arXiv subjects

Hao Jiang

Publications and source records attributed to Hao Jiang.

At least 19 recordsLinked to original sources

arXiv2026

Efficient Token-Level Policy Optimization for Masked Diffusion Models via Velocity Alignment

Reinforcement Learning with Verifiable Rewards (RLVR) successfully aligns autoregressive models in rigorous domains like mathematical reasoning and code generation. However, applying RLVR to Diffusion Large Language Models (dLLMs) is hindered by intractable sequence-level likelihoods, forcing existing methods to rely on unstable, high-variance approximations. We propose Sequence Likelihood-Free Policy Optimization (LFPO), an efficient framework that circumvents this bottleneck by formulating policy alignment as discrete velocity alignment. Inspired by continuous flow matching, LFPO eschews trajectory-level estimation and directly optimizes denoising logits via contrastive token-level updates, providing a stable, low-variance surrogate. Furthermore, predicting terminal solutions from intermediate states implicitly straightens the generation trajectory, enabling high-quality sampling with significantly fewer denoising iterations. Extensive experiments on reasoning and coding benchmarks demonstrate that LFPO matches or exceeds the accuracy of likelihood-based baselines while substantially accelerating inference.

cs.LG↗

arXiv2026

Flexible Beamforming via Fluid Antenna Array: From Discrete to Continuous Array

Fluid antenna arrays (FAAs) reconfigure their effective aperture by activating radiation ports from a dense set of candidate positions within a finite physical region, thereby offering geometry diversity beyond conventional fixed arrays. This paper studies arbitrary-direction far-field beam synthesis with a planar FAA. Since a fixed array adapts its beamforming weights on a predetermined spatial sampling structure, its ability to match desired beam patterns is limited when both the aperture and the number of active elements are constrained. To exploit the geometry diversity of the FAA, we propose a coarse-to-fine scheme that couples discrete port selection with continuous joint refinement. In the coarse stage, a fast Fourier transform (FFT) based phase retrieval step converts the amplitude-only beam specification into an aperture-compatible complex target, after which an FFT-tailored greedy method selects the active ports and least squares initializes their complex weights. In the fine stage, a projected adaptive moment estimation procedure jointly refines the continuous port positions and beamforming weights while enforcing the minimum interelement spacing. For the evaluated two-beam configuration, the proposed scheme increases the mainlobe energy fraction from 35.73% to 48.92% and reduces the peak sidelobe level from -2.30 dB to -3.83 dB relative to a fixed array with the same aperture and active-port count.

cs.IT↗

arXiv2026

A counterexample to Tokareva's conjecture on sums of two bent functions

It is well known that the sum of two bent functions on $\mathbb{F}^n$ has algebraic degree at most $n/2$ for even $n\ge 4$. Tokareva conjectured that the converse holds: every Boolean function with degree at most $n/2$ can be written as a sum of two bent functions. The conjecture has been verified for even $n\leq 6$ and for several special classes of Boolean functions. We disprove this conjecture by showing that $x_1x_2(x_3x_4+x_5x_6+x_7x_8)$ is not a sum of two bent functions on $\mathbb{F}^8$. We extend this construction to obtain counterexamples in every even dimension $n\ge 8$. Using Shaporenko's reduction, we also obtain explicit counterexamples to Tokareva's derivative conjecture in every even dimension $n\ge 10$.

math.CO↗

arXiv2026

Beyond Pattern Matching in 3D MLLMs: Part-Level Functional Reasoning for Generalizable Recognition

