作者Yuanda Xu, Zhengze Zhou, Kayhan Behdin, Jelena Markovic-Voronov, Hejian Sang, Xiaomin Li, Wenhui Zhu, Xinchen Du, Aida Rahmattalabi, Ran He, Sen Na, Zhipeng Wang, Alborz Geramifard
Group Relative Policy Optimization (GRPO) is effective when the current policy already samples useful reasoning trajectories, but it stalls on hard prompts whose correct solution modes lie outside the student's on-policy support. We propose TREK (Teacher-Routed Exploration via Forward KL), a simple staged procedure that uses distillation not for imitation but for exploration support expansion. A key advantage of TREK is its generality: because it only consumes verified output trajectories, it can use an external black-box teacher, a white-box teacher, or the same model given additional inference-time context, and it can efficiently identify which hard-prompt samples are most worth consolidating even when teacher internals are unavailable. TREK first identifies prompts where the unaided student has very low pass rate, queries a proposal source to produce verified candidate solutions, keeps the top-$r$ proposals ranked by current student likelihood, applies a short forward-KL phase to pull those verified modes into the student's support, and then returns to standard on-policy GRPO refinement. On mathematical reasoning, TREK with DeepSeek-V4 proposals improves Qwen3 models across all tested scales on AIME 2024 and AIME 2025; for Qwen3-8B, it improves AIME 2025 from 36.9 to 40.3 and AIME 2024 from 47.9 to 51.1 (avg@16), while the self-context variant reaches 38.5 and 49.6 without an external teacher. On agentic tasks, TREK raises ALFWorld success rate from 75.8 to 82.8 and ScienceWorld success rate from 12.5 to 26.7; notably, on the hardest task types, TREK achieves high success rates early in training while unaided GRPO requires substantially more optimization steps to reach comparable levels.
展开完整摘要收起摘要↓
Group Relative Policy Optimization (GRPO) is effective when the current policy already samples useful reasoning trajectories, but it stalls on hard prompts whose correct solution modes lie outside the student's on-policy support. We propose TREK (Teacher-Routed Exploration via Forward KL), a simple staged procedure that uses distillation not for imitation but for exploration support expansion. A key advantage of TREK is its generality: because it only consumes verified output trajectories, it can use an external black-box teacher, a white-box teacher, or the same model given additional inference-time context, and it can efficiently identify which hard-prompt samples are most worth consolidating even when teacher internals are unavailable. TREK first identifies prompts where the unaided student has very low pass rate, queries a proposal source to produce verified candidate solutions, keeps the top-$r$ proposals ranked by current student likelihood, applies a short forward-KL phase to pull those verified modes into the student's support, and then returns to standard on-policy GRPO refinement. On mathematical reasoning, TREK with DeepSeek-V4 proposals improves Qwen3 models across all tested scales on AIME 2024 and AIME 2025; for Qwen3-8B, it improves AIME 2025 from 36.9 to 40.3 and AIME 2024 from 47.9 to 51.1 (avg@16), while the self-context variant reaches 38.5 and 49.6 without an external teacher. On agentic tasks, TREK raises ALFWorld success rate from 75.8 to 82.8 and ScienceWorld success rate from 12.5 to 26.7; notably, on the hardest task types, TREK achieves high success rates early in training while unaided GRPO requires substantially more optimization steps to reach comparable levels.
Reinforcement learning with verifiable rewards (RLVR) is a powerful recipe for improving language-model reasoning, but it is expensive to repeat on every new strong model because the target model must generate many rollouts during training. As models scale, post-training itself becomes a bottleneck. We study a weak-to-strong alternative: run RL on a smaller model where rollouts are cheaper, then reuse what that RL run learned to improve a stronger target model. Directly distilling the post-RL weak teacher is not enough, because the teacher's final policy mixes useful RL gains with the limitations of the smaller model. We propose Direct On-Policy Distillation (Direct-OPD), which transfers the teacher's RL-induced policy shift instead. Direct-OPD compares the post-RL teacher with its own pre-RL reference and treats their log-ratio as a dense implicit reward for the student. In plain terms, the checkpoint pair tells us which actions RL made the weak model more or less likely to take, and Direct-OPD applies that signal on the stronger student's own on-policy states. This directly reuses the weak model's RL supervision signal without running sparse-reward RL on the target model. Empirically, Direct-OPD consistently leverages weaker teachers to improve stronger target models; notably, it boosts Qwen3-1.7B from 48.3% to 58.3% on AIME 2024 in just 4 hours on 8 A100 GPUs. It outperforms step-matched direct RL and enables the sequential composition of multiple policy shifts. Our results show that RL outcomes can be reused across model scales as implicit reward signals, not merely as final models to imitate.
展开完整摘要收起摘要↓
Reinforcement learning with verifiable rewards (RLVR) is a powerful recipe for improving language-model reasoning, but it is expensive to repeat on every new strong model because the target model must generate many rollouts during training. As models scale, post-training itself becomes a bottleneck. We study a weak-to-strong alternative: run RL on a smaller model where rollouts are cheaper, then reuse what that RL run learned to improve a stronger target model. Directly distilling the post-RL weak teacher is not enough, because the teacher's final policy mixes useful RL gains with the limitations of the smaller model. We propose Direct On-Policy Distillation (Direct-OPD), which transfers the teacher's RL-induced policy shift instead. Direct-OPD compares the post-RL teacher with its own pre-RL reference and treats their log-ratio as a dense implicit reward for the student. In plain terms, the checkpoint pair tells us which actions RL made the weak model more or less likely to take, and Direct-OPD applies that signal on the stronger student's own on-policy states. This directly reuses the weak model's RL supervision signal without running sparse-reward RL on the target model. Empirically, Direct-OPD consistently leverages weaker teachers to improve stronger target models; notably, it boosts Qwen3-1.7B from 48.3% to 58.3% on AIME 2024 in just 4 hours on 8 A100 GPUs. It outperforms step-matched direct RL and enables the sequential composition of multiple policy shifts. Our results show that RL outcomes can be reused across model scales as implicit reward signals, not merely as final models to imitate.
作者Junze Ye, Jiayi Cheng, Miao Lu, Michal Mankowski, Jose Blanchet, Mohsen Bayati
For LLM agents, supervised fine-tuning is not only about teacher labels' quality, but also about which interaction contexts those labels condition on. Pure behavioral cloning uses full teacher demonstrations, creating a mismatch between teacher-induced contexts seen in training and student-induced contexts encountered at test time. Recent work addresses this mismatch by querying a teacher at contexts reached by the student, often with increasingly elaborate filtering of the teacher's continuations. We instead frame on-policy data construction as a budget-allocation problem: under matched supervision resources, should teacher output be spent on more start-to-finish demos, longer continuations, outcome filtering, or broader coverage of learner-induced contexts? We formalize this design space through the rollout policy, switch-time distribution, continuation horizon, filtering rules, and two complementary costs: teacher inference generated before filtering and teacher supervision retained for SFT. Across HotpotQA, ALFWorld, and Terminal-Bench-Dev, bounded unfiltered teacher continuations at learner-induced contexts improve over pure behavioral cloning at matched budgets. On HotpotQA and ALFWorld, where we run the full comparison, few-step continuations match or exceed success-filtered and critical-context-filtered alternatives. Our findings suggest that a few teacher steps, placed at learner-induced contexts, can be a more cost-efficient supervision allocation than longer or more heavily curated teacher completions.
