Post-training quantization of speech language models is often summarized with text-output scores and nominal bit widths. Those numbers alone do not establish behavior that depends on information missing from a transcript, or efficiency for a particular runtime. We introduce an evaluation protocol that separately tests lexical output, a transcript-insufficient endpoint, and a measured packed implementation. In a Qwen2-Audio case study, a translation-selected 6-bit allocation improves chrF by 2.36 on a frozen English-to-German replay, with paired 95% bootstrap interval [1.04, 3.62], but loses 3.91 percentage points on speaker-disjoint emotion recognition. At the same 6-bit budget, the uniform structural control reaches higher emotion accuracy than the selected allocation, and the front-layer control is also higher by point estimate on the same frozen set. At 7 bits, chrF improves by 3.28 with interval [2.08, 4.59], the emotion interval against FP16 includes zero, and a same-budget front-layer control still exceeds the selected allocation. A separate matched-budget 4.08-bit study finds roughly 10-point emotion deficits for every tested low-bit allocation and no selected-allocation advantage over frozen controls. Finally, a dequantized average-6-bit simulation retains the FP16 peak memory. This case study identifies a precision-dependent mismatch between lexical output, waveform-dependent behavior, and nominal precision. It does not establish a general failure of low-bit speech models or a deployment benefit for the selected allocation.
展开完整摘要收起摘要↓
Post-training quantization of speech language models is often summarized with text-output scores and nominal bit widths. Those numbers alone do not establish behavior that depends on information missing from a transcript, or efficiency for a particular runtime. We introduce an evaluation protocol that separately tests lexical output, a transcript-insufficient endpoint, and a measured packed implementation. In a Qwen2-Audio case study, a translation-selected 6-bit allocation improves chrF by 2.36 on a frozen English-to-German replay, with paired 95% bootstrap interval [1.04, 3.62], but loses 3.91 percentage points on speaker-disjoint emotion recognition. At the same 6-bit budget, the uniform structural control reaches higher emotion accuracy than the selected allocation, and the front-layer control is also higher by point estimate on the same frozen set. At 7 bits, chrF improves by 3.28 with interval [2.08, 4.59], the emotion interval against FP16 includes zero, and a same-budget front-layer control still exceeds the selected allocation. A separate matched-budget 4.08-bit study finds roughly 10-point emotion deficits for every tested low-bit allocation and no selected-allocation advantage over frozen controls. Finally, a dequantized average-6-bit simulation retains the FP16 peak memory. This case study identifies a precision-dependent mismatch between lexical output, waveform-dependent behavior, and nominal precision. It does not establish a general failure of low-bit speech models or a deployment benefit for the selected allocation.
作者Hyunsun Chung, Taewan Noh, Minji Kim, Joo-Young Hwang, Hong-Yeon Kim, Youngjae Kim
NAND-backed storage offers the capacity needed to scale LLM prefix caching, but its block I/O path incurs CPU cache contention and host-DRAM staging in addition to NAND latency. Our characterization shows that these interface costs persist even with DRAM as the storage medium, motivating CXL-SSDs for byte-addressable access to NAND-backed capacity. Surprisingly, however, a stock CXL-SSD remains about 3$\times$ slower than local DRAM and no faster than an NVMe SSD, while generic prefetching provides little benefit. We present LM-CXD, a CXL-SSD specialized for LLM prefix caching. LM-CXD bridges the semantic gap between the serving engine, which knows which KV chunks will be consumed, and the device, which controls their placement and movement. It makes KV chunks device-visible I/O units, exposes NAND-to-DRAM progress to the serving engine, and uses device DRAM as a GPU-accessible buffer. LM-CXD further coordinates request scheduling with windowed prefetching and pipelines layerwise KV movement with GPU computation to hide NAND latency under limited device DRAM. Across five LLM models, LM-CXD reduces average TTFT over a stock CXL-SSD by up to 2.6$\times$ with compute asynchronous prefetching and 4.03$\times$ with layerwise prefetching, achieving TTFT within 1.5$\times$ of local DRAM on average.
展开完整摘要收起摘要↓
NAND-backed storage offers the capacity needed to scale LLM prefix caching, but its block I/O path incurs CPU cache contention and host-DRAM staging in addition to NAND latency. Our characterization shows that these interface costs persist even with DRAM as the storage medium, motivating CXL-SSDs for byte-addressable access to NAND-backed capacity. Surprisingly, however, a stock CXL-SSD remains about 3$\times$ slower than local DRAM and no faster than an NVMe SSD, while generic prefetching provides little benefit. We present LM-CXD, a CXL-SSD specialized for LLM prefix caching. LM-CXD bridges the semantic gap between the serving engine, which knows which KV chunks will be consumed, and the device, which controls their placement and movement. It makes KV chunks device-visible I/O units, exposes NAND-to-DRAM progress to the serving engine, and uses device DRAM as a GPU-accessible buffer. LM-CXD further coordinates request scheduling with windowed prefetching and pipelines layerwise KV movement with GPU computation to hide NAND latency under limited device DRAM. Across five LLM models, LM-CXD reduces average TTFT over a stock CXL-SSD by up to 2.6$\times$ with compute asynchronous prefetching and 4.03$\times$ with layerwise prefetching, achieving TTFT within 1.5$\times$ of local DRAM on average.
作者Tung Sum Thomas Kwok, Yidong Ouyang, Yingjia Wan, Ying Nian Wu, Zhijiang Guo, Oscar Leong
Uniform discrete flow permits repeated updates at every generation position. While continued revision supports correction of wrong tokens, it also exposes correct intermediate predictions to later errors. An experiment on Sudoku puzzles shows that 9.4% of generated cells are correct at an intermediate step but incorrect in the final output. We introduce generation order into uniform discrete flow through selective absorption, which fixes chosen predictions while preserving the uniform flow velocity at active positions. To prioritize reliable predictions for absorption, we propose Low-Entropy Discrete Flow (LEDFlow), a training-free sampler that adaptively orders absorption by local entropy. By decomposing absorption error into joint dependence and conditional prediction terms, we show that, under entropy-error regularity, selecting the lowest-entropy positions under a fixed absorption count minimizes an upper bound on the conditional term. We further analyze sensitivity of global lookahead, whose worst-case decision-error bound grows with lookahead window under an imperfect denoiser. Across reasoning benchmarks, LEDFlow attains 0.845 Nikoli Sudoku solve accuracy, with largest gains on strongly constrained tasks. On a text-to-image generation benchmark it attains the best overall score among decode-time samplers, and on multimodal understanding it improves over the default sampler on all six benchmarks, at an inference cost comparable to standard flow sampling.
展开完整摘要收起摘要↓
Uniform discrete flow permits repeated updates at every generation position. While continued revision supports correction of wrong tokens, it also exposes correct intermediate predictions to later errors. An experiment on Sudoku puzzles shows that 9.4% of generated cells are correct at an intermediate step but incorrect in the final output. We introduce generation order into uniform discrete flow through selective absorption, which fixes chosen predictions while preserving the uniform flow velocity at active positions. To prioritize reliable predictions for absorption, we propose Low-Entropy Discrete Flow (LEDFlow), a training-free sampler that adaptively orders absorption by local entropy. By decomposing absorption error into joint dependence and conditional prediction terms, we show that, under entropy-error regularity, selecting the lowest-entropy positions under a fixed absorption count minimizes an upper bound on the conditional term. We further analyze sensitivity of global lookahead, whose worst-case decision-error bound grows with lookahead window under an imperfect denoiser. Across reasoning benchmarks, LEDFlow attains 0.845 Nikoli Sudoku solve accuracy, with largest gains on strongly constrained tasks. On a text-to-image generation benchmark it attains the best overall score among decode-time samplers, and on multimodal understanding it improves over the default sampler on all six benchmarks, at an inference cost comparable to standard flow sampling.
作者Qisheng Li, Shunchao Jiang, Chen Qi, Xin Su, Da Han, Guangyong Chen
Retrosynthetic planning advances through decisions that reshape the remaining chemical problem: a locally plausible disconnection can leave precursors whose chemoselectivity constraints complicate the rest of the route. Existing planners often channel model proposals through search or template procedures, leaving open whether a general-purpose large language model (LLM) can itself sustain and revise route strategy. We developed Rachel, a stateful environment that executes and checks LLM-directed chemistry but prescribes neither a search policy nor a stopping rule. Without supplied reference routes or route-level solutions, GPT-5.5 achieved strict closure for 111 of 120 PaRoutes120 targets and 24 of 25 targets in the separate RF25 difficult-target cohort. RF25 was drawn largely from studies published after GPT-5.5's reported knowledge cutoff. Closure required complete routes and independent source resolution of every terminal precursor after planning. On a shared PaRoutes subset, forward-model support exceeded that of most comparator methods, and Rachel received the highest mean overall route score from both method-blinded LLM evaluators. Recorded trajectories showed continued model-proposed chemistry, with revised strategies carried into subsequent steps. Replacing LLM route decisions with fixed policies reduced strict closure to 6-15/120 despite continued local chemical execution; restricting planning support also reduced closure in RF25. Within Rachel, a general-purpose LLM coordinated successive chemical choices and revised its strategy as earlier decisions reshaped the remaining problems.