3D MLLMs recognize objects via pattern matching, which fails when geometrically similar objects require functional reasoning to disambiguate (e.g., dresser vs. wardrobe). Our per-sample Geometric Confusion Score (GCS) exposes this failure: baseline accuracy is 33.7% on ModelNet40's most ambiguous 20% versus 84.0% on its least ambiguous 20%. We identify a function-as-annotation pattern, in which functional language decorates a known identity instead of supporting inference, and hypothesize that reversing this information flow should concentrate improvements on the most ambiguous samples. PFIR (Part-level Functional Inference for Recognition) reverses that flow. Built on PartNeXt part segmentation, its data pipeline produces 507K QA pairs with functional CoTs (geometry $\to$ function $\to$ identity), and its progressive curriculum culminates in dual-objective GRPO: strict functional reasoning and descriptive grounding for Objaverse's diverse distribution. The two RL objectives yield nearly orthogonal task vectors ($\cos = 0.056$), enabling composition via Task Arithmetic. Under a strict dual OOD protocol (Objaverse $\to$ ModelNet40; XYZRGB $\to$ XYZ-only), PFIR raises zero-shot accuracy from 60.5% to 66.7%, with gains concentrated $2.6\times$ more on geometrically ambiguous samples---supporting part-level functional reasoning as an effective route to more generalizable 3D recognition. Code: https://github.com/this233/pfir-3d.

cs.CV↗

arXiv2026

WhisperRec: Latent Reasoning for Efficient Foundation Recommendation Models

Large language models (LLMs) have demonstrated strong reasoning capabilities, motivating their adoption as backbones for foundation recommendation models (FRMs). Existing approaches typically enhance recommendation with explicit Chain-of-Thought (CoT) under the Think-then-Answer paradigm. However, generating lengthy rationales introduces substantial inference overhead, while fixed CoT templates struggle to model diverse, dynamic, and context-dependent user interests. We propose WhisperRec, an efficient latent reasoning framework for FRMs. WhisperRec compresses teacher-generated CoT into learnable latent reasoning tokens, enabling a Latent-Reason-then-Answer paradigm that performs reasoning in latent space without producing verbose rationales. This design retains decision-relevant reasoning information while avoiding the latency bottleneck of autoregressive rationale generation. Specifically, it first introduces Multi-View Adaptive CoT (MV-ACoT) to construct diverse, high-quality supervision from complementary perspectives on user interests. MV-ACoT also adapts reasoning complexity to each instance, applying lightweight analysis to clear cases and targeted multi-factor reasoning to challenging ones. Building on a pre-trained FRM, WhisperRec then employs a three-stage Latent Reasoning Alignment procedure to progressively internalize teacher CoT into latent representations. Finally, curriculum-based post-training activates latent-token reasoning for downstream recommendation while preserving standard recommendation capability. Experiments on an industrial-scale Kuaishou dataset and the public Kuaishou LLM-Rec benchmark show that WhisperRec consistently outperforms explicit-CoT methods and conventional baselines. Compared with explicit CoT Think and No-Think variants, WhisperRec improves SID@64 by 17.44% and 9.33%, respectively, and achieves over 10x higher online inference throughput.

cs.IR↗

arXiv2026

TrajCFG: Token-Wise Classifier-Free Guidance via Denoising Trajectory Feedback

Classifier-free guidance (CFG) typically applies a shared scale to all latent tokens, although their velocity predictions evolve differently during sampling. We introduce Trajectory-Adaptive Classifier-Free Guidance (TrajCFG), a training-free method for flow-matching models that adapts token-wise guidance using historical changes in guided velocity direction. TrajCFG maps this feedback to bounded scales, assigning stronger guidance to larger directional changes and weaker guidance to more consistent directions. It combines adjacent-step and interval-normalized longer-span feedback, reusing standard CFG predictions without additional model evaluations. Controlled diagnostics and ablations examine the roles of directional feedback, token-wise allocation, and history-span fusion. Experiments on text-to-video, image-to-video, text-to-audio-video, text-to-image, and image-to-3D generation show improvements over base samplers across multiple backbones. For text-to-video generation, TrajCFG improves VBench Total scores by 0.30-0.75 points over the strongest evaluated baseline on each of three backbones. These results support historical directional feedback as a broadly applicable signal for fine-grained guidance allocation. The project page is available at https://jyliu-98.github.io/trajcfg-project-page/.

