Search arXiv⌕ Search

arXiv subjects

Yang Song

Publications and source records attributed to Yang Song.

At least 19 recordsLinked to original sources
arXiv2026

Test-Time Prompt-Agnostic Decomposition

Test-Time Prompt Tuning (TPT) adapts large pretrained models to unseen target distribution shifts by updating only lightweight prompt tokens. Existing TPT methods are effective, but they mainly optimize each prompt update locally. Across test-time steps, prompts updated at the current step are reused for the corresponding predictions and updates, so noisy or biased unlabeled objectives can accumulate and destabilize the prompt-update trajectory. We identify two failure modes: magnitude expansion, where updates move the prompt too far, and directional drift, where consecutive updates point in inconsistent directions. To analyze and control these failures, we propose Test-Time Prompt-Agnostic Decomposition ($\mathtt{TPD}$), which characterizes observed prompt-update trajectories from a dynamical-systems view and decomposes base prompt updates into magnitude and direction. $\mathtt{TPD}$ computes a spectral radius to measure update expansion and decomposes update directions into persistent, oscillatory, and residual components. Building on this decomposition, Adaptive Koopman Control ($\mathtt{AKC}$) regulates update magnitude by shrinking updates under expansive recent dynamics, while Hankel Update Router ($\mathtt{HUR}$) refines update direction by preserving persistent components and suppressing oscillatory and residual components. Together, $\mathtt{AKC}$ and $\mathtt{HUR}$ produce a stabilized prompt update. As a prompt-agnostic framework, $\mathtt{TPD}$ can be plugged into visual and text TPT methods without modifying the backbone, prompt architecture, or adaptation loss. Experiments on 15 datasets show that $\mathtt{TPD}$ consistently improves 10 baselines, with accuracy gains of $3$--$8\%$ for visual prompts and over $2\%$ for text prompts.

cs.CV↗
arXiv2026

SpeeDiff: Scalable Pixel-Anchored End-to-End Latent Diffusion Model

We present Scalable Pixel-anchored End-to-end Diffusion (SpeeDiff), a latent diffusion method that jointly trains the VAE and the diffusion model from scratch. In principle, joint training allows the diffusion loss gradient to directly guide the VAE encoder, encouraging the formation of a generation-friendly latent space and potentially yielding faster convergence than the conventional two-stage approach with a pretrained frozen VAE. However, a naive end-to-end implementation severely degrades performance, as unrestricted backpropagation of the diffusion loss leads to latent space collapse. Our main technical contribution is a simple yet effective Tweedie Pixel Reconstruction (TPR) loss, which provides additional pixel-level feedback by decoding a predicted clean latent from an intermediate noisy state using Tweedie's formula, thereby alleviating collapse. Furthermore, our method enables jointly scaling a fully transformer-based architecture and enhances representation alignment within the end-to-end framework. Our SpeeDiff-XL model achieves over 140x and 61x faster training compared to Vanilla SiT and REPA, respectively, while attaining an FID of 1.50 without guidance on ImageNet 256x256 generation. With a more efficient 32x compressed VAE, our model further reaches an FID of 1.53 without guidance on ImageNet 512x512 generation.

cs.CV↗
arXiv2026

Evidence Acquisition and Stopping via Lattice Distillation

Sequential evidence acquisition requires deciding what to observe and when to stop from partial information. We introduce a complete-record view of policy supervision: each fully observed training instance defines a lattice of evidence subsets, with a stop/query decision at every node. For a modest number of evidence groups, this structure can be enumerated even when individual groups contain rich representations. Our method, Subset Lattice Distillation (SLD), scores subsets with a shared masked predictor, constructs multi-step decision targets by cost-sensitive backward planning, and distills them into an observed-history controller. Uniformly resampled masks train the student across the subset space, while combinatorial planning remains offline. Separating subset scores from acquisition prices allows the predictor to be reused when the policy is replanned for new costs. Our analysis identifies action-margin conditions under which privileged supervision preserves optimal online decisions. On retrospective influenza monitoring, SLD achieves higher phase F1 at lower acquisition cost than both the closest-cost adaptive baselines and the fixed full-evidence model. A geographic and seasonal holdout, price perturbations, and ablation studies provide further evaluation.