展开完整摘要收起摘要↓
Retrosynthetic planning advances through decisions that reshape the remaining chemical problem: a locally plausible disconnection can leave precursors whose chemoselectivity constraints complicate the rest of the route. Existing planners often channel model proposals through search or template procedures, leaving open whether a general-purpose large language model (LLM) can itself sustain and revise route strategy. We developed Rachel, a stateful environment that executes and checks LLM-directed chemistry but prescribes neither a search policy nor a stopping rule. Without supplied reference routes or route-level solutions, GPT-5.5 achieved strict closure for 111 of 120 PaRoutes120 targets and 24 of 25 targets in the separate RF25 difficult-target cohort. RF25 was drawn largely from studies published after GPT-5.5's reported knowledge cutoff. Closure required complete routes and independent source resolution of every terminal precursor after planning. On a shared PaRoutes subset, forward-model support exceeded that of most comparator methods, and Rachel received the highest mean overall route score from both method-blinded LLM evaluators. Recorded trajectories showed continued model-proposed chemistry, with revised strategies carried into subsequent steps. Replacing LLM route decisions with fixed policies reduced strict closure to 6-15/120 despite continued local chemical execution; restricting planning support also reduced closure in RF25. Within Rachel, a general-purpose LLM coordinated successive chemical choices and revised its strategy as earlier decisions reshaped the remaining problems.
作者Futian Wang, Yuhan Qiao, Xiao Wang, Dan Xu, Yuehang Li, Zhixiang Guo, Yaowei Wang, Jin Tang
Despite the remarkable progress of LLM-based and knowledge graph-augmented Radiology Report Generation (RRG) methods, existing techniques still suffer from inherent defects. Conventional LLM-only models lack structured medical prior knowledge, resulting in frequent medical hallucinations and low diagnostic interpretability. Current knowledge graph-enhanced schemes adopt static one-round knowledge fusion with single-source knowledge, incapable of dynamic knowledge updating according to generation feedback. This paper proposes a novel Multi-Agent Collaborative iterative framework for X-ray Radiology Report Generation, termed MAC-RRG. Inspired by multi-agent technology, our framework constructs a closed-loop optimization paradigm based on task decoupling and collaborative reasoning. Specifically, the framework first generates a preliminary radiology report from input X-ray images via a vision encoder and a basic LLM. Subsequently, a multimodal knowledge graph (MM-KG) agent mines structured disease correlation and anatomical knowledge from medical knowledge graphs, while an auxiliary knowledge agent extracts unstructured domain knowledge from public medical databases. The multi-source knowledge acquired by dual agents is fused and embedded to guide the LLM in iteratively refining the initial report. Extensive quantitative and qualitative experiments on mainstream X-ray RRG datasets, including IU X-ray, MIMIC, and CheXpert Plus, fully verify the superiority of our proposed method. The source code and pre-trained models have been released on https://github.com/Event-AHU/Medical_Image_Analysis
展开完整摘要收起摘要↓
Despite the remarkable progress of LLM-based and knowledge graph-augmented Radiology Report Generation (RRG) methods, existing techniques still suffer from inherent defects. Conventional LLM-only models lack structured medical prior knowledge, resulting in frequent medical hallucinations and low diagnostic interpretability. Current knowledge graph-enhanced schemes adopt static one-round knowledge fusion with single-source knowledge, incapable of dynamic knowledge updating according to generation feedback. This paper proposes a novel Multi-Agent Collaborative iterative framework for X-ray Radiology Report Generation, termed MAC-RRG. Inspired by multi-agent technology, our framework constructs a closed-loop optimization paradigm based on task decoupling and collaborative reasoning. Specifically, the framework first generates a preliminary radiology report from input X-ray images via a vision encoder and a basic LLM. Subsequently, a multimodal knowledge graph (MM-KG) agent mines structured disease correlation and anatomical knowledge from medical knowledge graphs, while an auxiliary knowledge agent extracts unstructured domain knowledge from public medical databases. The multi-source knowledge acquired by dual agents is fused and embedded to guide the LLM in iteratively refining the initial report. Extensive quantitative and qualitative experiments on mainstream X-ray RRG datasets, including IU X-ray, MIMIC, and CheXpert Plus, fully verify the superiority of our proposed method. The source code and pre-trained models have been released on https://github.com/Event-AHU/Medical_Image_Analysis
This paper presents a probabilistic perspective on large language models (LLMs), developed with the aim of bringing together, in a single self-contained account, tools that are usually treated separately across the literature. LLMs are described through probability measures on the set of sequences of tokens, specified via their autoregressive conditional distributions. Training is formulated as a maximum-likelihood estimation problem, addressed by stochastic gradient methods, while text generation is viewed as the sequential simulation of the resulting stochastic process. The role of the asymmetry of the Kullback--Leibler divergence in text generation is examined in relation with characteristic phenomena such as hallucination and the distinction between statistical plausibility and truth. As a complementary illustration of the same viewpoint, we also discuss diffusion models, built around the score function, which cast generation not as sequential token prediction but as the simulation of a reverse-time stochastic process transforming noise into data both in discrete and continuous time.
展开完整摘要收起摘要↓
This paper presents a probabilistic perspective on large language models (LLMs), developed with the aim of bringing together, in a single self-contained account, tools that are usually treated separately across the literature. LLMs are described through probability measures on the set of sequences of tokens, specified via their autoregressive conditional distributions. Training is formulated as a maximum-likelihood estimation problem, addressed by stochastic gradient methods, while text generation is viewed as the sequential simulation of the resulting stochastic process. The role of the asymmetry of the Kullback--Leibler divergence in text generation is examined in relation with characteristic phenomena such as hallucination and the distinction between statistical plausibility and truth. As a complementary illustration of the same viewpoint, we also discuss diffusion models, built around the score function, which cast generation not as sequential token prediction but as the simulation of a reverse-time stochastic process transforming noise into data both in discrete and continuous time.
作者Chenxiao Li, Yunhe Feng, Dongfang Liu, Dong Nie, Yan Huang, Heng Fan
The visual modality, i.e., images, plays a key role in multi-modal entity alignment (MMEA). Existing approaches often directly fuse the image with other modalities to align different entities. Although simple, such strategies overlook the potential noise in the images and their semantic misalignment with corresponding entities, resulting in suboptimal fusion and degraded performance. Addressing this, we propose a novel Reliability-Aware framework for MMEA (RA-MMEA), which assesses visual reliability and adaptively improves unreliable visual representations for robust entity alignment. The core lies in two modules, including dependency-aware visual reliability prediction (DA-VRP) and stability-regularized visual embedding generation (SR-VEG). The former aims to estimate the reliability of an image by leveraging multi-modal dependency within the entity, while the latter focuses on producing alternative visual representation conditioned on semantics encoded in textual modalities for multi-modal fusion. Compared to current methods, RA-MMEA enables more reliable visual representations for modality fusion, thereby improving performance. In extensive experiments, RA-MMEA achieves state-of-the-art results, verifying the importance of reliable visual modality for entity alignment and the effectiveness of RA-MMEA. The code and results will be released.
展开完整摘要收起摘要↓
The visual modality, i.e., images, plays a key role in multi-modal entity alignment (MMEA). Existing approaches often directly fuse the image with other modalities to align different entities. Although simple, such strategies overlook the potential noise in the images and their semantic misalignment with corresponding entities, resulting in suboptimal fusion and degraded performance. Addressing this, we propose a novel Reliability-Aware framework for MMEA (RA-MMEA), which assesses visual reliability and adaptively improves unreliable visual representations for robust entity alignment. The core lies in two modules, including dependency-aware visual reliability prediction (DA-VRP) and stability-regularized visual embedding generation (SR-VEG). The former aims to estimate the reliability of an image by leveraging multi-modal dependency within the entity, while the latter focuses on producing alternative visual representation conditioned on semantics encoded in textual modalities for multi-modal fusion. Compared to current methods, RA-MMEA enables more reliable visual representations for modality fusion, thereby improving performance. In extensive experiments, RA-MMEA achieves state-of-the-art results, verifying the importance of reliable visual modality for entity alignment and the effectiveness of RA-MMEA. The code and results will be released.
作者Suqin Yuan, Runqi Lin, Muyang Li, Guanzhe Hong, Jindong Gu, Lei Feng, Chris Russell, Tongliang Liu
Human-feedback alignment has made language models useful assistants and is commonly described as aligning them with humans. However, the responses people prefer from an AI need not be the responses they themselves would give. We distinguish alignment with human preferences from alignment with human behavior, and show that alignment with human preferences can make model behavior less human-like even when both preferences and responses come entirely from humans. We call this the Turing-test gap. We show that preference alignment preserves the human response distribution only under a restrictive condition, and find no consistent evidence that real human preferences satisfy it. Empirically, the loss of human-response likelihood increases with the strength of preference weighting, regardless of its direction, and the gap also appears under standard DPO. These results establish human-likeness as an explicit dimension of alignment rather than something assumed to follow from preference alignment.