cs.CV↗

arXiv2026

OneLatent: Latent Reasoning for Efficient Foundation Recommendation Models

Large language models (LLMs) have demonstrated strong reasoning capabilities, motivating their use as the backbone of foundation recommendation models (FRMs). Existing methods enhance recommendations through explicit Chain-of-Thought (CoT) reasoning under a Think-then-Answer paradigm. However, explicit CoT incurs substantial inference overhead by generating lengthy reasoning traces and relies on manually designed templates that struggle to capture diverse, dynamic user interests. We propose OneLatent, an efficient latent reasoning framework that compresses explicit reasoning traces into several learnable latent tokens, enabling Latent-Reason-then-Answer inference without generating verbose traces. OneLatent first introduces Multi-View Adaptive CoT (MV-ACoT), which creates diverse, high-quality teacher-generated supervision by exploring user interests from multiple perspectives and automatically adapting reasoning complexity to each instance. Building on pretrained FRMs, it then uses a three-stage latent-token alignment paradigm to progressively internalize CoT traces into learnable latent tokens. Finally, a multistage curriculum-based post-training strategy activates latent-token reasoning for downstream recommendation tasks. Experiments on an industrial-scale Kuaishou dataset and the public Kuaishou LLM-Rec benchmark show that OneLatent consistently outperforms explicit CoT-based methods and traditional baselines. Compared with the Think and No-Think variants of FRMs, OneLatent improves SID@64 by 17.44% and 9.33%, respectively, while achieving over 17x higher online inference throughput. We further develop a production serving system for scalable, real-time FRM inference. An online A/B test in Kuaishou's local-services advertising scenario shows that deploying OneLatent with this system yields an estimated 9.6% revenue lift over strong online baselines, including OneRec and OneReason.

cs.IR↗

arXiv2026

V-CoLA: Vision Token Compression with Linear Attention

Vision-language models (VLMs) have demonstrated impressive capabilities but suffer from substantial computational overhead, as vision tokens dominate the input sequence. This motivates vision token compression as a key direction to alleviate the burden. However, with the emergence of hybrid architectures incorporating linear attention (\eg, Qwen3.5), prior methods designed for softmax attention struggle to generalize. Our analysis reveals that both attention- and similarity-based approaches suffer notable performance degradation, underscoring the urgent need for compression methods tailored to this regime. To this end, we propose \textbf{V-CoLA}, an efficient training-free token compression framework specifically designed for linear attention. V-CoLA introduces a novel \textit{uniqueness-aware importance criterion} for identifying critical vision tokens, coupled with an \textit{adaptive token merging strategy} that performs compression. All components are optimized at the implementation level to remain compatible with the chunk-wise parallelism of linear attention, ensuring strong practical value. Extensive experiments across multiple benchmarks demonstrate the superiority of V-CoLA: it achieves 99.5\% of the original performance with only 50.0\% of vision tokens, and over 88.0\% with as few as 12.5\%, while delivering a 1.86$\times$ to 6.15$\times$ prefill speedup. Our code is available at https://github.com/alibaba/EfficientRL.

cs.CV↗

arXiv2026

Efficient SAM3 Adaptation for Multi-Class Semantic Segmentation via Dense Competitive Representations

The Segment Anything Model 3 (SAM3) has achieved significant progress in Promptable Concept Segmentation (PCS) by processing short noun phrases to generate segmentation masks with unique instance identifiers. While SAM3 excels at single-concept segmentation, it falls short in complex, real-world environments where multiple concepts must be segmented simultaneously. In particular, SAM3 typically fails to simultaneously process concepts with semantic overlap, leading to semantic misclassifications and redundant mask predictions. Moreover, as the number of concepts increases, the independent inference paradigm introduces prohibitive inference latency. To address these limitations, we propose the Dense Competitive Representations SAM (DCR-SAM) framework. By decoupling cross-modal interactions, the proposed method alleviates the inference latency. To resolve semantic misclassifications, DCR-SAM dynamically aligns textual concepts with visual representations and incorporates a dense competition head to enforce explicit inter-class competition. Furthermore, a learnable background token is applied to absorb non-target objects dynamically. These mechanisms effectively suppress overlapping predictions and generate high-fidelity masks. Extensive experimental evaluations across multiple benchmarks demonstrate that the proposed framework: (1) achieves a 2.3% $\sim$ 10.7% performance improvement over state-of-the-art methods, (2) delivers a 33$\times$ inference speedup compared to the native SAM3, and (3) exhibits strong robustness across diverse scenarios.

