work, we isolate the source of the loss, finding that the positional mismatch has minor effect, and independently cached artifacts retain faithful representations: reading a provided artifact stays largely accurate, and performance degrades only when the model must select among multiple artifacts. Moreover, replacing PIC's attention scores with full-prefill scores recovers performance with the cached KV unchanged, localizing the failure to the attention rather than KV recomputation. Motivated by this, we propose \textsc{Attuner}, a query-side adaptation method that learns to read a frozen artifact cache. \textsc{Attuner} inserts low-rank adapters into the query projections and is trained by distilling full-prefill distribution into the student. It trains fewer than 0.05\% of the model parameters and, at inference, requires neither cache recomputation nor a full-context reference. On Qwen3-4B and Qwen3-8B across seven benchmarks covering skills, documents, memory, and code, \textsc{Attuner} substantially outperforms prior PIC baselines in both in-domain and out-of-domain settings, matches full-context prefill quality while providing up to $3.73\times$ speedup.

cs.CL↗

arXiv2026

Bifocal: Training-Free Mixed Granularity Visual Token Pruning for Multimodal LLMs

Visual token pruning for multimodal large language models (MLLMs) appears highly effective, with existing methods retaining over 95% of unpruned performance on common VQA benchmarks using a quarter of the visual tokens or fewer. However, our extensive evaluation of ten methods across three MLLM families and five ability groups reveals a critical gap: strong general performance often comes at the cost of severe degradation on fine-grained tasks such as OCR and visual grounding. We trace these failures to two limitations of the prevailing paradigm of retaining the highest-ranked native patches. A single importance ranking overlooks the complementary evidence required by different tasks, while sparse patch selection disrupts both image content and spatial integrity. To address these limitations, we propose Bifocal, a training-free, two-stage pruning method built on a mixed-granularity framework. It allocates separate budgets to query-focused and image-intrinsic evidence, complemented by a pooled coarse representation of the entire image. Before the LLM, Bifocal matches question words to visual regions through a soft conjunction balanced against norm-debiased image saliency, while selecting image-intrinsic evidence to maximize representational coverage. Inside the LLM, it further prunes tokens using density-adjusted text-to-image attention. With roughly a quarter of the original visual tokens, Bifocal is the only evaluated method to retain over 90% of unpruned performance in every ability group across all three models, and ranks first on average in all nine model-budget settings. The code is available at: https://github.com/EIT-NLP/Bifocal

cs.CV↗

arXiv2026

Perceive, Solve, Reason: Progressive Capability Formation in Time Series-LLMs

Large language models (LLMs) can now perceive, describe, and reason about images and audio, yet time series understanding remains underestimated: most work uses LLMs as forecasters and classifiers, reasoning studies are few and narrow, and no suite trains and evaluates the full range of capabilities. In this paper, we build the first such suite, with 337K training and 44K evaluation samples across three progressive capabilities, Perceive, Solve, and Reason, and use it to ask how time series should enter an LLM that must master all three, and in what order. We find that a dedicated encoder is essential: with one, a 4B model surpasses GPT-4o on every perception attribute we test, whereas textualizing the series costs 8--10 points in reasoning accuracy. The capabilities then build on one another: learning them in order beats mixing the same data and rivals reasoning-only training with over three times the reasoning data, but only when an encoder is used. Analyses further reveal that capabilities are transferred through distinct network depths rather than simply stacked, and that a stage converged on its own benchmark is not necessarily ready for the next. Distilled into a simple recipe, these findings yield 4B TS-LLMs that, without any target-domain data, outperform four recent 7B TS-LLMs on seven of eight out-of-domain metrics, by up to 11 points. We therefore advocate treating time series as a first-class modality for LLMs, read through dedicated encoders and taught to perceive before they reason.

cs.CL↗

arXiv2026

Nucleus Speculative Decoding: Plausibility-Aware Verification Beyond Exact Distribution