展开完整摘要收起摘要↓
For LLM agents, supervised fine-tuning is not only about teacher labels' quality, but also about which interaction contexts those labels condition on. Pure behavioral cloning uses full teacher demonstrations, creating a mismatch between teacher-induced contexts seen in training and student-induced contexts encountered at test time. Recent work addresses this mismatch by querying a teacher at contexts reached by the student, often with increasingly elaborate filtering of the teacher's continuations. We instead frame on-policy data construction as a budget-allocation problem: under matched supervision resources, should teacher output be spent on more start-to-finish demos, longer continuations, outcome filtering, or broader coverage of learner-induced contexts? We formalize this design space through the rollout policy, switch-time distribution, continuation horizon, filtering rules, and two complementary costs: teacher inference generated before filtering and teacher supervision retained for SFT. Across HotpotQA, ALFWorld, and Terminal-Bench-Dev, bounded unfiltered teacher continuations at learner-induced contexts improve over pure behavioral cloning at matched budgets. On HotpotQA and ALFWorld, where we run the full comparison, few-step continuations match or exceed success-filtered and critical-context-filtered alternatives. Our findings suggest that a few teacher steps, placed at learner-induced contexts, can be a more cost-efficient supervision allocation than longer or more heavily curated teacher completions.
作者Yifu Xiong, Wenhao Yu, Jiaxuan Lin, Bojun Zou, Jiahao Li, Lu Zhang, Yanyong Zhang, Jianmin Ji
Vision-Language-Action (VLA) models have become a promising paradigm for generalist robot manipulation, where visual-language representations are used to condition continuous action generation. However, these representations are not explicitly optimized for action conditioning, leaving the action expert to bridge the gap between multimodal understanding and precise motor control. Recent action-reasoning methods introduce additional modules to generate explicit action plans or action-space reasoning signals, demonstrating the benefit of action-level guidance but often requiring separate action-generation frameworks. We propose CAC-VLA, a Context-Gated Action Conditioning framework that learns a lightweight latent-action interface directly within the VLM. Instead of generating executable trajectories, CAC-VLA trains the VLM to predict coarse-to-fine latent actions, which are structured representations encoded from future action segments, and adaptively leverages them to condition the action expert via a context gate. This enables VLM-native action conditioning while calibrating the influence of latent-action guidance on expert action generation. Experiments on LIBERO and LIBERO-Plus demonstrate the effectiveness of CAC-VLA, achieving 98.3% average success rate on LIBERO and 89.5% LIBERO-Plus, suggesting that context-gated latent-action conditioning is an effective interface for continuous expert control.
展开完整摘要收起摘要↓
Vision-Language-Action (VLA) models have become a promising paradigm for generalist robot manipulation, where visual-language representations are used to condition continuous action generation. However, these representations are not explicitly optimized for action conditioning, leaving the action expert to bridge the gap between multimodal understanding and precise motor control. Recent action-reasoning methods introduce additional modules to generate explicit action plans or action-space reasoning signals, demonstrating the benefit of action-level guidance but often requiring separate action-generation frameworks. We propose CAC-VLA, a Context-Gated Action Conditioning framework that learns a lightweight latent-action interface directly within the VLM. Instead of generating executable trajectories, CAC-VLA trains the VLM to predict coarse-to-fine latent actions, which are structured representations encoded from future action segments, and adaptively leverages them to condition the action expert via a context gate. This enables VLM-native action conditioning while calibrating the influence of latent-action guidance on expert action generation. Experiments on LIBERO and LIBERO-Plus demonstrate the effectiveness of CAC-VLA, achieving 98.3% average success rate on LIBERO and 89.5% LIBERO-Plus, suggesting that context-gated latent-action conditioning is an effective interface for continuous expert control.
作者Ruslan Rakhimov, George Bredis, Yuriy Maksyuta, Daniil Gavrilov
Joint-Embedding Predictive Architectures (JEPAs) underpin a growing family of latent world models for control from raw pixels, but every existing JEPA world model commits at training time to a single inference paradigm: either trajectory optimisation in a learned dynamics model, or direct behaviour cloning. A single checkpoint that serves both would defer this choice to inference, when deployment constraints (rollout cost, observation accessibility) determine which path wins. We present Qantara, an end-to-end JEPA whose joint training objective pairs a Brownian-bridge interpolant between consecutive clean latents on the state axis with noise-to-data flow matching on the action axis. The same checkpoint serves three inference paradigms without retraining: latent planning, behaviour-cloning action sampling, and inverse dynamics, which we query through a video-inverse composition that first predicts the next latent without action conditioning, then extracts the action. Training concentrates mass on the edges of the (action-time, state-time) noise square, where inference queries the predictor: replacing it with uniform interior sampling drops Push-T planning from 90.1 to 53.3 SR at matched compute. On the LeWM control suite, Qantara reaches a 91.2 SR three-train-seed average and sets new SOTA on OGBench-Cube (+7.7 SR over DINO-WM, +19.7 over LeWM). From the same weights, the behaviour-cloning and video-inverse paths reach 82-83 SR on Push-T and 71-73 SR on Cube. These results move JEPA world models from single-paradigm planners to multi-paradigm controllers.
展开完整摘要收起摘要↓
Joint-Embedding Predictive Architectures (JEPAs) underpin a growing family of latent world models for control from raw pixels, but every existing JEPA world model commits at training time to a single inference paradigm: either trajectory optimisation in a learned dynamics model, or direct behaviour cloning. A single checkpoint that serves both would defer this choice to inference, when deployment constraints (rollout cost, observation accessibility) determine which path wins. We present Qantara, an end-to-end JEPA whose joint training objective pairs a Brownian-bridge interpolant between consecutive clean latents on the state axis with noise-to-data flow matching on the action axis. The same checkpoint serves three inference paradigms without retraining: latent planning, behaviour-cloning action sampling, and inverse dynamics, which we query through a video-inverse composition that first predicts the next latent without action conditioning, then extracts the action. Training concentrates mass on the edges of the (action-time, state-time) noise square, where inference queries the predictor: replacing it with uniform interior sampling drops Push-T planning from 90.1 to 53.3 SR at matched compute. On the LeWM control suite, Qantara reaches a 91.2 SR three-train-seed average and sets new SOTA on OGBench-Cube (+7.7 SR over DINO-WM, +19.7 over LeWM). From the same weights, the behaviour-cloning and video-inverse paths reach 82-83 SR on Push-T and 71-73 SR on Cube. These results move JEPA world models from single-paradigm planners to multi-paradigm controllers.
作者Paul Engstler, Iro Laina, Christian Rupprecht, Andrea Vedaldi
We present SynCity 3000, a framework for generating 3D scenes that are globally coherent while enabling fine-grained layout control. Building on the ability of current image-to-3D generators to produce complex 3D assets from a single image, we extend this capability to the scale of entire scenes by adapting the generator to be applicable as a convolutional operator. We achieve this by fine-tuning the model on scene-like data generated by a new synthetic data engine, which we propose to address the scarcity of 3D scene data for training. The convolutional generator is then applied to a dimetric image of the entire scene, generated from the user prompt, resulting in 3D scenes of arbitrary size and complexity. Across diverse prompts and layouts, SynCity 3000 produces large, coherent, and detailed scenes, addressing the shortcomings of prior approaches to 3D scene generation.
展开完整摘要收起摘要↓
We present SynCity 3000, a framework for generating 3D scenes that are globally coherent while enabling fine-grained layout control. Building on the ability of current image-to-3D generators to produce complex 3D assets from a single image, we extend this capability to the scale of entire scenes by adapting the generator to be applicable as a convolutional operator. We achieve this by fine-tuning the model on scene-like data generated by a new synthetic data engine, which we propose to address the scarcity of 3D scene data for training. The convolutional generator is then applied to a dimetric image of the entire scene, generated from the user prompt, resulting in 3D scenes of arbitrary size and complexity. Across diverse prompts and layouts, SynCity 3000 produces large, coherent, and detailed scenes, addressing the shortcomings of prior approaches to 3D scene generation.