cs.AI↗
arXiv2026

GVPO: Group Variance Policy Optimization for Large Language Model Post-Training

Post-training plays a crucial role in refining and aligning large language models to meet specific tasks and human preferences. While recent advancements in post-training techniques, such as Group Relative Policy Optimization (GRPO), leverage increased sampling with relative reward scoring to achieve superior performance, these methods often suffer from training instability that limits their practical adoption. As a next step, we present Group Variance Policy Optimization (GVPO). GVPO incorporates the analytical solution to KL-constrained reward maximization directly into its gradient weights, ensuring alignment with the optimal policy. The method provides intuitive physical interpretations: its gradient mirrors the mean squared error between the central distance of implicit rewards and that of actual rewards. GVPO offers two key advantages: (1) it guarantees a unique optimal solution, exactly the KL-constrained reward maximization objective, (2) it supports flexible sampling distributions that avoids on-policy and importance sampling limitations. By unifying theoretical guarantees with practical adaptability, GVPO establishes a new paradigm for reliable and versatile LLM post-training.

cs.AI↗
arXiv2026

VideoHarness-SRSI: Structured Recursive Self-Improvement for Video Understanding Harness

LLM system capabilities depend not only on model parameters but also on the executable harnesses that organize information acquisition, tool use, reasoning, and memory. Recursive harness evolution can automate their design, but how it can reliably discover generalizable systems from limited, noisy, and repeatedly reused feedback remains unclear. We study this question in agentic video understanding, where rich multimodal evidence and coupled mechanisms make harness optimization challenging. We first introduce VHarness-BRSI, a recursive self-improvement framework that evolves video harnesses around a frozen vision-language model. Analyzing over 4,000 evolved candidates, we find it competitive with hand-engineered systems but identify three interconnected failure modes: validation overfitting, search-budget dilution, and the early-experience trap. We trace all three to a common root: the same imperfect validation evidence drives both harness revisions and the proposer's search strategy. We therefore propose VHarness-SRSI, an evidence-governed framework in which every search decision rests on evidence that is transferable, sufficient, and gathered before the search commits. It restricts the proposer to content-free feedback, informs exploration with instrumented warm-up experiments, and develops mechanisms in isolated probe-race branches before factorial composition and joint refinement. Across four video understanding benchmarks, VideoHarness-SRSI improves the equally weighted mean held-out score by 6.2% relative to VideoHarness-BRSI and reduces the mean validation-test gap by 37.9%. Reliable harness evolution thus depends on how search is organized and informed, not only on how well it optimizes validation performance.

cs.CV↗
arXiv2026

Estimating Causal Treatment Effects in Placebo-Controlled Randomized Clinical Trials When High Placebo Response is Anticipated

In placebo-controlled randomized clinical trials (RCTs), the placebo response significantly modifies treatment effects and diminishes the intention-to-treat (ITT) treatment effect, $Δ_{ITT}$. This study presents a novel two-stage framework for estimating the standardized causal treatment effect, $Δ_{STD}$, among the ITT population, under the assumption that their placebo responses are similar to the levels of self-administering medication at home. The first stage employs a real-world, pragmatic, single-blinded placebo lead-in to measure placebo responses to levels expected during routine at-home use. This is achieved by preserving the participants' expectations and controlling for trial-related factors that inflate the responses. In the second stage, a double-blinded randomized phase is used to estimate the conditional average treatment effect (CATE) as a function of placebo response levels and other important effect modifiers. To facilitate CATE estimation, the prognostic scores, defined as the expected placebo responses, are used for dimension reduction. The causal estimand $Δ_{STD}$ is computed by integrating the CATE function over the distribution of the expected placebo response levels from stage one and other modifiers. We further derive theoretical values for $Δ_{ITT}-Δ_{STD}$ to quantify the underestimated treatment benefit due to high placebo responses. The validity and statistical performance of the proposed framework are evaluated through comprehensive simulations.

stat.ME↗
arXiv2026

Sparse All-Layer Connector for Domain Generalised Semantic Segmentation

Domain generalised semantic segmentation (DGSS) has recently benefited from combining multiple foundation models, such as vision-language and vision foundation models, to leverage their diverse representations for generalisation. Existing multi-model approaches share a common design choice: cross-model fusion is restricted to depth-aligned layers, implicitly assuming that mutually beneficial information resides at the same depth across models. However, foundation models pretrained under different objectives may develop distinct representation hierarchies, making the optimality of depth-aligned fusion questionable. Moreover, we observe that the foundation models exhibit different depth-wise domain sensitivity. Motivated by these observations, we propose a Sparse All-Layer Connector (SALC) that enables each layer to access information from preceding layers of another model. As the number of accessible layers grows with depth, dense aggregation may hurt generalisation. SALC therefore learns to adaptively select a sparse subset of informative layers. We further introduce a candidate dropout regularisation that strengthens sparsity and encourages SALC to explore diverse layers, leading to more robust selection. Across four foundation model combinations and three DGSS evaluation settings, the proposed design demonstrates stronger generalisation than conventional depth-aligned fusion. The code is available at https://github.com/RuoyuGuo/SALC.