Speculative decoding (SD) accelerates autoregressive generation by letting a lightweight draft model propose tokens that a target model verifies in parallel. However, its standard acceptance rule, designed for exact distribution correction, rejects tokens to which the draft assigns excess probability, even when the target finds them highly plausible: over half of the rejected tokens lie inside the target's own nucleus. Since verification dominates the cost of SD, these rejections waste its most expensive operation. We introduce Nucleus Speculative Decoding (NSD), a plausibility-aware verification rule that accepts a draft token if it passes the standard test or belongs to the target nucleus, while keeping SD's residual correction and bonus sampling unchanged. NSD requires no training or extra forward passes and is compatible with any drafting mechanism. Because it preserves SD's recovery rule, its deviation from the target distribution can be characterized exactly: we prove that the single-step total variation error equals the draft's excess probability inside the target nucleus, which is precisely NSD's gain in acceptance probability. NSD thus trades fidelity for acceptance at an exact one-to-one rate, and we further derive sequence-level fidelity bounds. Across six target models and five mathematical reasoning and code generation benchmarks, NSD achieves the highest throughput among all compared methods, with speedups of up to $5.16\times$ over autoregressive decoding and $3.15\times$ over standard SD, while improving accuracy over SD on five of the six models and nearly halving verification time per output token. Our code is available at https://github.com/EIT-NLP/Nucleus-Speculative-Decoding.

cs.CL↗

arXiv2026

TAME-Nav: Temporal Asymmetric Mixture-of-Experts Distillation for Agile and Versatile Robot Navigation

Agile and reliable navigation is a core capability for embodied robots navigating complex spatial constraints and dynamic interactions. However, a single policy often struggles to accommodate the distinct decision-making demands of diverse scenarios. We present TAME-Nav, an asymmetric Teacher-Student navigation framework built around a temporal soft mixture of experts (Soft-MoE). The Teacher combines clean observations with causally maintained dynamic-obstacle tracks to construct privileged representations, while the Student learns complementary navigation features from corrupted observation histories through temporal encoding, soft expert routing, and latent distillation. Procedural scenario generation and parallel task-level simulation broaden training coverage, while deployment-informed asymmetric domain randomization supports sim-to-real transfer. Across in-distribution and unseen compositional scenarios, TAME-Nav achieves higher success rates, lower collision rates, and higher navigation scores than representative baselines. On a real quadruped robot, it completes all twenty trials across local-trap escape and narrow-passage navigation with sudden dynamic obstacles, with only one minor contact.

cs.RO↗

arXiv2026

GeoGR: Spatial-Semantic Token Compression and Group Routing for Efficient 3D VLMs

Multi-view representations enable pretrained image and video vision-language models to understand 3D scenes, but their long visual token sequences make inference costly. Redundancy arises both from repeated observations of the same scene regions across views and from processing all visual tokens at every LLM layer, regardless of their relevance to the current computation. We propose GeoGR, a two-stage framework that jointly compresses scene representations and visual computation. At the projector level, GeoSemZip consolidates cross-view observations into 3D voxel tokens and combines semantic saliency with spatial coverage to construct a compact scene representation. Unselected voxel features are aggregated into contextual anchors to retain scene information that would otherwise be discarded. At the LLM level, we identify an effective visual computation window and observe that query-conditioned token relevance remains locally stable but changes over longer layer intervals. Guided by these findings, GroupRoute combines late entry and early exit with recoverable group-wise skipping. Low-relevance tokens bypass complete LLM layers while retaining their hidden states and can rejoin the active group at subsequent anchor layers selected under a computation budget. For compression-aware post-training, we construct a compact dataset of 67K samples by retaining 30% of the original training pool. With this protocol, GeoGR retains 95.2% and 96.8% of the dense post-trained models' performance on LLaVA-OneVision-7B and Video-3D LLM-7B, respectively, while accelerating LLM prefill by $6.8\times$ and $6.1\times$. These results demonstrate the consistent effectiveness of GeoGR across 3D VLMs with and without explicit 3D spatial encoding, highlighting a favorable performance-efficiency trade-off. The code is available at https://github.com/EIT-NLP/GeoGR.