作者Zhifeng Kong, Sang-gil Lee, Jaehyeon Kim, Boxin Wang, Zihan Liu, Sungwon Kim, Yang Chen, Arushi Goel, Rajarshi Roy, Wenliang Dai, Zhuolin Yang, Yangyi Chen, Dongfu Jiang, Sreyan Ghosh, Tuomas Rintamaki, Andrew Tao, Jonathan Raiman, Mohammad Shoeybi, Bryan Catanzaro, Wei Ping
Audio intelligence involves understanding, reasoning about, and generating both audio and speech. In this work, we introduce Nemotron-Labs-Audex-30B-A3B (Audex), a unified audio-text LLM built on Nemotron-Cascade-2-30B-A3B, a strong text-only MoE LLM. Audex adopts a simple unified design with a single Transformer decoder: audio inputs are encoded and projected into the text embedding space, while text tokens and quantized audio output tokens are treated uniformly during generation. This architecture enables strong audio-text fusion, seamless multimodal generation, and compatibility with standard LLM training and inference infrastructure. For training, we meticulously curate audio-text datasets comprising 157.4B audio tokens and 320.5B text tokens. We apply multi-stage supervised training on these datasets, followed by text-only Cascade RL and multi-domain on-policy distillation. Audex delivers state-of-the-art audio understanding, speech recognition and translation, text-to-speech, audio generation, and speech-to-speech generation, while preserving very compelling reasoning, alignment, knowledge, long-context, and agentic capabilities of its text-only LLM backbone with marginal or no regression. We release the model checkpoints to facilitate open research.
展开完整摘要收起摘要↓
Audio intelligence involves understanding, reasoning about, and generating both audio and speech. In this work, we introduce Nemotron-Labs-Audex-30B-A3B (Audex), a unified audio-text LLM built on Nemotron-Cascade-2-30B-A3B, a strong text-only MoE LLM. Audex adopts a simple unified design with a single Transformer decoder: audio inputs are encoded and projected into the text embedding space, while text tokens and quantized audio output tokens are treated uniformly during generation. This architecture enables strong audio-text fusion, seamless multimodal generation, and compatibility with standard LLM training and inference infrastructure. For training, we meticulously curate audio-text datasets comprising 157.4B audio tokens and 320.5B text tokens. We apply multi-stage supervised training on these datasets, followed by text-only Cascade RL and multi-domain on-policy distillation. Audex delivers state-of-the-art audio understanding, speech recognition and translation, text-to-speech, audio generation, and speech-to-speech generation, while preserving very compelling reasoning, alignment, knowledge, long-context, and agentic capabilities of its text-only LLM backbone with marginal or no regression. We release the model checkpoints to facilitate open research.
作者Ziqi Cai, Shuchen Weng, Kaiqi Liu, Zifeng Wang, Zhiquan Zhang, Minggui Teng, Han Jiang, Boxin Shi
Recovering dynamic environment maps from a single in-the-wild video is crucial for photorealistic rendering, yet remains a challenge. Recent video generation models can produce photorealistic scenes with complex lighting, possessing an inherent understanding of lighting. In this paper, we introduce V-LITE (Video generation models are inherent lighting estimators), a framework that unlocks this internal knowledge by reframing lighting estimation as a guided video inpainting task. Inspired by VFX industry practices, we insert a synthetic chrome ball into the scene to compel the model to generate physically plausible reflections from the surrounding spatio-temporal context. To bridge the gap from LDR-native models to the HDR domain, we design an HDR-aware VAE and employ an efficient LoRA-based fine-tuning strategy. We then construct a mixed dataset comprising high-fidelity HDR images to provide realistic HDR priors, and in-the-wild HDR videos to provide dynamic spatio-temporal context. Extensive experiments demonstrate that V-LITE produces temporally coherent HDR environment maps, revealing that modern video diffusion models are not merely synthesizers but also powerful, inherently capable estimators of physical scene lighting.
展开完整摘要收起摘要↓
Recovering dynamic environment maps from a single in-the-wild video is crucial for photorealistic rendering, yet remains a challenge. Recent video generation models can produce photorealistic scenes with complex lighting, possessing an inherent understanding of lighting. In this paper, we introduce V-LITE (Video generation models are inherent lighting estimators), a framework that unlocks this internal knowledge by reframing lighting estimation as a guided video inpainting task. Inspired by VFX industry practices, we insert a synthetic chrome ball into the scene to compel the model to generate physically plausible reflections from the surrounding spatio-temporal context. To bridge the gap from LDR-native models to the HDR domain, we design an HDR-aware VAE and employ an efficient LoRA-based fine-tuning strategy. We then construct a mixed dataset comprising high-fidelity HDR images to provide realistic HDR priors, and in-the-wild HDR videos to provide dynamic spatio-temporal context. Extensive experiments demonstrate that V-LITE produces temporally coherent HDR environment maps, revealing that modern video diffusion models are not merely synthesizers but also powerful, inherently capable estimators of physical scene lighting.
作者Marian Lupascu, Sebastian Ripa, Mihai Trascau, Mariana-Iuliana Georgescu, Ionut Mironica
Personalizing text-to-image diffusion models to render several specific subjects in a coherent image remains challenging: the model must preserve each subject's identity while keeping the scene spatially and visually coherent. Methods that fuse independently trained concept adapters in a shared weight space (via federated averaging, gradient fusion, or orthogonality constraints) suffer from identity confusion and style bleeding and require joint retraining. In this work, we show that composing concepts as separate image layers, instead of merging their adapters in a shared weight space, avoids parameter-level interference. We introduce LILAC, a framework that composes independently trained low-rank adapters at inference time: each subject is conditioned on the frozen composite of previously placed subjects, with exactly one adapter active at a time, therefore identities never interfere at the parameter level. LILAC composes the adapters without any joint training, scales linearly with the number of concepts, and is backbone-agnostic. Under the Orthogonal Adaptation protocol, LILAC applied on Qwen-Image-Edit reaches an ArcFace detection rate of 0.861, while Orthogonal Adaptation reports 0.745 in its original setting. Adaptation reports 0.745 in its original setting. Code is available at https://github.com/marianlupascu/LILAC.
展开完整摘要收起摘要↓
Personalizing text-to-image diffusion models to render several specific subjects in a coherent image remains challenging: the model must preserve each subject's identity while keeping the scene spatially and visually coherent. Methods that fuse independently trained concept adapters in a shared weight space (via federated averaging, gradient fusion, or orthogonality constraints) suffer from identity confusion and style bleeding and require joint retraining. In this work, we show that composing concepts as separate image layers, instead of merging their adapters in a shared weight space, avoids parameter-level interference. We introduce LILAC, a framework that composes independently trained low-rank adapters at inference time: each subject is conditioned on the frozen composite of previously placed subjects, with exactly one adapter active at a time, therefore identities never interfere at the parameter level. LILAC composes the adapters without any joint training, scales linearly with the number of concepts, and is backbone-agnostic. Under the Orthogonal Adaptation protocol, LILAC applied on Qwen-Image-Edit reaches an ArcFace detection rate of 0.861, while Orthogonal Adaptation reports 0.745 in its original setting. Adaptation reports 0.745 in its original setting. Code is available at https://github.com/marianlupascu/LILAC.
Recent advances in video diffusion models have enabled either long single-view generation through temporal autoregression, or short multi-view synthesis through bidirectional attention. However, generating long, multi-view consistent videos of dynamic scenes remains unsolved. In this work, we present MV-Forcing, a framework that composes temporal and view-wise autoregression within a single diffusion model by introducing a 4D geometric bridge between sequentially generated views. Our key insight is that an autoregressive 3D reconstruction model naturally interfaces between autoregressively generated views. Given a completed source view, we reconstruct its 3D structure and render a geometric prior of the next target viewpoint, which the diffusion model refines into a high-quality video. To extend generation beyond the teacher's fixed temporal window, we introduce a joint denoising regime where both view slots are initialized from noise during training, enabling temporally unbounded generation. We distill the model via Distribution Matching Distillation with Spatio-Temporal Self-Forcing, closing the train-inference exposure bias gap for both temporal and view-sequential autoregression. Extensive experiments on both synthetic and real-world data demonstrate that MV-Forcing produces geometrically consistent multi-view videos of dynamic scenes at arbitrary lengths and viewpoint counts using a single few-step student model.