cs.CV↗

arXiv2026

Balancing Image Compression and Generation with Bootstrapped Tokenization

Despite progress in image tokenization, standard methods encode redundant information by mixing all granularities within each token, thus redundancy persists between tokens. The mix of information of different granularity also complicates the training of generators. This paper introduces SelfBootTok, a method that resolves this by cleanly decomposing information into global and local token groups. Through self-bootstrapped learning, the model predicts local details exclusively from global tokens, shifting the burden of visual details from the generator to the tokenizer. Consequently, our generator is far more efficient, requiring only global tokens and reducing computation by approximately 40%, while delivering superior reconstruction and generation. Moreover, this paradigm scales elegantly: by leveraging more data or parameters to self-supervise local representation learning, SelfBootTok achieves a new state-of-the-art gFID score of 1.56 using only 64 tokens.

cs.LG↗

arXiv2026

Covert Digital Semantic Communication With Movable Antennas: Constellation Shaping, Spatial Optimization, and Semantic Mapping

Digital semantic communication (SemCom) is emerging as a bandwidth-efficient paradigm for next-generation wireless networks by delivering the meaning of a source over finite constellations. However, the broadcast nature of wireless channels exposes transmissions to adversarial wardens, while discrete symbols deviate from the Gaussian signaling assumed by conventional covert communication. In this paper, we propose a covert digital SemCom system assisted by movable antennas (MAs), where the constellation, the antenna positions, and the semantic transceiver are jointly designed to conceal the transmission while preserving the semantic fidelity. Specifically, the maximum covert rate against the optimal detector with unknown noise level is derived in closed-form, which reveals that the legitimate signal-to-noise ratio (SNR) scales with the moment-matching order of the constellation and the spatial gain provided by the MA array. A Gauss-Laguerre ring constellation is then designed to raise the moment-matching order under the peak power constraints, and a covert symbol mapping with rank-to-ring assignment and scrambling is developed to carry the semantics in the ordering of the latent magnitudes. Guided by the derived semantic distortion bound, the semantic encoder and decoder are trained end-to-end through an invertible flow over the designed constellation and MA array. Simulation results demonstrate that the proposed scheme supports a higher achievable covert rate and improves the reconstruction quality by 24% compared to state-of-the-art SemCom schemes.

eess.SP↗

arXiv2026

MiCo: Mutual Information Coverage Optimization through Semantic Erasure Modeling for Efficient MLLM Inference

Multimodal large language models (MLLMs) have demonstrated impressive performance in multimodal understanding, but processing large numbers of visual tokens results in high computational costs. While many methods have been proposed to reduce the number of visual tokens, most of them rely on heuristics and are prone to discarding substantial visual information during pruning, leading to degradation in model performance. In this work, by using a semantic erasure model, we derive a general mutual information coverage objective from task log-loss and propose MiCo, a training-free two-stage pruning method. MiCo first uses visual signals to select a representative candidate pool before visual tokens enter the language model, then performs task-aware subset selection within it. At each stage, suitable observable proxies instantiate the derived objective as a monotone submodular coverage function, which MiCo greedily optimizes under the token budget. MiCo is evaluated on diverse MLLMs ranging from 7B to 13B parameters across a broad range of image and video benchmarks spanning general visual reasoning, fine-grained OCR and grounding, hallucination detection, and long-video understanding. MiCo consistently achieves the best performance across nearly all evaluated models under all pruning ratios. On LLaVA-NEXT-13B, MiCo uses only 5.6% visual tokens, retains 97.5% of baseline performance, and achieves a 3.8-fold inference speedup. Our experiments demonstrate the effectiveness of MiCo and our mutual information coverage objective for visual token pruning. Our code is available at https://github.com/MAX-7237/MiCo.