cs.CV↗

arXiv2026

DIAL-OPD: Learning More from Fewer Tokens in On-Policy Distillation

On-policy distillation (OPD) supervises student-generated trajectories with token-level teacher signals. Its sampled-token variant avoids the cost of full-vocabulary probabilities. Yet we find that training on fewer tokens can outperform full-token OPD, challenging the intuition that more supervision improves learning. This motivates selecting tokens by learning value. Existing disagreement-based criteria ignore probability scale: tokens assigned negligible probability by both models, termed low-low tokens, can receive large log-ratio rewards and hinder learning. We propose DIAL-OPD, a token-selection method that bridges log-probability and probability spaces by weighting reward magnitude with the logarithmic mean of teacher and student probabilities. A parameter beta controls this weighting, and the highest-scoring tokens are retained. Across 4 teacher-student pairs and 7 mathematical reasoning benchmarks, we compare DIAL-OPD with 9 baselines. Retaining only 40% of tokens, it outperforms Vanilla OPD and its full-token variants, with mean accuracy gains reaching 5.25 percentage points over Vanilla OPD, and doubles AIME25 Pass@16 from 13.33% to 26.67%. It also achieves up to an 18% relative improvement in mean accuracy over the strongest token-selection baseline at matched retention ratios. With a 4B teacher, DIAL-OPD surpasses the strongest full-token baseline using an 8B teacher at both student scales, showing that effective supervision allocation can outweigh teacher scaling. Further analysis shows that moderate beta balances suppressing low-low tokens against preserving useful disagreements. Token-level evidence reveals that DIAL-OPD filters high-reward tokens with limited reasoning value while preserving supervision critical to reasoning correctness.

cs.CL↗

arXiv2026

The Trace Is the State: Exact Credit Assignment for LLM Agent Teams

Credit assignment for a team of LLM agents, what each message was worth, has mostly been treated as prediction: a learned critic, a trajectory-level score, or an agent ablation stands in for the counterfactual that defines credit, which is rarely run. Teams that communicate through a shared context are different: when everything a downstream agent reads is written into the trace, the trace is the state, and that counterfactual can be executed. A credit signal can then be judged as any estimator is: by bias, variance, and agreement with an independent reference. C3, credit assignment by counterfactual continuation, substitutes one message at a decision point and continues the run to the terminal reward, so its credit is unbiased, exact up to Monte Carlo error. Given the sampled alternatives, that error's variance follows a derived law with no term for the number of agents, and the observed noise follows the law on 6 workflows of 2 to 10 decision points. On the 2-agent chain, from 4 continuations, C3 ranks alternatives at 0.69 rank correlation against a 16-continuation reference, near the 0.73 at which 2 such references agree; a critic trained on the same rollouts reaches 0.29. Used as the advantage in training a 2-agent team, C3 beats MAGRPO, the stronger baseline in our comparison, and spends 37% fewer training tokens than MAPPO, since only the messages downstream of a decision point are regenerated. When the trace is the state, credit need not be predicted; it can be exact.

cs.LG↗

arXiv2026

Action Shaping: Policies Absorb What They Can Express

Reward shaping has a theorem: a potential-based term can be removed without changing the optimal policy. The same practice on the action channel, an offset added in training and dropped at deployment, has no theorem. Nothing cancels an action offset, so the correction is kept at deployment or removed without a guarantee. We call it action shaping and state its principle. A trainable policy absorbs an offset its own output layer can reproduce exactly, which is what we mean by express; what is absorbed can be removed with the return intact. Its minimal instance is a zero-initialized linear head behind a learnable gate, added to an actor that trains through a learned action-value function, with no penalty or schedule. The gate rises and then falls on its own, for deterministic and stochastic actors alike, and on 20 tasks removing the head costs almost nothing. The condition is exact reproduction, not capacity: a nonlinear head with more parameters is not absorbed, and in a paired control, one linear path added to a nonlinear base head restores absorption. Exact reproduction gives the loss a flat direction that gradient noise drifts along, and the offset's amplitude indicates, before removal, what dropping the head will cost. Action shaping thus gains the counterpart of the shaping theorem, a condition for absorption, together with the mechanism behind it and a diagnostic that reads it. Policies absorb what they can express, and only that.