展开完整摘要收起摘要↓
Recent advances in video diffusion models have enabled either long single-view generation through temporal autoregression, or short multi-view synthesis through bidirectional attention. However, generating long, multi-view consistent videos of dynamic scenes remains unsolved. In this work, we present MV-Forcing, a framework that composes temporal and view-wise autoregression within a single diffusion model by introducing a 4D geometric bridge between sequentially generated views. Our key insight is that an autoregressive 3D reconstruction model naturally interfaces between autoregressively generated views. Given a completed source view, we reconstruct its 3D structure and render a geometric prior of the next target viewpoint, which the diffusion model refines into a high-quality video. To extend generation beyond the teacher's fixed temporal window, we introduce a joint denoising regime where both view slots are initialized from noise during training, enabling temporally unbounded generation. We distill the model via Distribution Matching Distillation with Spatio-Temporal Self-Forcing, closing the train-inference exposure bias gap for both temporal and view-sequential autoregression. Extensive experiments on both synthetic and real-world data demonstrate that MV-Forcing produces geometrically consistent multi-view videos of dynamic scenes at arbitrary lengths and viewpoint counts using a single few-step student model.
作者Mengmeng Liu, Diankun Zhang, Jiuming Liu, Jianfeng Cui, Hongwei Xie, Guang Chen, Hangjun Ye, Francesco Nex, Hao Cheng, Michael Ying Yang
World Action Models (WAMs) have shown strong potential for improving action generalization in autonomous driving by using future video prediction as dense supervision for scene dynamics and temporal causality. However, it remains unclear which architecture better transfers video-modeling benefits to trajectory generation. Existing cascaded or dual-DiT designs separate video imagination from action prediction, weakening the transfer of video-learned world dynamics to the trajectory branch: the action model may still overfit dataset-specific driving priors, while the video model only indirectly regularizes planning. We propose UNIVERSE, a unified video-action model built upon a single mask-modulated Diffusion Transformer. By co-training future video latents and ego-trajectory tokens within shared generative parameters, UNIVERSE allows dense video supervision to directly shape trajectory denoising, leading to stronger cross-domain action generalization. To ensure causal validity and efficient deployment, we introduce a Modality-Decoupling Visibility Mask, which shares historical context across modalities while blocking mutual attention between future video and trajectory tokens. This prevents future-target leakage and enables trajectory-only inference by removing future-video denoising at test time, achieving a $4.3\times$ speedup over joint video-action rollout while maintaining comparable planning accuracy. The same model also supports video-only and joint video-action rollouts. Experiments show that UNIVERSE achieves 91.0 PDMS on NAVSIM (vs. 89.6 for the Two-DiT variant), and demonstrates strong zero-shot transfer to nuScenes and Bench2Drive without fine-tuning, while ablations confirm the importance of single-DiT unification, video co-training, and mask-based modality decoupling.
展开完整摘要收起摘要↓
World Action Models (WAMs) have shown strong potential for improving action generalization in autonomous driving by using future video prediction as dense supervision for scene dynamics and temporal causality. However, it remains unclear which architecture better transfers video-modeling benefits to trajectory generation. Existing cascaded or dual-DiT designs separate video imagination from action prediction, weakening the transfer of video-learned world dynamics to the trajectory branch: the action model may still overfit dataset-specific driving priors, while the video model only indirectly regularizes planning. We propose UNIVERSE, a unified video-action model built upon a single mask-modulated Diffusion Transformer. By co-training future video latents and ego-trajectory tokens within shared generative parameters, UNIVERSE allows dense video supervision to directly shape trajectory denoising, leading to stronger cross-domain action generalization. To ensure causal validity and efficient deployment, we introduce a Modality-Decoupling Visibility Mask, which shares historical context across modalities while blocking mutual attention between future video and trajectory tokens. This prevents future-target leakage and enables trajectory-only inference by removing future-video denoising at test time, achieving a $4.3\times$ speedup over joint video-action rollout while maintaining comparable planning accuracy. The same model also supports video-only and joint video-action rollouts. Experiments show that UNIVERSE achieves 91.0 PDMS on NAVSIM (vs. 89.6 for the Two-DiT variant), and demonstrates strong zero-shot transfer to nuScenes and Bench2Drive without fine-tuning, while ablations confirm the importance of single-DiT unification, video co-training, and mask-based modality decoupling.
作者Zhi Song, Ximing Xing, Zhenchao Tang, hanbo Huang, Tianxu Lv, minghao Yang, Zhongzheng Niu, He Bing, Lusheng Wang, Jianhua Yao
JEPA world models predict the next latent state with a single deterministic predictor trained by latent regression. We show that this fails structurally when the environment is stochastic: at a branching transition, the regression-optimal predictor outputs the conditional mean of the successor embeddings, a point between the true next states that corresponds to no state at all. We prove this collapse for deterministic and gated mixture-of-experts predictors, and prove that MoP-JEPA's hard-assigned predictors converge instead to a quantizer of the transition distribution: one head per successor mode, enumerable in a single forward pass, which is the interface a planner consumes. On official OGBench offline data with leak-free evaluation, planning over single-predictor rollouts performs poorly ($0.02$--$0.09$ success) while planning over our predicted modes reaches up to $0.85$, ahead of deterministic, gated-MoE, and variational predictors on every task. Because multi-prediction evaluation invites coverage freeloading, a verification protocol is part of the method: an input-agnostic codebook control, a shuffled-context test, router-gated readouts, transition-precision guards, and a verified-route criterion in which the model proposes its transition graph blind and ground truth is used only to check the result. Under this criterion our method outperforms the strongest soft alternative on all three mazes ($2$--$5\times$), and the protocol identifies the remaining gap in that baseline's raw scores as routes through predicted transitions that do not exist. The same model executes in the real environment, placing second of seven against the published OGBench baselines on the hardest maze. Multimodal dynamics decide whether a JEPA world model can plan at all; a mixture of predictors with hard assignment is a minimal and verifiable fix.
展开完整摘要收起摘要↓
JEPA world models predict the next latent state with a single deterministic predictor trained by latent regression. We show that this fails structurally when the environment is stochastic: at a branching transition, the regression-optimal predictor outputs the conditional mean of the successor embeddings, a point between the true next states that corresponds to no state at all. We prove this collapse for deterministic and gated mixture-of-experts predictors, and prove that MoP-JEPA's hard-assigned predictors converge instead to a quantizer of the transition distribution: one head per successor mode, enumerable in a single forward pass, which is the interface a planner consumes. On official OGBench offline data with leak-free evaluation, planning over single-predictor rollouts performs poorly ($0.02$--$0.09$ success) while planning over our predicted modes reaches up to $0.85$, ahead of deterministic, gated-MoE, and variational predictors on every task. Because multi-prediction evaluation invites coverage freeloading, a verification protocol is part of the method: an input-agnostic codebook control, a shuffled-context test, router-gated readouts, transition-precision guards, and a verified-route criterion in which the model proposes its transition graph blind and ground truth is used only to check the result. Under this criterion our method outperforms the strongest soft alternative on all three mazes ($2$--$5\times$), and the protocol identifies the remaining gap in that baseline's raw scores as routes through predicted transitions that do not exist. The same model executes in the real environment, placing second of seven against the published OGBench baselines on the hardest maze. Multimodal dynamics decide whether a JEPA world model can plan at all; a mixture of predictors with hard assignment is a minimal and verifiable fix.