cs.CV↗

arXiv2026

Designing Transformers for Continuous Diffusion Language Models

Continuous diffusion language models (CDLMs) offer an alternative approach to parallel language generation; however, most existing studies directly use transformer architectures from either causal large language models (LLMs) or Diffusion Transformers (DiTs) for image generation, which were not optimized for CDLMs. In this paper, we conduct a systematic study on Transformer architectures for CDLMs and introduce Language Diffusion Transformer (LangDiT), combining feature and query-key normalization, wider noise conditioning, likelihood-shortcut gradient routing and calibration, and confidence-gated training self-conditioning. LangDiT significantly reduces generation perplexity in matched LangFlow comparisons, establishing a new state of the art in continuous language generation across different model formulations. Using LangFlow as a representative CDLM, LangDiT as a plug-and-play backbone leads to a new state-of-the-art model with a generation perplexity (GenPPL) of 21.72 using 1024 NFEs and 44.72 GenPPL using 128 NFEs on the unconditional generation of 1024-token sequences on the OpenWebText dataset. When adapted to Plaid-style CDLM formulations, LangDiT results in a new leading GenPPL of 11.33 using 1024 denoising NFEs, and 27.51 GenPPL using 128 denoising NFEs. These gains are achieved while retaining each model's underlying representation, training objective, and inference sampler, isolating transformer architecture as a fundamental source of improvement. Our study identifies architectural design as an important source of improvement in continuous diffusion language models.

cs.AI↗

arXiv2026

Share the Symbol, Not the Vector: Training-Free Fact Communication Across Language Models

Multi-agent systems allow language models to solve tasks collaboratively by sharing complementary information. These systems typically communicate through text, while recent approaches also explore exchanging internal model states. However, sharing these states across architectures often requires learned translators or assumptions about model compatibility. We introduce SYNC (Symbol-Indexed Neural Communication), a word-indexed interface that connects heterogeneous models without fine-tuning or a learned translator. From descriptions only it can see, the sender selects words from a shared list and sends their indices for each object, without seeing the questions. The receiver writes the described facts into its hidden states using its own word directions, vectors fitted once from its activations. We test selective use by swapping one attribute between two objects and requiring correct answers about both changed and unchanged attributes before and after the swap, four answers in total. We choose an early write layer on four development receivers and then fix it. With three different senders and nine newer receivers from five families, four-answer success on new scenes reaches 72.0-98.3%, up to 96 percentage points above middle-layer writing, and the two largest receivers match plain text. A task-free probe predicted in advance which depths each new receiver can use (45 of 45 cases). Without retuning, messages listing every fact keep 89.6-100% success under new layouts, and sender messages reach 23.9-64.0% on a landmark task with multi-token symbols, below text but from 0% without a message. These results show that shared word indices and receiver-local directions can support selective fact use across model families without additional training.

cs.CL↗

arXiv2026

When Does Touch Matter? Charting the Vision-Interaction Gap in Cluttered Dexterous Grasping

Dexterous grasping in clutter poses a basic sensing question: when do tactile measurements and external wrench estimates improve on visual geometry? Occlusion and contact can obscure grasp quality, motivating a controlled evaluation of these interaction signals. We present a controlled real-world study over five tabletop scene conditions on a dexterous hand (LEAP Hand) with vision, per-finger and wrist wrench estimates, and distributed fingertip taxels. With demonstrations, visual observations, action space, and compliant control fixed, we compare diffusion policies trained by imitation learning on vision only, wrench, taxels, or all combined, plus representation and fusion baselines. The combined policy succeeds in 24/25 trials versus 14/25 for vision only, and 15/15 versus 6/15 across the three confined conditions. Ablations show that wrench and taxel feedback are complementary. Behavioral comparisons show that interaction feedback enables earlier rejection of inadequate contacts, regrasping before lift, and more stable grasps. To our knowledge, this is the first real-world study to combine and separately evaluate these interaction modalities for target-oriented dexterous grasping in clutter. These results chart a widening vision-interaction gap and position cluttered dexterous grasping as a benchmark for determining when the learned policy needs interaction sensing. Project website: https://interaction-dex-grasp.github.io/