展开完整摘要收起摘要↓
Human-feedback alignment has made language models useful assistants and is commonly described as aligning them with humans. However, the responses people prefer from an AI need not be the responses they themselves would give. We distinguish alignment with human preferences from alignment with human behavior, and show that alignment with human preferences can make model behavior less human-like even when both preferences and responses come entirely from humans. We call this the Turing-test gap. We show that preference alignment preserves the human response distribution only under a restrictive condition, and find no consistent evidence that real human preferences satisfy it. Empirically, the loss of human-response likelihood increases with the strength of preference weighting, regardless of its direction, and the gap also appears under standard DPO. These results establish human-likeness as an explicit dimension of alignment rather than something assumed to follow from preference alignment.
作者Yeonho Yoo, Hyunho Lee, Hyunmok Choi, Chuck Yoo, Gyeongsik Yang
Trace-driven simulation is widely used to evaluate distributed training (DT) jobs in GPU clusters, but existing simulators either ignore network contention or approximate it with a fixed penalty. This misses how scheduling decisions determine which jobs share server network interfaces and inter-server links, thereby changing networking time during training. As a result, our motivating experiments demonstrate that they incur large errors, reaching up to 73.64% mean absolute percentage error (MAPE) in average job completion time (JCT). This paper introduces MoSim, a GPU-cluster simulator that models DT job execution under dynamic network contention. MoSim combines GPU-free characterization with network contention model: it obtains each job's compute time, networking time, and networking volume without GPUs, then uses the current worker assignment to estimate how shared network interfaces affect each job's iteration time. Our evaluation shows that, compared with existing simulators, MoSim reduces simulation error for average JCT by up to 3.28$\times$, tail (99th-percentile) JCT by up to 7.79$\times$, and makespan by up to 8.48$\times$, while modeling NIC contention factors with only 8.63% error on average. By avoiding real-GPU profiling, MoSim also reduces input construction overhead by 44.6$\times$.
展开完整摘要收起摘要↓
Trace-driven simulation is widely used to evaluate distributed training (DT) jobs in GPU clusters, but existing simulators either ignore network contention or approximate it with a fixed penalty. This misses how scheduling decisions determine which jobs share server network interfaces and inter-server links, thereby changing networking time during training. As a result, our motivating experiments demonstrate that they incur large errors, reaching up to 73.64% mean absolute percentage error (MAPE) in average job completion time (JCT). This paper introduces MoSim, a GPU-cluster simulator that models DT job execution under dynamic network contention. MoSim combines GPU-free characterization with network contention model: it obtains each job's compute time, networking time, and networking volume without GPUs, then uses the current worker assignment to estimate how shared network interfaces affect each job's iteration time. Our evaluation shows that, compared with existing simulators, MoSim reduces simulation error for average JCT by up to 3.28$\times$, tail (99th-percentile) JCT by up to 7.79$\times$, and makespan by up to 8.48$\times$, while modeling NIC contention factors with only 8.63% error on average. By avoiding real-GPU profiling, MoSim also reduces input construction overhead by 44.6$\times$.
Open-vocabulary audio-visual event localization (OV-AVEL) grounds a text-queried event in time from video, audio, and language. The supervision sources available to this task can differ in temporal-boundary reliability: on OV-AVEBench, our configured visual teacher gives more reliable boundary cues than the configured audio teacher, although the latter is a strong pretrained audio model and remains semantically informative. This is a setting-specific diagnostic rather than a universal ranking of vision and audio. We formulate the resulting challenge as supervision placement: which teacher signals may shape the localization decision, and which should remain auxiliary. Based on this view, we propose OV-OrthKD, a reliability-aware asymmetric distillation framework. Visual feature transfer shapes a decision-aligned representation, audio feature transfer enriches a complementary auxiliary subspace, a text prototype anchors seen/unseen category semantics, and an orthogonality loss limits directional overlap between the two teacher-specific projections. The student continues to use both modalities through query-aware fusion at inference, while the default training recipe keeps audio-teacher supervision off the segment-logit path. On OV-AVEBench, OV-OrthKD achieves 0.816 segment AP and improves F1@0.5 over the official fine-tuning baseline by 2.7 points overall and 3.4 points on unseen categories. Path-assignment, role-swap, corruption, and transfer analyses consistently support supervision placement as a task-specific design axis for OV-AVEL.
展开完整摘要收起摘要↓
Open-vocabulary audio-visual event localization (OV-AVEL) grounds a text-queried event in time from video, audio, and language. The supervision sources available to this task can differ in temporal-boundary reliability: on OV-AVEBench, our configured visual teacher gives more reliable boundary cues than the configured audio teacher, although the latter is a strong pretrained audio model and remains semantically informative. This is a setting-specific diagnostic rather than a universal ranking of vision and audio. We formulate the resulting challenge as supervision placement: which teacher signals may shape the localization decision, and which should remain auxiliary. Based on this view, we propose OV-OrthKD, a reliability-aware asymmetric distillation framework. Visual feature transfer shapes a decision-aligned representation, audio feature transfer enriches a complementary auxiliary subspace, a text prototype anchors seen/unseen category semantics, and an orthogonality loss limits directional overlap between the two teacher-specific projections. The student continues to use both modalities through query-aware fusion at inference, while the default training recipe keeps audio-teacher supervision off the segment-logit path. On OV-AVEBench, OV-OrthKD achieves 0.816 segment AP and improves F1@0.5 over the official fine-tuning baseline by 2.7 points overall and 3.4 points on unseen categories. Path-assignment, role-swap, corruption, and transfer analyses consistently support supervision placement as a task-specific design axis for OV-AVEL.
Deploying efficient neural networks is essential in resource-constrained environments, yet compact models often sacrifice interpretability - a critical in safety-critical domains such as autonomous driving and medicine. This study investigates whether Knowledge Distillation transfers the spatial feature attribution of a large teacher network to a compact student. To assess the influence of the KD scheme on interpretability, we distill a ResNet-152 teacher into a ResNet-34 student on ImageNet-1K across five configurations by systematically varying the distillation temperature and soft-label loss weight. Models are evaluated on top-1 accuracy, along with two interpretability metrics: Relevance Mass Accuracy and Relevance Rank Accuracy. These metrics are computed via Grad-CAM heatmaps benchmarked against ground-truth object masks. Our results show that top-1 accuracy ranges from 71.6% to 74.0%. For Grad-CAM, RMA ranges from 7.7% to 9.7% and RRA from 7.3% to 10.1%; for Guided Grad-CAM, RMA ranges from 16.1% to 18.6% and RRA from 15.9% to 21.5%. Interpretability proves far more sensitive to the soft-label weight than to the temperature: keeping the student anchored to hard labels preserves both accuracy and coarse localization, whereas weighting the teacher heavily degrades both. Fine-grained attribution, however, fell below the undistilled baseline in every configuration tested, indicating that logit distillation transmits where a model attends more readily than the pixel-level structure of that attention. We evaluate 12 cross-architecture combinations of convolutional and transformer-based models, revealing that the inheritance of fine-grained spatial reasoning is fundamentally bottlenecked by the student's intrinsic structural biases. To our knowledge, this is the first application of this interpretability-aware evaluation framework - previously used for neural network pruning - to KD.
展开完整摘要收起摘要↓
Deploying efficient neural networks is essential in resource-constrained environments, yet compact models often sacrifice interpretability - a critical in safety-critical domains such as autonomous driving and medicine. This study investigates whether Knowledge Distillation transfers the spatial feature attribution of a large teacher network to a compact student. To assess the influence of the KD scheme on interpretability, we distill a ResNet-152 teacher into a ResNet-34 student on ImageNet-1K across five configurations by systematically varying the distillation temperature and soft-label loss weight. Models are evaluated on top-1 accuracy, along with two interpretability metrics: Relevance Mass Accuracy and Relevance Rank Accuracy. These metrics are computed via Grad-CAM heatmaps benchmarked against ground-truth object masks. Our results show that top-1 accuracy ranges from 71.6% to 74.0%. For Grad-CAM, RMA ranges from 7.7% to 9.7% and RRA from 7.3% to 10.1%; for Guided Grad-CAM, RMA ranges from 16.1% to 18.6% and RRA from 15.9% to 21.5%. Interpretability proves far more sensitive to the soft-label weight than to the temperature: keeping the student anchored to hard labels preserves both accuracy and coarse localization, whereas weighting the teacher heavily degrades both. Fine-grained attribution, however, fell below the undistilled baseline in every configuration tested, indicating that logit distillation transmits where a model attends more readily than the pixel-level structure of that attention. We evaluate 12 cross-architecture combinations of convolutional and transformer-based models, revealing that the inheritance of fine-grained spatial reasoning is fundamentally bottlenecked by the student's intrinsic structural biases. To our knowledge, this is the first application of this interpretability-aware evaluation framework - previously used for neural network pruning - to KD.