作者Anthony Hu, Václav Volhejn, Adrien Ramanana Rahary, Chris Mulder, Aditya Makkar, Alyx Liao, Amélie Royer, Manu Orsini, Adam Jelley, Eloi Alonso, Florian Laurent, Fredrik Norén, James Swingos, Jan Hünermann, Kent Rollins, Lucas Hosseini, Matthieu Le Cauchois, Maxim Peter, Pim de Witte, Tim Brown, Vincent Micheli, Moritz Böhle, Gabriel de Marmiesse, Viktoriia Sharmanska, Lucia Specia, Michael Black, Patrick Pérez
We introduce the first multiplayer world model for highly dynamic environments governed by complex physical interactions. Whereas single-player world models treat the other agents as part of the environment, ours conditions on the action streams of multiple agents, learning to attribute changes in the scene to the correct player and to stay coherent under arbitrary combinations of their actions. We study this problem in the game of Rocket League, where players compete and cooperate under fast, tightly coupled dynamics. Trained on 10,000 hours of gameplay collected with publicly available bots, our 5-billion-parameter latent diffusion model generates four-player matches in real time, producing 20 frames per second on a single Nvidia B200 GPU. Although trained only on short clips, its rollouts stay stable far beyond the training horizon: distributional quality holds steady out to five minutes, the longest horizon we measure, and in practice we observe rollouts continuing for hours with no sign of collapse. We systematically investigate the central design choices: the video codec, the generative objective, and the multiplayer conditioning scheme. In addition, we characterize how behavior changes with model and data scale, including the capabilities that emerge and the failure modes that persist. We further develop targeted evaluations that probe the model's physical understanding rather than visual appearance alone. To support continued research on multiplayer world models, we release our dataset, our full training and inference codebase, and a live demo.
展开完整摘要收起摘要↓
We introduce the first multiplayer world model for highly dynamic environments governed by complex physical interactions. Whereas single-player world models treat the other agents as part of the environment, ours conditions on the action streams of multiple agents, learning to attribute changes in the scene to the correct player and to stay coherent under arbitrary combinations of their actions. We study this problem in the game of Rocket League, where players compete and cooperate under fast, tightly coupled dynamics. Trained on 10,000 hours of gameplay collected with publicly available bots, our 5-billion-parameter latent diffusion model generates four-player matches in real time, producing 20 frames per second on a single Nvidia B200 GPU. Although trained only on short clips, its rollouts stay stable far beyond the training horizon: distributional quality holds steady out to five minutes, the longest horizon we measure, and in practice we observe rollouts continuing for hours with no sign of collapse. We systematically investigate the central design choices: the video codec, the generative objective, and the multiplayer conditioning scheme. In addition, we characterize how behavior changes with model and data scale, including the capabilities that emerge and the failure modes that persist. We further develop targeted evaluations that probe the model's physical understanding rather than visual appearance alone. To support continued research on multiplayer world models, we release our dataset, our full training and inference codebase, and a live demo.
作者Jian Zhu, Jianjun Zhang, Taiyi Su, Tianbin Liu, Zhangyuan Wang, Kai Xie, Zitai Huang, Chong Ma, Youzhang He, Tianjian Wang, Hanyang Wang, Weihao Ding, Yi Xu
World Action Models (WAMs) provide a promising alternative to Vision-Language-Action (VLA) policies by using video-based world modeling as dense supervision for robot action learning. Existing WAMs excel at physically grounded execution, but typically lack the explicit language-level planning interface in VLM-based VLAs for decomposing coarse instructions. Such decomposition becomes important when household tasks involve complex multi-step goals, where coarse user commands need to be converted into sequences of fine-grained executable subtasks. Meanwhile, the field still lacks a fair real-robot comparison between VLA and WAM execution capabilities, since existing systems often differ in data, robot embodiments, and task protocols. To address both the decomposition gap and the need for a controlled WAM-VLA comparison, we introduce DSWAM, a Dual-System World Action Foundation Model for fine-grained robot manipulation. DSWAM keeps a System 1 WAM executor as the default control path and optionally activates a System 2 vision-language subtask planner only when task decomposition is useful. The planner predicts executable subtasks from short-term visual history and a global task prompt, while the WAM executor performs world-aware action generation for each instruction or subtask. The executor is trained with action prediction and video co-training, but inference directly predicts action chunks without explicit future video generation. To make this execution path practical on real robots, we further integrate TensorRT acceleration, asynchronous execution, and real-time chunking (RTC) so that policy queries do not block robot control. To provide a fair real-robot comparison with VLA policies, we build and evaluate DSWAM under the DeMaVLA real-world deformable manipulation setting with matched robot platform, pretraining data, post-training data, and evaluation criteria.
展开完整摘要收起摘要↓
World Action Models (WAMs) provide a promising alternative to Vision-Language-Action (VLA) policies by using video-based world modeling as dense supervision for robot action learning. Existing WAMs excel at physically grounded execution, but typically lack the explicit language-level planning interface in VLM-based VLAs for decomposing coarse instructions. Such decomposition becomes important when household tasks involve complex multi-step goals, where coarse user commands need to be converted into sequences of fine-grained executable subtasks. Meanwhile, the field still lacks a fair real-robot comparison between VLA and WAM execution capabilities, since existing systems often differ in data, robot embodiments, and task protocols. To address both the decomposition gap and the need for a controlled WAM-VLA comparison, we introduce DSWAM, a Dual-System World Action Foundation Model for fine-grained robot manipulation. DSWAM keeps a System 1 WAM executor as the default control path and optionally activates a System 2 vision-language subtask planner only when task decomposition is useful. The planner predicts executable subtasks from short-term visual history and a global task prompt, while the WAM executor performs world-aware action generation for each instruction or subtask. The executor is trained with action prediction and video co-training, but inference directly predicts action chunks without explicit future video generation. To make this execution path practical on real robots, we further integrate TensorRT acceleration, asynchronous execution, and real-time chunking (RTC) so that policy queries do not block robot control. To provide a fair real-robot comparison with VLA policies, we build and evaluate DSWAM under the DeMaVLA real-world deformable manipulation setting with matched robot platform, pretraining data, post-training data, and evaluation criteria.
Unified models for robot manipulation aim to equip one policy with both the semantic priors of pretrained VLMs and the physical dynamics learned through future prediction. In practice, existing designs tend to erode the semantics of the pretrained backbone, suffer interference among heterogeneous objectives, and learn future prediction from scratch in pixel space, leaving the dynamics priors of pretrained video generators unexploited. We present InternVLA-A1.5, which builds the policy on a native VLM backbone that keeps training on VQA and subtask prediction, and attaches a lightweight unified expert for continuous action generation. Future prediction is recast as a latent-querying problem, where a small set of learnable foresight tokens condenses the task-relevant future into a compact latent code under the supervision of a frozen pretrained video generation model, so the policy inherits world-model dynamics priors without ever learning pixel-level generation. The video branch is discarded at inference, keeping real-time control. Pretrained on 1.2M robot episodes and 3M multimodal samples, InternVLA-A1.5 achieves the best overall results on all six simulation benchmarks. In the real world, the preserved semantics deliver the strongest compositional generalization on held-out instruction bindings, and the two designs together sustain long-horizon execution.
展开完整摘要收起摘要↓
Unified models for robot manipulation aim to equip one policy with both the semantic priors of pretrained VLMs and the physical dynamics learned through future prediction. In practice, existing designs tend to erode the semantics of the pretrained backbone, suffer interference among heterogeneous objectives, and learn future prediction from scratch in pixel space, leaving the dynamics priors of pretrained video generators unexploited. We present InternVLA-A1.5, which builds the policy on a native VLM backbone that keeps training on VQA and subtask prediction, and attaches a lightweight unified expert for continuous action generation. Future prediction is recast as a latent-querying problem, where a small set of learnable foresight tokens condenses the task-relevant future into a compact latent code under the supervision of a frozen pretrained video generation model, so the policy inherits world-model dynamics priors without ever learning pixel-level generation. The video branch is discarded at inference, keeping real-time control. Pretrained on 1.2M robot episodes and 3M multimodal samples, InternVLA-A1.5 achieves the best overall results on all six simulation benchmarks. In the real world, the preserved semantics deliver the strongest compositional generalization on held-out instruction bindings, and the two designs together sustain long-horizon execution.