cs.CV↗
arXiv2026

Towards Autopilot Labeling for Short Video Content Governance

The explosive growth of short videos has presented significant challenges to content governance. Traditional manual labeling is inefficient, slow to iterate, and costly. Therefore, we explore and develop a multi-stage autopilot labeling framework aimed at progressively increasing the level of automation in short video content annotation. To move beyond the heavy reliance on human expert labeling at the initial L0 stage, we first introduce a discriminative MLLM labeler at the L1 stage, which is fine-tuned with supervised data to directly predict labels for short videos. Advancing to the L2 stage, we shift to a domain-adaptive MLLM pretrained on short video data and use it for zero-shot labeling, enabling flexible generalization without task-specific supervision. In the L3 stage, we propose an agentic annotation workflow, and equip the agent with several key tools, which empower the agent to handle complex, unfamiliar, or vaguely defined cases, significantly reducing the need for human intervention. We conduct extensive experiments across various content issues, and results demonstrate that each stage of the autopilot labeling framework progressively improves labeling accuracy, and has significantly reduced manual labor costs in real-world applications by up to 85%.

cs.CL↗
arXiv2026

AdaptDuplex: From Static to Adaptive Full-Duplex Spoken Dialogue

Full-duplex spoken dialogue requires simultaneous listening and speaking at sub-second latency, under conversational timing and cognitive demands that change moment to moment. Yet current models mostly impose static operating points, lacking a systematic mechanism for adaptive decisions. We present AdaptDuplex, which extends Qwen3-Omni with such a mechanism, co-designed across three layers. A compact token-level protocol represents every window as a canonical sequence, trains dual-stream alignment through a bounded text lead over speech, and exposes every behavioral decision as an explicit token for training-free runtime control via logits bias. Adaptive mechanisms dynamically predict among discrete window durations and augment direct response as needed with non-blocking cognitive consolidation and multi-flight external reasoning. A progressive pipeline introduces these behaviors through a three-stage Thinker curriculum, then Talker-only and joint SFT, with GRPO as a preliminary increment. On Full-Duplex-Bench v1 and v1.5, AdaptDuplex outperforms DuplexOmni on 18 of 21 comparable turn-taking, overlap-behavior, and timing metrics and MiniCPM-o 4.5 on 19 of 22, with gains in both interaction decisions and response timing. On human-recorded HumDial-FDBench, it attains the top Final score (69.6) of the compared duplex models.

cs.SD↗
arXiv2026

Recurrent Harness for Agentic Continual Learning

Language model agents increasingly operate over long interaction streams, requiring online adaptation to distribution shifts through continual learning. However, existing agent harnesses typically either evolve their states only through forward information flow or rely on additional replay or validation data for harness optimization. To address this, we propose ReHarness, a recurrent harness framework for agentic continual learning that complements forward read and write with backward revision under a strictly online streaming protocol. We ask what information could have been captured differently in the harness state to better inform the current instance in the stream. Specifically, after each instance, we leverage the resulting trajectory as feedback to revise the reflections retrieved for that instance, thereby propagating feedback backward through the textual harness state. Across three continual learning benchmarks and three language model backbones, ReHarness consistently outperforms strong raw-context baselines, memory- and reflection-based methods, and prompt optimization approaches. Further analysis shows that ReHarness produces source-grounded and meaningful revisions, corrects overgeneralized or outdated experience under distribution shifts, and improves both stability and plasticity during continual learning.

cs.LG↗
arXiv2026

GR-LIO: A Local Ground-Aware LiDAR-Inertial Odometry System Using Body-to-Ground Height

LiDAR-inertial odometry (LIO) is widely used for state estimation in ground-based autonomous mobile robots. However, the geometric constraints provided by the local ground surface remain largely underexploited in existing LIO systems. This paper proposes a filter-based local ground-aware LIO framework that explicitly incorporates local ground plane geometry into the state estimation process to improve both localization accuracy and computational efficiency. Specifically, a local ground plane is parameterized by the robot orientation and the body-to-ground (B-G) height and continuously propagated within the state estimation process. Based on the proposed B-G geometry model, a propagated local ground plane enables efficient and reliable ground segmentation. The segmented ground points are then incorporated into the filter update through point-to-plane geometric constraints, improving both state estimation accuracy and efficiency. Furthermore, a planar motion update is introduced to exploit the propagated local ground plane as an additional geometric constraint, effectively suppressing vertical drift and improving estimation robustness. To address the initially unknown B-G height, an efficient initialization strategy is developed, followed by an online calibration procedure for continuous refinement. The proposed system is evaluated on several public benchmark datasets and self-collected real-world datasets covering diverse operating scenarios. Experimental results demonstrate that the proposed method consistently outperforms representative LIO methods in terms of both localization accuracy and computational efficiency.