作者Mikhail Sannikov, Ilya Mikhalchuk, Konstantin Gubernatorov, Petr Kovalev, Ogunwoye Faith Oluwatobi, Dzmitry Tsetserukou
Contact-rich manipulation requires estimating forces, slip and contact geometry that can remain ambiguous in scene images. Optical tactile sensors provide both visual observations of the contact surface and mechanical measurements, yet learning from these signals raises two challenges: representing contact beyond appearance and transferring its benefits to a policy that does not require fingertip observations at deployment. We introduce HapticWAM, a world-action model that combines heterogeneous tactile encoding, structured contact prediction and teacher-student distillation. Its teacher encodes gel images together with deformation, shear, distributed forces, resultant wrench and derived contact state into a frozen video backbone. Rather than predicting tactile pixels alone, the model jointly generates actions and a contact package describing future events and mechanics. Anticipatory Contact Coupling uses the previously imagined package to condition attention, preserving a contact-related input when direct tactile observations are unavailable. Haptic-Imagination Distillation transfers both contact futures and action predictions to a student that retains the generative contact head but removes its fingertip input branches. On a real-world setup, across three contact-rich pick-and-place tasks, HapticWAM Student achieves a 77% per-task mean success rate (41 of 50 starts, 82% pooled), reaching 95% on one of the tasks, outperforming the evaluated teacher and baseline configurations.
展开完整摘要收起摘要↓
Contact-rich manipulation requires estimating forces, slip and contact geometry that can remain ambiguous in scene images. Optical tactile sensors provide both visual observations of the contact surface and mechanical measurements, yet learning from these signals raises two challenges: representing contact beyond appearance and transferring its benefits to a policy that does not require fingertip observations at deployment. We introduce HapticWAM, a world-action model that combines heterogeneous tactile encoding, structured contact prediction and teacher-student distillation. Its teacher encodes gel images together with deformation, shear, distributed forces, resultant wrench and derived contact state into a frozen video backbone. Rather than predicting tactile pixels alone, the model jointly generates actions and a contact package describing future events and mechanics. Anticipatory Contact Coupling uses the previously imagined package to condition attention, preserving a contact-related input when direct tactile observations are unavailable. Haptic-Imagination Distillation transfers both contact futures and action predictions to a student that retains the generative contact head but removes its fingertip input branches. On a real-world setup, across three contact-rich pick-and-place tasks, HapticWAM Student achieves a 77% per-task mean success rate (41 of 50 starts, 82% pooled), reaching 95% on one of the tasks, outperforming the evaluated teacher and baseline configurations.
Data acquisition (DAQ) systems at future particle physics experiments stand to benefit from the extremes of AI/ML development: large-scale foundation models can enhance the performance of feature extraction algorithms, and small-scale on-detector deployments can enable real-time intelligent data handling. This work provides the first fine-tuning of an industrial foundation model for particle physics DAQ. Starting from the backbone of Google Research's TimesFM (Time Series Foundation Model), we demonstrate fine-tuning on real-time regression tasks for drift chamber trackers and dual-readout calorimeters. Furthermore, the fine-tuned TimesFM model is distilled into a student and co-designed with FPGA implementation to enable these models to run in real-time at future colliders. The fine-tuned distillations meet or exceed the performance of previously published AI/ML solutions for each task. Further, the pipeline of distillation and model compression from TimesFM is generic and can be easily adapted to a variety of 1D waveform tasks across domains.
展开完整摘要收起摘要↓
Data acquisition (DAQ) systems at future particle physics experiments stand to benefit from the extremes of AI/ML development: large-scale foundation models can enhance the performance of feature extraction algorithms, and small-scale on-detector deployments can enable real-time intelligent data handling. This work provides the first fine-tuning of an industrial foundation model for particle physics DAQ. Starting from the backbone of Google Research's TimesFM (Time Series Foundation Model), we demonstrate fine-tuning on real-time regression tasks for drift chamber trackers and dual-readout calorimeters. Furthermore, the fine-tuned TimesFM model is distilled into a student and co-designed with FPGA implementation to enable these models to run in real-time at future colliders. The fine-tuned distillations meet or exceed the performance of previously published AI/ML solutions for each task. Further, the pipeline of distillation and model compression from TimesFM is generic and can be easily adapted to a variety of 1D waveform tasks across domains.
Multi-teacher on-policy distillation allows a student to learn from complementary specialists on its own trajectories. Domain-routed approaches, however, select one teacher per example and keep it fixed throughout the response. This design both depends on domain labels that mixed training corpora often lack and cannot adapt teacher selection when the expertise required changes within a trajectory. We observe that each specialist deviates more from a shared reference on in-domain prompts than on out-of-domain prompts, on average. Based on this observation, we propose TrustMOPD, which replaces example-level teacher selection with label-free, token-level supervision allocation. At each student-generated prefix, TrustMOPD measures this displacement in next-token preferences, calibrates its magnitude across teachers, and uses the resulting scores as proxies for local reliability to weight teacher-specific distillation losses. Evaluated across mathematics, code, and instruction following, TrustMOPD closes 91.5% and 98.0% of the overall-score gap between the initial student and oracle-routed teachers when trained on \textsc{SingleCap} and \textsc{MultiCap}, respectively, compared with 54.4% and 54.5% for the strongest label-free baseline in each setting. On \textsc{SingleCap}, it approaches label-based MOPD without using domain labels.
展开完整摘要收起摘要↓
Multi-teacher on-policy distillation allows a student to learn from complementary specialists on its own trajectories. Domain-routed approaches, however, select one teacher per example and keep it fixed throughout the response. This design both depends on domain labels that mixed training corpora often lack and cannot adapt teacher selection when the expertise required changes within a trajectory. We observe that each specialist deviates more from a shared reference on in-domain prompts than on out-of-domain prompts, on average. Based on this observation, we propose TrustMOPD, which replaces example-level teacher selection with label-free, token-level supervision allocation. At each student-generated prefix, TrustMOPD measures this displacement in next-token preferences, calibrates its magnitude across teachers, and uses the resulting scores as proxies for local reliability to weight teacher-specific distillation losses. Evaluated across mathematics, code, and instruction following, TrustMOPD closes 91.5% and 98.0% of the overall-score gap between the initial student and oracle-routed teachers when trained on \textsc{SingleCap} and \textsc{MultiCap}, respectively, compared with 54.4% and 54.5% for the strongest label-free baseline in each setting. On \textsc{SingleCap}, it approaches label-based MOPD without using domain labels.
High-dimensional count data are common in scientific applications, but most diffusion and flow models are designed for continuous or categorical data, and generation often requires many sequential model evaluations. We propose Count Flow Map, a generative model that learns finite-time transitions directly in count space for one- or few-step generation. Our model directly learns stochastic transitions over finite time intervals, using Poisson births and Binomial deaths to preserve nonnegative integer counts without a predefined maximum. These transition models are trained to match the underlying local birth--death dynamics and to maintain consistency across step sizes. We characterize the connection between local dynamics and finite-time transition consistency and derive a bound on the generation error. After validating Count Flow Map in several simulations, including a high-dimensional, high-count setting, we apply it to single-cell drug perturbation prediction and neural population forecasting, where it captures perturbation responses and supports forecasts of high-activity events with only one or a few model evaluations. Together, these experiments demonstrate that Count Flow Map enables high-quality generation directly in count space across inference budgets, from one-step to few-step generation, using a single trained model.
展开完整摘要收起摘要↓
High-dimensional count data are common in scientific applications, but most diffusion and flow models are designed for continuous or categorical data, and generation often requires many sequential model evaluations. We propose Count Flow Map, a generative model that learns finite-time transitions directly in count space for one- or few-step generation. Our model directly learns stochastic transitions over finite time intervals, using Poisson births and Binomial deaths to preserve nonnegative integer counts without a predefined maximum. These transition models are trained to match the underlying local birth--death dynamics and to maintain consistency across step sizes. We characterize the connection between local dynamics and finite-time transition consistency and derive a bound on the generation error. After validating Count Flow Map in several simulations, including a high-dimensional, high-count setting, we apply it to single-cell drug perturbation prediction and neural population forecasting, where it captures perturbation responses and supports forecasts of high-activity events with only one or a few model evaluations. Together, these experiments demonstrate that Count Flow Map enables high-quality generation directly in count space across inference budgets, from one-step to few-step generation, using a single trained model.