Recently, diffusion models have achieved considerable success in the text-guided video editing domain. However, existing works often struggle to balance the trade-off between temporal consistency and editability in video editing, with consistency and editability typically being inversely related. To address this, we propose a high-quality video editing framework enhanced for consistency and editability, named EquiEdit, which improves coordinatively the temporal consistency and editability of the edited videos while achieving a balance between the two. In terms of temporal consistency, the proposed temporal Mamba module with a tailored temporal-aware scanning scans fused video sequences following four designed directions, effectively enhancing the inter-frame consistency of edited videos. For editability, we design a noise injection strategy based on the spectral transformation to increase editing flexibility, where the Fourier transform is used to preserve the hidden structure in the initial latent noise used for editing, ensuring inter-frame consistency of the edited video and fidelity to the input video. Extensive qualitative and quantitative experiments demonstrate the effectiveness of our method in terms of temporal consistency and editability, as well as its great fidelity to the input video itself.
展开完整摘要收起摘要↓
Recently, diffusion models have achieved considerable success in the text-guided video editing domain. However, existing works often struggle to balance the trade-off between temporal consistency and editability in video editing, with consistency and editability typically being inversely related. To address this, we propose a high-quality video editing framework enhanced for consistency and editability, named EquiEdit, which improves coordinatively the temporal consistency and editability of the edited videos while achieving a balance between the two. In terms of temporal consistency, the proposed temporal Mamba module with a tailored temporal-aware scanning scans fused video sequences following four designed directions, effectively enhancing the inter-frame consistency of edited videos. For editability, we design a noise injection strategy based on the spectral transformation to increase editing flexibility, where the Fourier transform is used to preserve the hidden structure in the initial latent noise used for editing, ensuring inter-frame consistency of the edited video and fidelity to the input video. Extensive qualitative and quantitative experiments demonstrate the effectiveness of our method in terms of temporal consistency and editability, as well as its great fidelity to the input video itself.
作者Yan Zeng, Yusuke Hosoya, Huyen T. T. Tran, Takayuki Okatani
Text-guided diffusion inversion is central to image editing, where an image is mapped to an initial latent and then edited by replaying the denoising process under a modified prompt. In practice, however, inversion is often performed with a lower classifier-free guidance(CFG) scale than the one used for generation or editing. This mismatch is empirically useful but leaves a basic question unresolved: when a target image is generated by a high-CFG trajectory, when can that trajectory actually be inverted? We study this question in a controlled generation--inversion--reconstruction setting, where the true initial latent and denoising trajectory are known. Using prompts taken from an existing diffusion-editing benchmark, we generate images under high CFG and reconstruct them with fixed-point inversion using the same prompt and guidance setting. The results reveal three types of prompt-level reconstruction behavior: easy prompts that reconstruct for most initial latents, hard prompts that fail for most initial latents, and intermediate prompts whose success depends on the prompt--latent pairing. To analyze the generation side, we define prompt pressure, a step-wise measure of how strongly CFG moves the denoising update away from the unconditional trajectory. Total pressure correlates with reconstruction quality and separates easy from hard prompts, but it does not explain the success or failure of intermediate prompt--latent pairs. Text-side analyses further show that the main visual subject and wording can change inversion difficulty. Finally, we evaluate a compact trajectory-consistency intervention that relaxes guidance only at locally unstable inverse steps. This diagnostic check improves reconstruction and Prompt-to-Prompt editing in our controlled setting, supporting the view that high-CFG inversion failure requires local, trajectory-aware analysis.
展开完整摘要收起摘要↓
Text-guided diffusion inversion is central to image editing, where an image is mapped to an initial latent and then edited by replaying the denoising process under a modified prompt. In practice, however, inversion is often performed with a lower classifier-free guidance(CFG) scale than the one used for generation or editing. This mismatch is empirically useful but leaves a basic question unresolved: when a target image is generated by a high-CFG trajectory, when can that trajectory actually be inverted? We study this question in a controlled generation--inversion--reconstruction setting, where the true initial latent and denoising trajectory are known. Using prompts taken from an existing diffusion-editing benchmark, we generate images under high CFG and reconstruct them with fixed-point inversion using the same prompt and guidance setting. The results reveal three types of prompt-level reconstruction behavior: easy prompts that reconstruct for most initial latents, hard prompts that fail for most initial latents, and intermediate prompts whose success depends on the prompt--latent pairing. To analyze the generation side, we define prompt pressure, a step-wise measure of how strongly CFG moves the denoising update away from the unconditional trajectory. Total pressure correlates with reconstruction quality and separates easy from hard prompts, but it does not explain the success or failure of intermediate prompt--latent pairs. Text-side analyses further show that the main visual subject and wording can change inversion difficulty. Finally, we evaluate a compact trajectory-consistency intervention that relaxes guidance only at locally unstable inverse steps. This diagnostic check improves reconstruction and Prompt-to-Prompt editing in our controlled setting, supporting the view that high-CFG inversion failure requires local, trajectory-aware analysis.
High-resolution velocity models are crucial for reservoir characterization and subsurface delineation. However, the band limited nature of our surface recorded data limits resolution. Utilizing well measurements to enhance the resolution of our subsurface models is an important objective. To this end, we present a diffusion-guided framework for structurally preconditioned velocity-model reconstruction from sparse well-log information. The proposed approach combines plane-wave PDE regularization, structurally preconditioned inversion, and measurement-guided diffusion posterior sampling within a unified formulation. Local structural slopes estimated through plane-wave destruction are used both to propagate well information along geological dip directions and to guide the diffusion sampling process through a joint velocity--slope generative prior. Numerical experiments on the Volve synthetic model and the Viking Graben field dataset demonstrate that the proposed framework improves structural continuity, lateral consistency, and geological realism compared with conventional structurally preconditioned inversion approaches while maintaining computationally practical inference through DDIM sampling.
展开完整摘要收起摘要↓
High-resolution velocity models are crucial for reservoir characterization and subsurface delineation. However, the band limited nature of our surface recorded data limits resolution. Utilizing well measurements to enhance the resolution of our subsurface models is an important objective. To this end, we present a diffusion-guided framework for structurally preconditioned velocity-model reconstruction from sparse well-log information. The proposed approach combines plane-wave PDE regularization, structurally preconditioned inversion, and measurement-guided diffusion posterior sampling within a unified formulation. Local structural slopes estimated through plane-wave destruction are used both to propagate well information along geological dip directions and to guide the diffusion sampling process through a joint velocity--slope generative prior. Numerical experiments on the Volve synthetic model and the Viking Graben field dataset demonstrate that the proposed framework improves structural continuity, lateral consistency, and geological realism compared with conventional structurally preconditioned inversion approaches while maintaining computationally practical inference through DDIM sampling.
作者Amanda Wang, Tudor Manole, Florentina Bunea, John Thickstun
As language modeling technology matures, there is an increasing research focus on the composition and curation of datasets used to train these models. For instance, practitioners commonly seek to augment high-quality datasets with additional text to enhance the performance of models trained on that data. However, informed decisions about data augmentation require more nuanced assessments about data quality. We build on work measuring the precision and recall of generative models to develop a pair of metrics that quantify (1) fidelity, capturing how closely candidate text resembles reference data, and (2) diversity, capturing how well it covers the modes of the reference dataset. Our metrics are based on optimal transport divergence functionals between discrete text summaries. In experiments on M2D2 text datasets, we show that these metrics are able to disentangle a lack of fidelity from a lack of diversity in deficient candidate text. In further experiments, our metrics detect diversity deficits in synthetic GSM8K-style math datasets, which correlate with degradations in downstream accuracy of language models finetuned on this synthetic data.