cs.AI↗

arXiv2026

FlashNav: Training Deployable Robot Navigation Policies in Seconds

Training Deep Reinforcement Learning (DRL) navigation policies for different robot configurations remains time-consuming. We present FlashNav, a GPU-based framework that trains robot-specific navigation policies within tens of seconds. A unified robot specification configures a lightweight simulator for batched motion updates, range sensing, and footprint collision checking over a shared occupancy map. The framework supports nonconvex footprints, different sensor configurations and drive types. Blockwise ray queries and selective observation recomputation after resets reduce simulation overhead, while GPU-resident replay and overlapping experience collection and learner updates support efficient off-policy training. Experiments covered five robot configurations and three computing platforms. With FastDSAC on a single RTX 5090 GPU, FlashNav can train a deployable navigation policy in under 30 seconds. FlashNav achieved the highest success rate and score in the benchmark comparison. The selected policies were deployed on wheeled, quadrupedal, humanoid, and irregularly shaped robots without additional policy training.

cs.RO↗

arXiv2026

MWOP: Modality-aware Width-wise Operation Pruning for Efficient MLLMs

Multimodal large language models (MLLMs) incur substantial inference costs when processing long visual-textual sequences. While existing operation compression methods exploit modality-level redundancy, they largely treat computation within attention heads and shared feed-forward network (FFN) channels as unified units, leaving finer-grained redundancy underexplored. We find that redundancy varies both across modality-interaction paths within the same attention head and across visual and textual executions of the same FFN channel. Based on these findings, we propose Modality-aware Width-wise Operation Pruning (MWOP), which independently prunes visual-to-visual (V2V), text-to-visual (T2V), and text-to-text (T2T) attention paths within each layer, and separately selects FFN channels for visual and textual inputs. A first-order Taylor criterion guides the pruning process, with FFN importance re-evaluated after attention pruning and LoRA-based recovery training. To translate the resulting fine-grained sparsity into practical acceleration, we further develop path-sparse Triton attention kernels and compact visual-side FFN execution. MWOP preserves the token sequence while reducing attention and FFN computation, making it complementary to token compression and enabling simultaneous reduction of sequence length and per-token computation. On LLaVA-OneVision-7B, MWOP alone achieves a $1.6\times$ prefill speedup with 99.7\% average performance retention across 12 benchmarks. Combined with two representative token compression methods, it further increases their prefill speedups from $2.0\times$ and $1.9\times$ to $2.9\times$ and $2.7\times$, respectively. Results on Qwen2.5-VL-7B further demonstrate its applicability across architectures. The code is available at https://github.com/EIT-NLP/MWOP.

cs.CV↗

arXiv2026

Adaptive Mass-Segmented KV Compression for Long-Context Reasoning

The linear growth of the Key-Value (KV) cache is a critical bottleneck in long-form LLM inference. Existing KV compression methods mitigate this by evicting tokens based on importance scores. However, we show that their reliance on global Top-k selection triggers Region Wipe-out: the severe eviction of contiguous reasoning blocks that derails logical coherence. To address this, we propose Adaptive Mass-Segmented (AMS) KV Compression, a framework that shifts the paradigm from token-level competition to region-aware quota allocation. AMS adaptively partitions the KV cache based on the spatial distribution of attention mass, ensuring structurally vital reasoning segments receive guaranteed memory quotas. To ensure stability during iterative decoding, an EMA-based smoothing mechanism is incorporated to prevent jitter in segment boundaries. Crucially, AMS is a universal plug-and-play layer that is orthogonal to existing scorers. It can be seamlessly integrated into representative methods such as TOVA, Expected Attention, KeyDiff, R-KV and TriAttention. AMS is also system-compatible with modern paged-KV serving frameworks such as vLLM, supporting efficient gather-and-compact KV execution without introducing additional steady-state attention overhead. Extensive experiments across a diverse suite of tasks, including mathematical reasoning (MATH500, AIME, GSM8K), code completion, open-domain QA, and sparse retrieval, demonstrate that AMS consistently mitigates structural fragmentation and boosts model performance.