Single-image 3D object generation can now produce high-fidelity assets, yet accurately placing them into a coherent scene layout remains an open challenge. A central difficulty lies in how object layout is represented. Holistic methods absorb placement into a scene-level generation process, sacrificing object-level detail. Compositional methods preserve object fidelity by decoupling geometry from layout, but typically parameterize layout as sparse, unbounded pose variables that are difficult to learn and generalize poorly under scarce scene-level supervision. We present Mira-Scene, a compositional 3D scene reconstruction framework that replaces sparse pose regression with dense, bounded correspondence recovery. At its core is the Canonical Coordinate Map (CCM), a pixel-aligned field that maps each visible object pixel to a surface coordinate in the object's bounded canonical space. When paired with a scene-space Point Cloud Map (PCM) from monocular geometry estimation, CCM induces dense canonical-to-scene correspondences from which object transformations are recovered through robust geometric alignment. Because CCM operates in bounded canonical space, it provides a stable prediction target that can be trained from scalable object-level 3D data without requiring scene-level layout annotations. Mira-Scene further introduces a multimodal diffusion transformer that jointly generates object geometry and CCMs, using modality-specific expert streams with shared attention and positional encoding to promote geometry-layout consistency. Experiments on indoor, outdoor, synthetic, and in-the-wild scenes show that Mira-Scene substantially outperforms strong baselines in layout accuracy, achieving relative gains of 39.8% in 3D-IoU and 16.5% in 2D-IoU over SAM3D, using limited open-source training data.
展开完整摘要收起摘要↓
Single-image 3D object generation can now produce high-fidelity assets, yet accurately placing them into a coherent scene layout remains an open challenge. A central difficulty lies in how object layout is represented. Holistic methods absorb placement into a scene-level generation process, sacrificing object-level detail. Compositional methods preserve object fidelity by decoupling geometry from layout, but typically parameterize layout as sparse, unbounded pose variables that are difficult to learn and generalize poorly under scarce scene-level supervision. We present Mira-Scene, a compositional 3D scene reconstruction framework that replaces sparse pose regression with dense, bounded correspondence recovery. At its core is the Canonical Coordinate Map (CCM), a pixel-aligned field that maps each visible object pixel to a surface coordinate in the object's bounded canonical space. When paired with a scene-space Point Cloud Map (PCM) from monocular geometry estimation, CCM induces dense canonical-to-scene correspondences from which object transformations are recovered through robust geometric alignment. Because CCM operates in bounded canonical space, it provides a stable prediction target that can be trained from scalable object-level 3D data without requiring scene-level layout annotations. Mira-Scene further introduces a multimodal diffusion transformer that jointly generates object geometry and CCMs, using modality-specific expert streams with shared attention and positional encoding to promote geometry-layout consistency. Experiments on indoor, outdoor, synthetic, and in-the-wild scenes show that Mira-Scene substantially outperforms strong baselines in layout accuracy, achieving relative gains of 39.8% in 3D-IoU and 16.5% in 2D-IoU over SAM3D, using limited open-source training data.
Robots operating in human-centered environments are typically designed to execute explicit instructions, and most robot-learning datasets likewise pair observations with task instructions or low-level actions. Although recent work has begun to explore proactive embodied assistance, existing resources target different settings and action levels, leaving real-world human-centered multimodal decision-making underexplored. We formulate this problem as Proactive Robot Action Reasoning (ProRobo), an upstream cognitive decision problem in which a robot must determine which action to take based on multimodal human and environmental cues without explicit action instructions. To support ProRobo, we introduce ProAction, a real-world multimodal dataset containing 10K samples of visual observations, audio signals, and text inputs across 12 daily-life scenarios in five common scenes. To construct cognitively grounded high-level action supervision, we develop a two-stage human-in-the-loop pipeline that combines appraisal-guided candidate generation with Affective Theory-of-Mind-guided human refinement, explicitly incorporating contextual judgment about human states, urgency, feasibility, and potential risk into action annotation. Based on this supervision, we benchmark representative Multimodal Large Language Models (MLLMs) and introduce MMC2Act, a reference model that implicitly learns the mapping from multimodal observations to cognitively grounded high-level actions. Experiments across modality settings, subject-disjoint generalization, cross-dataset transfer, and human evaluation show that general-purpose MLLMs struggle with proactively reasoning high-level actions from multimodal cues, whereas training on ProAction substantially improves performance.
展开完整摘要收起摘要↓
Robots operating in human-centered environments are typically designed to execute explicit instructions, and most robot-learning datasets likewise pair observations with task instructions or low-level actions. Although recent work has begun to explore proactive embodied assistance, existing resources target different settings and action levels, leaving real-world human-centered multimodal decision-making underexplored. We formulate this problem as Proactive Robot Action Reasoning (ProRobo), an upstream cognitive decision problem in which a robot must determine which action to take based on multimodal human and environmental cues without explicit action instructions. To support ProRobo, we introduce ProAction, a real-world multimodal dataset containing 10K samples of visual observations, audio signals, and text inputs across 12 daily-life scenarios in five common scenes. To construct cognitively grounded high-level action supervision, we develop a two-stage human-in-the-loop pipeline that combines appraisal-guided candidate generation with Affective Theory-of-Mind-guided human refinement, explicitly incorporating contextual judgment about human states, urgency, feasibility, and potential risk into action annotation. Based on this supervision, we benchmark representative Multimodal Large Language Models (MLLMs) and introduce MMC2Act, a reference model that implicitly learns the mapping from multimodal observations to cognitively grounded high-level actions. Experiments across modality settings, subject-disjoint generalization, cross-dataset transfer, and human evaluation show that general-purpose MLLMs struggle with proactively reasoning high-level actions from multimodal cues, whereas training on ProAction substantially improves performance.
Point cloud completion aims to infer a complete 3D shape from a partial point cloud and serves as a fundamental building block for downstream tasks such as reconstruction, editing, and simulation. Despite the recent progress, existing learning-based methods often implicitly rely on access to the ground-truth shape scale (GT-scale) during both training- and testing-time normalization, assuming privileged information that is unavailable in real-world inference. This hidden assumption limits practical deployment and can lead to severe completion artifacts, e.g., over- or under-completion and nested shells, once the oracle GT-scale cue is removed. We observe that the recent foundation image generation models exhibit a strong capability of understanding objects and geometries, and producing multi-view consistent renderings, making them promising priors for GT-scale-free 3D completion. Motivated by this insight, we propose ScaleBlind, a novel framework that leverages foundation-model-based image completion to recover global scale directly from partial inputs and then faithfully produces the 3D completion. Specifically, ScaleBlind dreams out complete multi-view appearances from rendered partial views, lifts the inferred missing regions back into 3D to obtain a geometry-aware coarse completion, and further refines it via a powerful cross-modal fusion network with the original partial point cloud. By harnessing 2D foundation priors, our method eliminates the need for accessing GT-scale information at inference. Moreover, it provides a principled bridge between 2D generative priors and 3D point cloud completion. Extensive experiments demonstrate the superiority of our framework, making ScaleBlind the new state-of-the-art for the point cloud completion task.
展开完整摘要收起摘要↓
Point cloud completion aims to infer a complete 3D shape from a partial point cloud and serves as a fundamental building block for downstream tasks such as reconstruction, editing, and simulation. Despite the recent progress, existing learning-based methods often implicitly rely on access to the ground-truth shape scale (GT-scale) during both training- and testing-time normalization, assuming privileged information that is unavailable in real-world inference. This hidden assumption limits practical deployment and can lead to severe completion artifacts, e.g., over- or under-completion and nested shells, once the oracle GT-scale cue is removed. We observe that the recent foundation image generation models exhibit a strong capability of understanding objects and geometries, and producing multi-view consistent renderings, making them promising priors for GT-scale-free 3D completion. Motivated by this insight, we propose ScaleBlind, a novel framework that leverages foundation-model-based image completion to recover global scale directly from partial inputs and then faithfully produces the 3D completion. Specifically, ScaleBlind dreams out complete multi-view appearances from rendered partial views, lifts the inferred missing regions back into 3D to obtain a geometry-aware coarse completion, and further refines it via a powerful cross-modal fusion network with the original partial point cloud. By harnessing 2D foundation priors, our method eliminates the need for accessing GT-scale information at inference. Moreover, it provides a principled bridge between 2D generative priors and 3D point cloud completion. Extensive experiments demonstrate the superiority of our framework, making ScaleBlind the new state-of-the-art for the point cloud completion task.
作者Yueyan Li, Haibo Wang, Caixia Yuan, Xiaojie Wang
Despite impressive visual quality, state-of-the-art video diffusion models often generate content that violates real-world physical laws. While existing solutions rely on external priors or specialized data, we investigate the root cause by exploring the internal mechanisms of these models. Specifically, we present the first interpretability study on the ''motion planning'' process of text-to-video diffusion models, revealing how motion trajectories form during early denoising stages. Building upon the ''first shape, then details'' finding, we combine cross-attention trajectory patterns with causal head contributions to identify a specific subset of attention heads driving motion planning. Further, our self-attention analysis shows that Rotary Position Embedding (RoPE) induces excessive spatial attention decay. This causes early candidate regions to prematurely lock into physically implausible positions, suppressing reasonable trajectories in adjacent frames and triggering generation failure modes. To address this fundamental flaw, we propose a lightweight architectural modification that scales the frequency of RoPE across different denoising steps. This strategy reduces excessive attention decay, helping the model explore better candidate regions to establish coherent physical motion. Finally, training-free and training-based experiments confirm the effectiveness of our approach in enhancing the physical commonsense of generated videos.