展开完整摘要收起摘要↓
As language modeling technology matures, there is an increasing research focus on the composition and curation of datasets used to train these models. For instance, practitioners commonly seek to augment high-quality datasets with additional text to enhance the performance of models trained on that data. However, informed decisions about data augmentation require more nuanced assessments about data quality. We build on work measuring the precision and recall of generative models to develop a pair of metrics that quantify (1) fidelity, capturing how closely candidate text resembles reference data, and (2) diversity, capturing how well it covers the modes of the reference dataset. Our metrics are based on optimal transport divergence functionals between discrete text summaries. In experiments on M2D2 text datasets, we show that these metrics are able to disentangle a lack of fidelity from a lack of diversity in deficient candidate text. In further experiments, our metrics detect diversity deficits in synthetic GSM8K-style math datasets, which correlate with degradations in downstream accuracy of language models finetuned on this synthetic data.
Persona-Trained Monte Carlo (PTMC) estimates distributions of market-outcome functionals by repeatedly simulating limit-order-book interaction among $K$ neural policy bots whose behavioral personas are drawn from a learned heterogeneity distribution $\mathcal{P}$. This paper develops the statistical theory that makes the word "reliable" precise for such estimators. We decompose estimator variance into a persona-draw component $σ_P^2$ and a within-run component $σ_w^2$, give unbiased ANOVA estimators of both, and derive the variance-optimal allocation of a fixed compute budget between outer persona draws and inner replications. A coupling-based stability bound quantifies how misestimation of $\mathcal{P}$ and error in the trained policy propagate into the estimand, yielding a three-term total-error budget whose terms are separately estimable; a uniform-in-horizon version holds under a Doeblin condition on the market chain. The main contribution is an identification theory for heterogeneous news reaction: under a fixed response nonlinearity, the aggregate impact curve $A(z)=\mathbb{E}_Q[g(ηz)]$ detects heterogeneous news sensitivity through a strict Jensen gap and identifies the distribution $Q$ locally via odd moments and Hausdorff determinacy, with sharp failure when the response family is unknown. We provide $\sqrt{n}$-consistent estimators and a boundary-corrected test of homogeneous news reaction. Two separation theorems delimit when PTMC is provably preferable to homogeneous-population simulators and reduced-form forecasters, formalizing an irreducible Jensen bias floor and the Lucas critique as a minimax limit on intervention extrapolation. All proofs are given in full; guarantees are classified as unconditional (Monte Carlo convergence), conditional worst-case (the error budget), or open (the large-$K$ mean-field limit).
展开完整摘要收起摘要↓
Persona-Trained Monte Carlo (PTMC) estimates distributions of market-outcome functionals by repeatedly simulating limit-order-book interaction among $K$ neural policy bots whose behavioral personas are drawn from a learned heterogeneity distribution $\mathcal{P}$. This paper develops the statistical theory that makes the word "reliable" precise for such estimators. We decompose estimator variance into a persona-draw component $σ_P^2$ and a within-run component $σ_w^2$, give unbiased ANOVA estimators of both, and derive the variance-optimal allocation of a fixed compute budget between outer persona draws and inner replications. A coupling-based stability bound quantifies how misestimation of $\mathcal{P}$ and error in the trained policy propagate into the estimand, yielding a three-term total-error budget whose terms are separately estimable; a uniform-in-horizon version holds under a Doeblin condition on the market chain. The main contribution is an identification theory for heterogeneous news reaction: under a fixed response nonlinearity, the aggregate impact curve $A(z)=\mathbb{E}_Q[g(ηz)]$ detects heterogeneous news sensitivity through a strict Jensen gap and identifies the distribution $Q$ locally via odd moments and Hausdorff determinacy, with sharp failure when the response family is unknown. We provide $\sqrt{n}$-consistent estimators and a boundary-corrected test of homogeneous news reaction. Two separation theorems delimit when PTMC is provably preferable to homogeneous-population simulators and reduced-form forecasters, formalizing an irreducible Jensen bias floor and the Lucas critique as a minimax limit on intervention extrapolation. All proofs are given in full; guarantees are classified as unconditional (Monte Carlo convergence), conditional worst-case (the error budget), or open (the large-$K$ mean-field limit).
Many real-world systems evolve continuously, yet most machine learning models interpret time series as discrete sequences. Continuous-time approaches instead treat time series as samples from an underlying input path, a formulation that naturally accommodates irregularly sampled or oversampled data. Among these, Neural Controlled Differential Equations (NCDEs) are a maximally expressive class of models that parametrise a vector field using a neural network and evolve their hidden state by solving a dynamical system driven by the input path. NCDEs typically use a non-linear vector field, so their expressive power and continuous-time flexibility come at the cost of a forward pass that is both computationally expensive and inherently sequential, limiting their scalability and practical applicability. This thesis advances the training and scalability of NCDEs through three complementary contributions. First, building on neural rough differential equations, Log-NCDEs apply the Log-ODE method to efficiently approximate an NCDE's solution during training, improving both computational speed and empirical performance. Second, Linear NCDEs replace the non-linear vector field with a linear one, enabling closed-form solutions and parallel-in-time computation without sacrificing theoretical expressivity. Third, Structured Linear NCDEs use structured linear vector fields to further enhance efficiency while maintaining theoretical expressiveness and empirical performance. Collectively, these methods reduce the time per training step for an NCDE by up to three orders of magnitude while achieving state-of-the-art performance across diverse time series benchmarks.
展开完整摘要收起摘要↓
Many real-world systems evolve continuously, yet most machine learning models interpret time series as discrete sequences. Continuous-time approaches instead treat time series as samples from an underlying input path, a formulation that naturally accommodates irregularly sampled or oversampled data. Among these, Neural Controlled Differential Equations (NCDEs) are a maximally expressive class of models that parametrise a vector field using a neural network and evolve their hidden state by solving a dynamical system driven by the input path. NCDEs typically use a non-linear vector field, so their expressive power and continuous-time flexibility come at the cost of a forward pass that is both computationally expensive and inherently sequential, limiting their scalability and practical applicability. This thesis advances the training and scalability of NCDEs through three complementary contributions. First, building on neural rough differential equations, Log-NCDEs apply the Log-ODE method to efficiently approximate an NCDE's solution during training, improving both computational speed and empirical performance. Second, Linear NCDEs replace the non-linear vector field with a linear one, enabling closed-form solutions and parallel-in-time computation without sacrificing theoretical expressivity. Third, Structured Linear NCDEs use structured linear vector fields to further enhance efficiency while maintaining theoretical expressiveness and empirical performance. Collectively, these methods reduce the time per training step for an NCDE by up to three orders of magnitude while achieving state-of-the-art performance across diverse time series benchmarks.
作者Rodrigo Casado Noguerales, Bernhard Schölkopf, Thomas Hofmann, Aran Raoufi
What does a discrete diffusion model learn: a denoiser, a score ratio, or a bridge plug-in predictor? At the level of jump rates, these are one object in different coordinates, and reading a neural network in the wrong coordinate changes the process being trained and sampled. Starting with a rigorous derivation of the continuous-time Markov chain (CTMC) ELBO for any noising process, boundary terms included, we prove the Oracle Distance theorem: the negative ELBO is exactly equal to the data entropy plus the path KL from the oracle reverse process to the learned one, not merely a bound. Its unique optimizer is therefore the conditional expectation of the true reverse jump rate given the current noisy state, and its irreducible cost is the rate at which the forward process $Z_t$ destroys information about the clean data $Z_0$, $-\tfrac{d}{dt}I(Z_0; Z_t)$, so every noising process shares the same best achievable negative ELBO: the data entropy. For sequences with token-factorizing noise, the oracle projection yields three exact coordinates for the optimizer: denoiser, cavity (bridge plug-in), and score, with closed-form conversions among them. This framework identifies which law each loss in the literature actually optimizes, recovering MDM, UDM, SEDD, and GIDD as special cases; explains why denoiser and cavity coincide for masked diffusion but not for uniform diffusion; proves that a denoiser parameterization makes the uniform ELBO diverge at initialization while the bridge plug-in stays finite; and calibrates ELBO implementations exactly at initialization. Every identity is verified numerically, without approximation, on an exactly solvable model.