cs.RO↗
arXiv2026

Identifying Temporal Features within Transcoders for Time Sensitive Factual Recall

Large Language Models (LLMs) suffer from temporal misalignment, often due to the contradictory nature of their training corpora. While current mitigation strategies rely on computationally expensive fine-tuning or context-heavy retrieval augmented generation (RAG), the internal mechanisms governing time-sensitive recall remain under-explored. Unlike prior studies that identify temporal components such as attention heads and MLP layers, we provide the first feature-level map of temporal recall by isolating individual MLP features via transcoder circuit tracing. We identify three node categories (common temporal, common to the year, and chrono-semantic) which interact to generate a temporal filter during factual recall. By analysing Gemma 2 2B, LLaMA 3.2 1B, and Qwen3-4B, we show that these features do not follow a simple linear pipeline but represent time through a parallel and mixed syntactic-semantic interplay across layers. We additionally discover a class of higher-layer temporal components invisible to existing EAP-IG methods, establishing transcoders as a more complete lens for temporal interpretability in time-sensitive factual recall. These findings present MLP components for potential targeted interventions in time-sensitive factual recall

cs.CL↗
arXiv2026

Dinomaly2: A Unified Framework for Unsupervised Image Anomaly Detection

Unsupervised anomaly detection (UAD) has evolved from specialized single-class models to unified multi-class models, yet existing multi-class models significantly underperform the most advanced one-for-one counterparts. Moreover, the field has fragmented into specialized methods tailored to specific scenarios (multi-class, 3D, few-shot, etc.), creating deployment barriers and highlighting the need for a unified solution. In this paper, we present Dinomaly2, a unified framework for image UAD. It bridges the performance gap in multi-class models by addressing over-generalization and contextual ambiguity, while extending across diverse data modalities and task settings. Guided by the "less is more" philosophy, we demonstrate that the orchestration of simple elements achieves superior performance in a standard reconstruction-based framework. This methodological minimalism enables natural extension across diverse tasks with minimal task-specific adaptations, establishing that simplicity is the foundation of unification. Extensive experiments on 15 UAD benchmarks demonstrate strong performance across multiple modalities, task settings, and application domains. For example, our unified multi-class model achieves unprecedented 99.9% and 99.3% image-level (I-) AUROC on MVTec-AD and VisA, respectively, with its effectiveness further extending to large-scale datasets containing hundreds of categories. Moreover, using only 8 normal examples per class, our method surpasses previous full-shot models, achieving 98.7% and 97.4%. For multi-view (Real-IAD, MANTA) and multi-modal (MVTec3D, MulSen-AD) inspection, Dinomaly2 demonstrates state-of-the-art performance through lightweight adaptations. The combination of minimalistic design, computational scalability, and broad applicability positions Dinomaly2 as a shared framework for a wide range of real-world anomaly detection applications.

cs.CV↗
arXiv2026

Learning from Next-Frame Prediction: Autoregressive Video Modeling Encodes Effective Representations

Recent advances in pretraining general foundation models have significantly improved performance across diverse downstream tasks. While autoregressive (AR) generative models like GPT have revolutionized NLP, most visual generative pretraining methods still rely on BERT-style masked modeling, which often disregards the temporal information essential for video analysis. The few existing autoregressive visual pretraining methods suffer from issues such as inaccurate semantic localization and poor generation quality, leading to poor semantics. In this work, we propose NExT-Vid, a novel autoregressive visual generative pretraining framework that utilizes masked next-frame prediction to jointly model images and videos. NExT-Vid introduces a context-isolated autoregressive predictor to decouple semantic representation from target decoding, and a conditioned flow-matching decoder to enhance generation quality and diversity. Through context-isolated flow-matching pretraining, our approach achieves strong representations. Extensive experiments on large-scale pretrained models demonstrate that our proposed method consistently outperforms previous generative pretraining methods for visual representation learning via attentive probing in downstream classification.

cs.CV↗
arXiv2026

DuplexDrama: A Synthesized Dialogue Dataset with Scenarios, Full-Duplex Behaviors, Expressive Speech, and Sound Events