展开完整摘要收起摘要↓
Despite impressive visual quality, state-of-the-art video diffusion models often generate content that violates real-world physical laws. While existing solutions rely on external priors or specialized data, we investigate the root cause by exploring the internal mechanisms of these models. Specifically, we present the first interpretability study on the ''motion planning'' process of text-to-video diffusion models, revealing how motion trajectories form during early denoising stages. Building upon the ''first shape, then details'' finding, we combine cross-attention trajectory patterns with causal head contributions to identify a specific subset of attention heads driving motion planning. Further, our self-attention analysis shows that Rotary Position Embedding (RoPE) induces excessive spatial attention decay. This causes early candidate regions to prematurely lock into physically implausible positions, suppressing reasonable trajectories in adjacent frames and triggering generation failure modes. To address this fundamental flaw, we propose a lightweight architectural modification that scales the frequency of RoPE across different denoising steps. This strategy reduces excessive attention decay, helping the model explore better candidate regions to establish coherent physical motion. Finally, training-free and training-based experiments confirm the effectiveness of our approach in enhancing the physical commonsense of generated videos.
As generative AI becomes increasingly used in anime-style image creation, distinguishing human-drawn, AI-inpainted, and text-to-image images is important for copyright attribution, visual provenance, and content governance. Existing AI-generated image detectors mainly target real-world photographs and often overlook anime-specific cues such as flat coloring, exaggerated structures, and artistic line control. To address this gap, we propose AniPrO, a multi-dimensional description-enhanced framework for interpretable anime image provenance. Built upon AnimeDL-2M, AniPrO contains 15,000 balanced samples from a 35,000-image candidate pool, covering Real, Inpainting, and Text2Image categories with structured five-dimensional descriptions. We further introduce AniPrO-SFD-Bench and AniPrO-MFR-Bench to evaluate provenance detection from statistical feature discrimination and multimodal fusion reasoning perspectives. Experiments show that structured semantic guidance reveals systematic AI-generation biases, such as the gap between global visual plausibility and local detail coherence, and improves the detection of challenging inpainting samples. The dataset and code will be released at: https://github.com/YAN-LIU05/AniPrO.
展开完整摘要收起摘要↓
As generative AI becomes increasingly used in anime-style image creation, distinguishing human-drawn, AI-inpainted, and text-to-image images is important for copyright attribution, visual provenance, and content governance. Existing AI-generated image detectors mainly target real-world photographs and often overlook anime-specific cues such as flat coloring, exaggerated structures, and artistic line control. To address this gap, we propose AniPrO, a multi-dimensional description-enhanced framework for interpretable anime image provenance. Built upon AnimeDL-2M, AniPrO contains 15,000 balanced samples from a 35,000-image candidate pool, covering Real, Inpainting, and Text2Image categories with structured five-dimensional descriptions. We further introduce AniPrO-SFD-Bench and AniPrO-MFR-Bench to evaluate provenance detection from statistical feature discrimination and multimodal fusion reasoning perspectives. Experiments show that structured semantic guidance reveals systematic AI-generation biases, such as the gap between global visual plausibility and local detail coherence, and improves the detection of challenging inpainting samples. The dataset and code will be released at: https://github.com/YAN-LIU05/AniPrO.
作者Alexander Christie, Miguel Moscoso, Alexei Novikov, George Papanicolaou, Chrysoula Tsogka
Dictionary learning seeks to recover an unknown dictionary $A$ from observations ${\bf y}_i = A{\bf x}_i$ with sparse coefficient vectors ${\bf x}_i$. We introduce the Iterative Atom Refinement (IAR) algorithm, a simple procedure for recovering individual dictionary atoms. Starting from a random direction, IAR repeatedly selects the observations most strongly correlated with the current iterate and updates the direction by averaging the selected data. Our main contribution is a rigorous convergence theory of IAR. Using high-dimensional probabilistic estimates and a novel monotonicity principle for atom-selection probabilities, we show that a small initial advantage of one atom is amplified until that atom is isolated. Under our model assumptions, IAR identifies a generating atom after only three refinement steps. Numerical experiments support the theory and show that the resulting dynamics accurately capture the behavior observed in dictionary refinement.
展开完整摘要收起摘要↓
Dictionary learning seeks to recover an unknown dictionary $A$ from observations ${\bf y}_i = A{\bf x}_i$ with sparse coefficient vectors ${\bf x}_i$. We introduce the Iterative Atom Refinement (IAR) algorithm, a simple procedure for recovering individual dictionary atoms. Starting from a random direction, IAR repeatedly selects the observations most strongly correlated with the current iterate and updates the direction by averaging the selected data. Our main contribution is a rigorous convergence theory of IAR. Using high-dimensional probabilistic estimates and a novel monotonicity principle for atom-selection probabilities, we show that a small initial advantage of one atom is amplified until that atom is isolated. Under our model assumptions, IAR identifies a generating atom after only three refinement steps. Numerical experiments support the theory and show that the resulting dynamics accurately capture the behavior observed in dictionary refinement.
作者Wenhong Huang, Jianwei Fei, Benedetta Tondi, Bin Ma, Fangjun Huang
The rapid development of video generative models (VGMs) has enabled the generation of highly realistic synthetic videos, raising concerns about the intellectual property rights (IPR) of these models. In particular, two closely related forensic tasks remain largely unaddressed: synthetic video verification (determining whether a video was generated by a protected VGM) and model ownership verification (determining whether a suspect VGM is an unauthorized copy of a protected VGM). In this paper, we propose a new in-generation watermarking scheme that can address the two verification tasks. First, a novel video watermarking network named VidMark is presented, which incorporates a two-scale discrete wavelet transform (DWT) decomposition and a global temporal attention block (GTAB) to enhance watermark robustness and imperceptibility. Second, we present a decoder-guided fine-tuning procedure. By leveraging the frozen VidMark decoder, this process enables VGMs to synthesize videos carrying an imperceptible, robust, and model-specific watermark. Finally, two verification frameworks are established to perform synthetic video verification and model ownership verification. Extensive experiments on representative VGMs demonstrate that the proposed scheme achieves over 99% watermark extraction accuracy and 100% verification accuracy on both tasks, with negligible impact on video generation quality. Furthermore, the watermarks exhibit strong robustness against a comprehensive range of video-level and model-level attacks.
展开完整摘要收起摘要↓
The rapid development of video generative models (VGMs) has enabled the generation of highly realistic synthetic videos, raising concerns about the intellectual property rights (IPR) of these models. In particular, two closely related forensic tasks remain largely unaddressed: synthetic video verification (determining whether a video was generated by a protected VGM) and model ownership verification (determining whether a suspect VGM is an unauthorized copy of a protected VGM). In this paper, we propose a new in-generation watermarking scheme that can address the two verification tasks. First, a novel video watermarking network named VidMark is presented, which incorporates a two-scale discrete wavelet transform (DWT) decomposition and a global temporal attention block (GTAB) to enhance watermark robustness and imperceptibility. Second, we present a decoder-guided fine-tuning procedure. By leveraging the frozen VidMark decoder, this process enables VGMs to synthesize videos carrying an imperceptible, robust, and model-specific watermark. Finally, two verification frameworks are established to perform synthetic video verification and model ownership verification. Extensive experiments on representative VGMs demonstrate that the proposed scheme achieves over 99% watermark extraction accuracy and 100% verification accuracy on both tasks, with negligible impact on video generation quality. Furthermore, the watermarks exhibit strong robustness against a comprehensive range of video-level and model-level attacks.
作者Md Sayed Tanveer, Mohammed A. Mostajo-Radji, Ge Wang
Generative models of neural activity could help characterize tissue dynamics, compare experimental conditions, and simulate population activity for applications ranging from disease and drug-response studies to closed-loop experimentation. Existing approaches, however, typically assume a fixed set of sorted neurons, whereas high-density microelectrode arrays produce extremely sparse, array-wide binary spike volumes in which the observed subset of electrodes varies across assays. We introduce a discrete generative model that represents this activity using a shared vocabulary of spatiotemporal motifs. A residual vector-quantized autoencoder learns the motif vocabulary, while a factorized masked transformer predicts where activity occurs and which motif appears at each active location. We evaluate the model on 31 assays spanning human brain organoids and acute ex vivo human hippocampal tissue. The learned motifs are broadly reused: assay identity explains only $9%$ of the entropy in motif use, and motif overlap across tissue types is comparable to overlap within them. When representation quality is evaluated independently of the generative prior, our approach achieves $5.2\times$ the voxel-level reconstruction average precision of a matched flat tokenizer. For masked completion and free generation, the full model achieves $1.4$--$2.6\times$ the site-level average precision of the matched generative baseline and outperforms it across all four families of generation metrics. These results establish a compact, reusable representation for array-wide spiking activity without learned assay-specific parameters, providing a scalable foundation for generative modeling across diverse neural preparations.