cs.LG↗

arXiv2026

Disentangling Long-Term Memory via Latent Neuro-Symbolic Reasoning

Personalized agents are required to reason over long-term history interactions to infer both explicit preferences and implicit behavioral evidence. While early flat retrieval methods score memory fragments independently and neglect the distributed information, current structured memory frameworks rely on query-agnostic static graphs that fail to capture the context-dependent relations. Crucially, raw textual memories are inherently entangled and noisy, making fine-grained personalization and cross-session reasoning computationally prohibitive. To this end, we present LGM, a novel neuro-symbolic framework that shifts long-term memory disentanglement into a continuous latent space. Specifically, (i) instead of persisting fixed graphs, we design a tailored latent graph construction with a sparse autoencoder. Subject to each query, it maps historical interactions into latent memory nodes and disentangles the memory traces into sparse concept activations, dynamically synthesizing query-aware relational edge weights. (ii) A graph encoder then treats the query embedding as a conditioning preference to direct non-linear message passing across the task-specific latent subgraph. This yields a highly expressive memory representation for effective activations. Extensive experiments on long-term personalization benchmarks demonstrate that LGM significantly outperforms state-of-the-art baselines in capturing both explicit and implicit preferences while enabling personalized responses.

cs.AI↗

arXiv2026

Multimodal Duplex Interaction Agent

In this work, we present Gander, a native multimodal duplex interaction model that builds on MiniCPM-o 4.5 and is further adapted for realtime interaction with an asynchronous agent loop. In contrast to conventional turn based systems, Gander continuously processes streaming user inputs, enabling full-duplex interaction in both everyday conversations and complex workflow agent scenarios. Users can interrupt an ongoing response, while the model can proactively provide intermediate feedback or ask follow up questions. To natively support these capabilities, Gander adopts two key architectural designs: 1) a Cerebellum-Brain collaborative framework, Cerebellum is responsible for realtime interaction while the Brain handles complex reasoning and higher level agentic tasks. The two components interact continuously through tool calling and the agent orchestration runtime. 2) The Cerebellum is built upon a streaming Thinker-Talker architecture, where user inputs and model outputs are flattened into an ordered token stream at the chunk level. We evaluate Gander across conversational ability, interactive capability, understanding, and tool assisted task execution. Internal human evaluations show that Gander maintains natural and expressive spoken dialogue, while benchmark results demonstrate effective turn taking capability and encouraging results on spoken question answering and related understanding tasks. Gander also supports a range of challenging interaction settings, including background noise interference, multi-party interactions, and backchannel communication. While our current evaluation focuses on tool assisted settings, broader long horizon agent tasks and more diverse deployment conditions remain promising directions for further study. We release Gander together with its models, code, and data to facilitate further research and development in the community.

eess.AS↗

arXiv2026

UniToolCall: Unifying Tool-Use Representation, Data, and Evaluation for LLM Agents

Tool-use capability is a fundamental component of LLM agents, enabling them to interact with external systems through structured function calls. However, existing research exhibits inconsistent interaction representations, largely overlooks the structural distribution of tool-use trajectories, and relies on incompatible evaluation benchmarks. We present UniToolCall, a unified framework for tool learning that standardizes the entire pipeline from toolset construction and dataset generation to evaluation. The framework curates a large tool pool of 22k+ tools and constructs a hybrid training corpus of 390k+ instances by combining 10 standardized public datasets with structurally controlled synthetic trajectories. It explicitly models diverse interaction patterns, including single-hop vs. multi-hop and single-turn vs. multi-turn, while capturing both serial and parallel execution structures. To support coherent multi-turn reasoning, we further introduce an Anchor Linkage mechanism that enforces cross-turn dependencies. Furthermore, we convert 7 public benchmarks into a unified Query--Action--Observation--Answer (QAOA) representation with fine-grained evaluation at the function-call, turn, and conversation levels. Experiments show that fine-tuning Qwen3-8B on our dataset substantially improves tool-use performance. Under the distractor-heavy Hybrid-20 setting, achieves 93.0% single-turn Strict Precision, outperforming commercial models including GPT, Gemini, and Claude.

cs.AI↗