cs.RO↗

arXiv2026

Many Right Answers, One Training Example: Removing Responder Noise from GRPO for LLM Recommenders

A recommendation request on Reddit draws several replies, each naming different titles. When an LLM recommender is trained on such threads with reinforcement learning (RL), which replies should a generated list be rewarded against? The state-of-the-art recipe, Rank-GRPO, picks one reply per training example, so the advantage mostly records which reply was drawn (85% of its variance) rather than how good the list is. We argue that RL should train on the request, not on each answer, and propose ReMaRk: it scores every list once against the union of all replies' titles and samples each request equally. This makes RL cheaper. On seven backbones and three datasets, ReMaRk reaches the validation recall of 2,000 per-reply RL steps in 25 to 53% of the steps and leads at that budget; trained until neither improves, it is never significantly behind Rank-GRPO and in four of seven pairs reaches Rank-GRPO's best with fewer GPU-hours. The same reward helps on ProtoQA, where about 100 people answer each question, which points to a simple rule for RL with several valid human answers: reward them jointly.

cs.CL↗

arXiv2026

Diff Diffusion: Predicting Edit Paths with Diffusion Models for Efficient Code Editing

Diffusion large language models (dLLMs) are appealing for code generation because they can update many tokens in parallel, yet many are optimized for sequence synthesis rather than edit prediction. This setting is poorly matched to modern AI-assisted IDEs, where edit prediction must infer the next code change from recent edits, the current file, and cursor-local context. Such changes are usually sparse and preserve most source tokens, so regenerating a full code segment wastes computation. We propose Diff Diffusion, a diffusion-inspired LLM framework for code editing that predicts edit paths from an original snippet to its revised form. Diff Diffusion represents denoising as structured edit operations, and combines this formulation with multi-stage training and efficient parallel decoding. By focusing computation on changed regions, Diff Diffusion supports accurate low-latency edits and multi-position parallel modification. Across public edit-completion benchmarks, edit-prediction evaluations, and online validation, Diff Diffusion matches or leads autoregressive baselines while delivering 2.78 times faster inference on public edit-completion benchmarks and a 39.0% P50 model-latency reduction in production. This work provides a new modeling perspective for AI applications that require high accuracy and low latency.

cs.SE↗

arXiv2026

ErrOPD: When Does On-Policy Distillation Transfer Teacher Errors?

A distilled student can copy its teacher's mistakes along with its skills. On-policy distillation (OPD), used in recent post-training recipes, has the teacher grade every token of the student's own outputs. Which teacher errors does it pass on? We build ErrOPD, a testbed whose teachers are fine-tuned to close a JSON list one item early, an error their students almost never make. The answer depends on the KL direction: across four model families, forward-KL OPD transmits the error within the training budget (+13-23 points), while reverse-KL OPD at our reference learning rate adds under 0.05 points. A duplicated ID and, where the decision arises, two SQL errors show the same split. The reverse-KL update scales with the student's own probability of a token, and an intervention tests this: giving reverse KL forward KL's gradient magnitude at the error decision alone makes it transmit on two model pairs. Reverse-KL OPD can delay such errors rather than block them: some runs adopt the error with longer training, and Qwen3 students that already make it, even rarely, adopt it within the budget. Because greedy evaluation on Qwen3 misses the inherited error, distilled students should also be audited on sampled outputs.

cs.LG↗