展开完整摘要收起摘要↓
Generative models of neural activity could help characterize tissue dynamics, compare experimental conditions, and simulate population activity for applications ranging from disease and drug-response studies to closed-loop experimentation. Existing approaches, however, typically assume a fixed set of sorted neurons, whereas high-density microelectrode arrays produce extremely sparse, array-wide binary spike volumes in which the observed subset of electrodes varies across assays. We introduce a discrete generative model that represents this activity using a shared vocabulary of spatiotemporal motifs. A residual vector-quantized autoencoder learns the motif vocabulary, while a factorized masked transformer predicts where activity occurs and which motif appears at each active location. We evaluate the model on 31 assays spanning human brain organoids and acute ex vivo human hippocampal tissue. The learned motifs are broadly reused: assay identity explains only $9%$ of the entropy in motif use, and motif overlap across tissue types is comparable to overlap within them. When representation quality is evaluated independently of the generative prior, our approach achieves $5.2\times$ the voxel-level reconstruction average precision of a matched flat tokenizer. For masked completion and free generation, the full model achieves $1.4$--$2.6\times$ the site-level average precision of the matched generative baseline and outperforms it across all four families of generation metrics. These results establish a compact, reusable representation for array-wide spiking activity without learned assay-specific parameters, providing a scalable foundation for generative modeling across diverse neural preparations.
Diffusion models generate high-quality images, yet often violate the physical laws governing mirror reflections. Reflections often suffer from geometric aberrations, including positional offsets, directional misalignment, proportional imbalance, and structural distortion. These failures remain evident even in contemporary state-of-the-art generative systems. Existing methods itigate this problem through synthetic data scaling or auxiliary depth conditioning, yet their merely reliance on latent-space noise reconstruction losses as implicit supervision prevents direct enforcement of reflection-specific geometric and perceptual constraints. To bridge this gap, we present PhysReflect, a geometry and perception guided diffusion framework that decodes the predicted clean latent into pixel space at each training step and applies annealed supervision through two complementary differentiable objectives. The Geometric Loss enforces mirror-induced spatial consistency through sparse epipolar correspondence and dense boundary projection alignment, where a SAM2-based TwinTrack mechanism provides stable in-mirror localization for boundary-aware supervision. The Perceptual Loss preserves reflected appearance by combining Semantic Consistency Loss, which maintains reflected identity and appearance via DINOv2 features, and Lighting Consistency Loss, which regularizes depth, surface-normal, and illumination coherence under monocular geometry priors. Experiments on synthetic and real-world benchmarks show that PhysReflect outperforms prior mirror-reflection methods in geometric, perceptual, and physical-plausibility metrics, as well as qualitative visual results.
展开完整摘要收起摘要↓
Diffusion models generate high-quality images, yet often violate the physical laws governing mirror reflections. Reflections often suffer from geometric aberrations, including positional offsets, directional misalignment, proportional imbalance, and structural distortion. These failures remain evident even in contemporary state-of-the-art generative systems. Existing methods itigate this problem through synthetic data scaling or auxiliary depth conditioning, yet their merely reliance on latent-space noise reconstruction losses as implicit supervision prevents direct enforcement of reflection-specific geometric and perceptual constraints. To bridge this gap, we present PhysReflect, a geometry and perception guided diffusion framework that decodes the predicted clean latent into pixel space at each training step and applies annealed supervision through two complementary differentiable objectives. The Geometric Loss enforces mirror-induced spatial consistency through sparse epipolar correspondence and dense boundary projection alignment, where a SAM2-based TwinTrack mechanism provides stable in-mirror localization for boundary-aware supervision. The Perceptual Loss preserves reflected appearance by combining Semantic Consistency Loss, which maintains reflected identity and appearance via DINOv2 features, and Lighting Consistency Loss, which regularizes depth, surface-normal, and illumination coherence under monocular geometry priors. Experiments on synthetic and real-world benchmarks show that PhysReflect outperforms prior mirror-reflection methods in geometric, perceptual, and physical-plausibility metrics, as well as qualitative visual results.
Modern conditional generative models face significant challenges when learning complex covariate dependencies. While sufficient dimension reduction (SDR) provides a principled approach to compress these dependencies, traditional SDR frameworks were not formulated for conditional generation. To bridge this gap, we propose Belted Engression, a unified and architecturally parameter-efficient framework for generative distributional regression. Our approach establishes an end-to-end compress-then-generate paradigm driven by sufficient representation learning, embedding a structural bottleneck into the generative architecture. Theoretically, we prove that the standard SDR condition is equivalent to a law-preserving generative factorization, which is achieved at the global optimum of the population Belted Engression objective. Furthermore, by uncovering a localized Bernstein-type control for the energy-score loss, we establish finite-sample convergence rates that are sharper than those of existing results. We also prove that this belted architecture is strictly smaller, operating with an asymptotically vanishing parameter count relative to the unstructured baseline. Extensive simulations and real-world applications demonstrate that Belted Engression achieves superior distributional prediction and SDR recovery with fewer trainable parameters.
展开完整摘要收起摘要↓
Modern conditional generative models face significant challenges when learning complex covariate dependencies. While sufficient dimension reduction (SDR) provides a principled approach to compress these dependencies, traditional SDR frameworks were not formulated for conditional generation. To bridge this gap, we propose Belted Engression, a unified and architecturally parameter-efficient framework for generative distributional regression. Our approach establishes an end-to-end compress-then-generate paradigm driven by sufficient representation learning, embedding a structural bottleneck into the generative architecture. Theoretically, we prove that the standard SDR condition is equivalent to a law-preserving generative factorization, which is achieved at the global optimum of the population Belted Engression objective. Furthermore, by uncovering a localized Bernstein-type control for the energy-score loss, we establish finite-sample convergence rates that are sharper than those of existing results. We also prove that this belted architecture is strictly smaller, operating with an asymptotically vanishing parameter count relative to the unstructured baseline. Extensive simulations and real-world applications demonstrate that Belted Engression achieves superior distributional prediction and SDR recovery with fewer trainable parameters.
Temporal logic is a formal language for reasoning about system behaviors over time. Signal temporal logic (STL), in particular, has been used to encode spatio-temporal requirements for control synthesis in multi-agent systems, often under the assumption that agents are cooperative and their dynamics are known. However, real-world multi-agent applications, such as autonomous driving, typically involve stochastic and uncontrollable agents. Recent work explored robust control with worst-case or probabilistic formulations, but remains limited in that it either (1) certifies strict satisfaction of STL constraints without addressing feasibility recovery, or (2) relaxes infeasible constraints with ego-centric objectives. In this paper, we propose a model predictive control (MPC) framework that treats feasibility repair as a Pareto optimization problem to explicitly characterize tradeoffs among agent objectives. We further provide a probabilistic certificate on STL violation rate to formally quantify uncertainty under stochastic and uncontrollable agents. The proposed framework is evaluated on two autonomous driving scenarios. Results show that the framework recovers feasible control with demonstrated safe behaviors.
展开完整摘要收起摘要↓
Temporal logic is a formal language for reasoning about system behaviors over time. Signal temporal logic (STL), in particular, has been used to encode spatio-temporal requirements for control synthesis in multi-agent systems, often under the assumption that agents are cooperative and their dynamics are known. However, real-world multi-agent applications, such as autonomous driving, typically involve stochastic and uncontrollable agents. Recent work explored robust control with worst-case or probabilistic formulations, but remains limited in that it either (1) certifies strict satisfaction of STL constraints without addressing feasibility recovery, or (2) relaxes infeasible constraints with ego-centric objectives. In this paper, we propose a model predictive control (MPC) framework that treats feasibility repair as a Pareto optimization problem to explicitly characterize tradeoffs among agent objectives. We further provide a probabilistic certificate on STL violation rate to formally quantify uncertainty under stochastic and uncontrollable agents. The proposed framework is evaluated on two autonomous driving scenarios. Results show that the framework recovers feasible control with demonstrated safe behaviors.
Memory systems are becoming a core component of LLM agents, but constructing and maintaining memory remains expensive because it relies on repeated calls to large proprietary language models. This cost creates a major barrier to deploying memory-enhanced agents at scale. In this paper, we present Pseudo Self-Distillation (PSD), a framework that enables small language models (SLMs) to construct hierarchical memory representations by distilling behavior from a strong black-box oracle through a multi-stage training pipeline. Standard distillation methods require access to teacher logits or hidden states, which closed models do not expose. Unlike conventional self-distillation settings, where supervision is derived from a model's own predictions, sampled rollouts, or aggregated outputs, PSD enables a single-model distillation setup while channeling external oracle knowledge through the prompt. PSD uses a single small model in two roles: a teacher that sees a privileged prompt containing the oracle's answer as reference context, and a student that sees only the task prompt. The student learns to reproduce the teacher's output distribution, absorbing oracle-guided behavior into its own weights without accessing the oracle's internals. On LoCoMo, PSD-trained Qwen3-0.6B, 1.7B, and 4B match or exceed GPT-4.1-mini on downstream retrieval at a fraction of the deployment cost, with off-policy PSD achieving the strongest results across most conditions. We further show that this memory-construction capability transfers out-of-distribution to LongMemEval, despite the students being trained exclusively on LoCoMo with no exposure to LongMemEval data.