We present DuplexDrama, the first synthesized spoken dialogue dataset that simultaneously covers four dimensions: (i) complete persona and scenario settings; (ii) three full-duplex behaviors (interruption, backchannel, incomplete); (iii) expressive speech with persona-aligned emotion labels; and (iv) script-aware sound events. DuplexDrama is built via a 4-stage pipeline; quality validation on both scripts and synthesized audio confirms its quality. We have produced more than 2,000 hours audio data with a 64-voice timbre pool spanning 13 personas and 5 age buckets; 3.8% of all turns carry at least one full-duplex behavior. This data has been validated through internal full-duplex model training. We will release a curated subset of 6,400 bilingual dialogues (800 h, Chinese ~500 h + English ~300 h) to advance full-duplex spoken dialogue model research. Data samples are available at our demo page and LLM-judge evaluation prompts will be released with the dataset.

cs.CL↗
arXiv2026

Same Words, Different Actions: Paired Turn-Taking Evaluation under Rewritten Dialogue Contexts

Real-time spoken dialogue systems must distinguish interruptions that require yielding the floor from backchannels that permit continued speaking. Existing benchmarks typically score events independently and may therefore assign high scores to systems with fixed action preferences rather than context-sensitive decision policies. We introduce ECHO, a paired diagnostic benchmark for Chinese turn-taking evaluation. ECHO pairs examples with the same overlap transcript but contrasting preceding multi-turn dialogue contexts, with one requiring Yield and the other Keep. It additionally includes off-talk examples for diagnosing unnecessary yielding. We introduce pair accuracy, which requires correct decisions on both members of a pair and assigns no credit to constant-action policies. Experiments on four speech systems show that three exhibit a severe over-yielding bias: they correctly keep the floor on fewer than 13% of backchannels, resulting in near-zero pairwise success rates equal or less than 4%. While the remaining system remains comparatively balanced across contexts, these findings broadly demonstrate that interruption-only evaluation can severely overestimate practical turn-taking reliability.

cs.CL↗
arXiv2026

LIGE-GR: A Smooth Leap from Ranking to Generative Recommendation in the LLM Era

The remarkable success of large language models (LLMs) has provided important inspiration for the next generation of recommender systems. Structurally, recommendation and language generation share a similarity: both aim to produce an ordered sequence that optimizes the user's experience. However, how to precisely absorb the essence of the LLM paradigm into mature industrial recommender systems remains an open problem. There are two challenges. First, it is unclear how to incorporate the LLM paradigm -- sequence-level generation and optimization -- into recommendation. Second, real-world recommender systems are mature systems that have been iteratively customized for years around specific products, business constraints, serving infrastructure, and organizational ownership. Replacing such systems wholesale is often technically risky and organizationally disruptive. In this paper, we propose LIGE-GR, a listwise generation and evaluation recommendation framework that upgrades from a traditional ranking system (itemwise recommendation) toward a generative recommendation paradigm. Instead of rebuilding the entire recommendation stack from scratch, LIGE-GR generalizes the existing pointwise recommendation system into a listwise generation system. This allows mature recommender systems to benefit from listwise optimization while preserving compatibility with existing models, value functions, and serving infrastructure. We validate LIGE-GR in short-video recommendation on Instagram Reels and Facebook Video. On these recommendation surfaces, LIGE-GR improves time spent by 1.14 percent on Instagram Reels and 0.72 percent on Facebook Video, while requiring only modest additional inference resources.

cs.LG↗
arXiv2026

GVPO++: Group Variance Policy Optimization for LLM Post-Training and On-Policy Distillation

Post-training plays a pivotal role in enhancing the reasoning capabilities and task-specific expertise of large language models (LLMs). Despite recent advances in post-training methods, such as Group Relative Policy Optimization (GRPO), their practical deployment remains impeded by training instability arising from the reliance on importance sampling. We introduce Group Variance Policy Optimization (GVPO), a novel post-training method that integrates the analytical solution of KL-constrained reward maximization into its gradient weighting scheme. This formulation provides an intuitive interpretation: GVPO's gradient corresponds to the mean squared error between the central distance of implicit rewards and that of actual rewards. GVPO offers two key advantages: (1) it guarantees a unique optimal solution, exactly to the KL-constrained reward maximization objective, and (2) it enables flexible sampling distributions without requiring importance sampling. Beyond general post-training, we show that GVPO naturally extends to on-policy distillation (OPD). Furthermore, GVPO enables the optimization of a broad family of extended OPD objectives, providing a principled foundation for diverse objective design. By unifying theoretical guarantees with practical adaptability, GVPO establishes a new paradigm for reliable and versatile LLM post-training and on-policy distillation.

cs.AI↗