展开完整摘要收起摘要↓
Memory systems are becoming a core component of LLM agents, but constructing and maintaining memory remains expensive because it relies on repeated calls to large proprietary language models. This cost creates a major barrier to deploying memory-enhanced agents at scale. In this paper, we present Pseudo Self-Distillation (PSD), a framework that enables small language models (SLMs) to construct hierarchical memory representations by distilling behavior from a strong black-box oracle through a multi-stage training pipeline. Standard distillation methods require access to teacher logits or hidden states, which closed models do not expose. Unlike conventional self-distillation settings, where supervision is derived from a model's own predictions, sampled rollouts, or aggregated outputs, PSD enables a single-model distillation setup while channeling external oracle knowledge through the prompt. PSD uses a single small model in two roles: a teacher that sees a privileged prompt containing the oracle's answer as reference context, and a student that sees only the task prompt. The student learns to reproduce the teacher's output distribution, absorbing oracle-guided behavior into its own weights without accessing the oracle's internals. On LoCoMo, PSD-trained Qwen3-0.6B, 1.7B, and 4B match or exceed GPT-4.1-mini on downstream retrieval at a fraction of the deployment cost, with off-policy PSD achieving the strongest results across most conditions. We further show that this memory-construction capability transfers out-of-distribution to LongMemEval, despite the students being trained exclusively on LoCoMo with no exposure to LongMemEval data.
Long-horizon conversational memory is especially challenging in multi-actor settings, where relevant evidence is distributed across participants and contexts and previously established information may later be revised. We introduce EGMEMORY, which formulates long-horizon multi-actor memory as a searchable state machine that separates persistent message-level evidence from an explicit active state. At write time, adaptive state resolution and an evidence-grounded propose-verify-commit protocol govern how this state evolves. At read time, adaptive evidence navigation iteratively resolves the state and supporting evidence required for a query, using conversational structure to narrow the search space and lexical-semantic relevance to rank candidates. The system operates through prompting and tool use without memory-specific policy training. EGMEMORY achieves 68.2% on GroupMemBench and 77.9% on EverMemBench, outperforming the strongest evaluated baselines by 22.7 and 21.4 percentage points, respectively. It further reaches 73.6% on the dyadic LoCoMo benchmark, demonstrating generalization beyond multi-actor conversations. We will release the codebase upon formal publication.
展开完整摘要收起摘要↓
Long-horizon conversational memory is especially challenging in multi-actor settings, where relevant evidence is distributed across participants and contexts and previously established information may later be revised. We introduce EGMEMORY, which formulates long-horizon multi-actor memory as a searchable state machine that separates persistent message-level evidence from an explicit active state. At write time, adaptive state resolution and an evidence-grounded propose-verify-commit protocol govern how this state evolves. At read time, adaptive evidence navigation iteratively resolves the state and supporting evidence required for a query, using conversational structure to narrow the search space and lexical-semantic relevance to rank candidates. The system operates through prompting and tool use without memory-specific policy training. EGMEMORY achieves 68.2% on GroupMemBench and 77.9% on EverMemBench, outperforming the strongest evaluated baselines by 22.7 and 21.4 percentage points, respectively. It further reaches 73.6% on the dyadic LoCoMo benchmark, demonstrating generalization beyond multi-actor conversations. We will release the codebase upon formal publication.
作者Suzhong Fu, Jingqi Dong, Xuan Ding, Rui Sun, Yiming Yang, Shuguang Cui, Zhen Li
The SYNTAX score is a clinically established tool for assessing anatomical lesion complexity in coronary artery disease and guiding subsequent treatment. However, automated SYNTAX scoring is commonly formulated as a direct regression problem from coronary angiography videos to patient-level scores. In this work, we reformulate SYNTAX scoring as a vessel segment identity-preserving anatomical reasoning problem and propose a hierarchical modeling framework that explicitly aligns learning with the clinical workflow. Our approach maintains vessel segment identity across frames and views, estimates stenosis severity at the segment level, and aggregates evidence hierarchically according to coronary anatomy. Simultaneously, to address the scarcity of domain-specific data, we integrate and complete multiple public coronary angiography datasets, constructing a large-scale resource featuring completed vessel segmentation and derived structural annotations. Experiments demonstrate that vessel segment-level stenosis embedding enhances explanatory power and reduces prediction variability compared to baseline models, with the R^2 score improving by 0.201 and dev STD decreasing by 18.4%. These results highlight the necessity of structure-aligned modeling for reliable and stable automated SYNTAX scoring from multi-view coronary angiography videos. The GitHub link is https://github.com/VersaceSu7/SYNTAX_score_777.
展开完整摘要收起摘要↓
The SYNTAX score is a clinically established tool for assessing anatomical lesion complexity in coronary artery disease and guiding subsequent treatment. However, automated SYNTAX scoring is commonly formulated as a direct regression problem from coronary angiography videos to patient-level scores. In this work, we reformulate SYNTAX scoring as a vessel segment identity-preserving anatomical reasoning problem and propose a hierarchical modeling framework that explicitly aligns learning with the clinical workflow. Our approach maintains vessel segment identity across frames and views, estimates stenosis severity at the segment level, and aggregates evidence hierarchically according to coronary anatomy. Simultaneously, to address the scarcity of domain-specific data, we integrate and complete multiple public coronary angiography datasets, constructing a large-scale resource featuring completed vessel segmentation and derived structural annotations. Experiments demonstrate that vessel segment-level stenosis embedding enhances explanatory power and reduces prediction variability compared to baseline models, with the R^2 score improving by 0.201 and dev STD decreasing by 18.4%. These results highlight the necessity of structure-aligned modeling for reliable and stable automated SYNTAX scoring from multi-view coronary angiography videos. The GitHub link is https://github.com/VersaceSu7/SYNTAX_score_777.
作者Liyang Fan, Yingcheng Shi, Yongbin Li, Chenghao Sun, Xin Chen, Xander Xu, Hu Wei, Shiwen Ni, Min Yang, Jieping Ye
Coding agents operate on real repository coding tasks, and persistent memory systems promise to reuse experience across tasks. Yet existing evaluations do not show whether those systems improve executable repository work. Repository benchmarks test code changes but do not isolate memory, while memory benchmarks score recall without measuring downstream coding outcomes. We introduce VibeMemBench, a benchmark for evaluating memory systems on 111 coding targets from 90 SWE-rebench V2 repositories and 3,634 history trajectories from the target repositories. The targets follow the SWE benchmark style and cover bug fixes, feature requests, interface changes, and configuration work. An agent edits each target codebase under a declared memory condition. Executable tests decide task resolution. Each target is retained only when injected history experience improves its executable outcome in a reference setting, so every target carries a prior experience whose usefulness is verified by execution in that setting. The frozen verified experience is then transferred to five held-out solvers. Direct injection raises observed task resolution on four of them by 1.1 to 4.5 percentage points while lowering agent steps on all five. Yet when four existing memory systems must construct and retrieve experience from the same history, eleven of twelve solver and system pairings fail to exceed the matched memory-off baseline. VibeMemBench exposes the gap between the useful experience that repository history holds and the experience existing memory systems deliver for repository coding tasks.
展开完整摘要收起摘要↓
Coding agents operate on real repository coding tasks, and persistent memory systems promise to reuse experience across tasks. Yet existing evaluations do not show whether those systems improve executable repository work. Repository benchmarks test code changes but do not isolate memory, while memory benchmarks score recall without measuring downstream coding outcomes. We introduce VibeMemBench, a benchmark for evaluating memory systems on 111 coding targets from 90 SWE-rebench V2 repositories and 3,634 history trajectories from the target repositories. The targets follow the SWE benchmark style and cover bug fixes, feature requests, interface changes, and configuration work. An agent edits each target codebase under a declared memory condition. Executable tests decide task resolution. Each target is retained only when injected history experience improves its executable outcome in a reference setting, so every target carries a prior experience whose usefulness is verified by execution in that setting. The frozen verified experience is then transferred to five held-out solvers. Direct injection raises observed task resolution on four of them by 1.1 to 4.5 percentage points while lowering agent steps on all five. Yet when four existing memory systems must construct and retrieve experience from the same history, eleven of twelve solver and system pairings fail to exceed the matched memory-off baseline. VibeMemBench exposes the gap between the useful experience that repository history holds and the experience existing memory systems deliver for repository coding tasks.