展开完整摘要收起摘要↓
What does a discrete diffusion model learn: a denoiser, a score ratio, or a bridge plug-in predictor? At the level of jump rates, these are one object in different coordinates, and reading a neural network in the wrong coordinate changes the process being trained and sampled. Starting with a rigorous derivation of the continuous-time Markov chain (CTMC) ELBO for any noising process, boundary terms included, we prove the Oracle Distance theorem: the negative ELBO is exactly equal to the data entropy plus the path KL from the oracle reverse process to the learned one, not merely a bound. Its unique optimizer is therefore the conditional expectation of the true reverse jump rate given the current noisy state, and its irreducible cost is the rate at which the forward process $Z_t$ destroys information about the clean data $Z_0$, $-\tfrac{d}{dt}I(Z_0; Z_t)$, so every noising process shares the same best achievable negative ELBO: the data entropy. For sequences with token-factorizing noise, the oracle projection yields three exact coordinates for the optimizer: denoiser, cavity (bridge plug-in), and score, with closed-form conversions among them. This framework identifies which law each loss in the literature actually optimizes, recovering MDM, UDM, SEDD, and GIDD as special cases; explains why denoiser and cavity coincide for masked diffusion but not for uniform diffusion; proves that a denoiser parameterization makes the uniform ELBO diverge at initialization while the bridge plug-in stays finite; and calibrates ELBO implementations exactly at initialization. Every identity is verified numerically, without approximation, on an exactly solvable model.
Training-free concept erasure is an attractive mechanism for controlling text-to-image diffusion models, but precise erasure often comes at the cost of damaging semantically related non-target concepts. Existing value-space methods remove the component of each cross-attention value along the target concept direction, implicitly treating target identity and shared visual structure as the same signal. We argue that this is the source of much of the collateral damage in prior preservation. We introduce CARE, a closed-form concept erasure operator that replaces the raw target direction with a kept-subspace-aware direction computed from a small bank of retained concept anchors. The resulting edit is applied directly in cross-attention value space, requires no model fine-tuning, and adds only a negligible offline computation. A single shrinkage parameter controls the erase-preserve trade-off. We further show that the operator admits a minimum-disturbance interpretation and, in its projection form, leaves the kept subspace invariant. Experiments under the standard concept-erasure protocol show that our method preserves non-target concepts more faithfully while maintaining competitive erasure across instance, style, and celebrity concepts. Code: https://github.com/parthupman/care
展开完整摘要收起摘要↓
Training-free concept erasure is an attractive mechanism for controlling text-to-image diffusion models, but precise erasure often comes at the cost of damaging semantically related non-target concepts. Existing value-space methods remove the component of each cross-attention value along the target concept direction, implicitly treating target identity and shared visual structure as the same signal. We argue that this is the source of much of the collateral damage in prior preservation. We introduce CARE, a closed-form concept erasure operator that replaces the raw target direction with a kept-subspace-aware direction computed from a small bank of retained concept anchors. The resulting edit is applied directly in cross-attention value space, requires no model fine-tuning, and adds only a negligible offline computation. A single shrinkage parameter controls the erase-preserve trade-off. We further show that the operator admits a minimum-disturbance interpretation and, in its projection form, leaves the kept subspace invariant. Experiments under the standard concept-erasure protocol show that our method preserves non-target concepts more faithfully while maintaining competitive erasure across instance, style, and celebrity concepts. Code: https://github.com/parthupman/care
While controllable image generation has made significant strides by incorporating visual reference conditions, existing methods predominantly operate as open-loop systems. They inject control signals in a strictly feed-forward manner, failing to guarantee strict fidelity to the reference due to the absence of active feedback and error correction mechanisms. To address this fundamental limitation, we propose a novel test-time iterative optimization framework that reformulates reference-consistent generation as a closed-loop dynamic tracking problem. By treating the pre-trained generative model as a control plant, our framework employs a sensor-controller architecture driven by a modified Proportional-Integral-Derivative (PID) algorithm. This mechanism iteratively optimizes the latent control signals at test time based on the sensed discrepancy between the generated output and the reference target. Notably, this approach is entirely training-free, model-agnostic, and integrates seamlessly around existing diffusion pipelines. Extensive evaluations across ID-preserving, pose-controlled, and depth-controlled generation tasks validate the universality of our method. Empirical results demonstrate improvements over computation-matched open-loop baselines, achieving relative performance gains of up to 25.36% for facial similarity, alongside spatial error reductions of up to 27.71% for pose alignment and 28.50% for depth consistency. More broadly, this work offers a new conceptual perspective: it demonstrates that controllable generation can be effectively managed as a dynamic feedback system, bringing the rigorous principles of classical control theory into the optimization of generative models. Code is available at https://github.com/zzdrill/From-Open-Loop-to-Closed-Loop.
展开完整摘要收起摘要↓
While controllable image generation has made significant strides by incorporating visual reference conditions, existing methods predominantly operate as open-loop systems. They inject control signals in a strictly feed-forward manner, failing to guarantee strict fidelity to the reference due to the absence of active feedback and error correction mechanisms. To address this fundamental limitation, we propose a novel test-time iterative optimization framework that reformulates reference-consistent generation as a closed-loop dynamic tracking problem. By treating the pre-trained generative model as a control plant, our framework employs a sensor-controller architecture driven by a modified Proportional-Integral-Derivative (PID) algorithm. This mechanism iteratively optimizes the latent control signals at test time based on the sensed discrepancy between the generated output and the reference target. Notably, this approach is entirely training-free, model-agnostic, and integrates seamlessly around existing diffusion pipelines. Extensive evaluations across ID-preserving, pose-controlled, and depth-controlled generation tasks validate the universality of our method. Empirical results demonstrate improvements over computation-matched open-loop baselines, achieving relative performance gains of up to 25.36% for facial similarity, alongside spatial error reductions of up to 27.71% for pose alignment and 28.50% for depth consistency. More broadly, this work offers a new conceptual perspective: it demonstrates that controllable generation can be effectively managed as a dynamic feedback system, bringing the rigorous principles of classical control theory into the optimization of generative models. Code is available at https://github.com/zzdrill/From-Open-Loop-to-Closed-Loop.
We study the problem of k-means clustering on large datasets. The state-of-the-art for the problem is given by coresets-based approaches, which build small weighted summaries of the input and derive approximate solutions with rigorous quality guarantees from them. One of the most popular and advanced approaches to derive coresets for k-means is sensitivity sampling. However, sensitivity sampling requires to compute the importance of each input point with respect to the whole dataset over all possible choices of centers. Since the exact computation of such quantities is unfeasible, current approaches work by approximating the sensitivity values. Nevertheless, the runtime of such approaches is still impractical for large datasets. In this work, we propose to reduce the runtime of sensitivity-based approaches for k-means by leveraging predictions to approximate the importance of input points. We first formally prove that current theoretical results on coresets construction via sensitivity sampling hold for coarser approximations of sensitivities compared to the one required by existing approaches. This implies that even fairly noisy predictors can be leveraged for sensitivity-sampling approaches. We then propose a natural predictor, which applies to the common scenario where clustering is performed (over time) on a sequence of datasets from the same problem. We prove that when the datasets in the sequence come from the same (unknown) distribution, centers resulting in a low error on one dataset can be used as predictions for sensitivity sampling in subsequent datasets, with guarantees on their quality. We perform an extensive experimental evaluation showing that our approach significantly improves, in terms of clustering cost vs runtime, over uniform sampling and state-of-the-art sensitivity sampling approaches when applied to sequences of datasets.
展开完整摘要收起摘要↓
We study the problem of k-means clustering on large datasets. The state-of-the-art for the problem is given by coresets-based approaches, which build small weighted summaries of the input and derive approximate solutions with rigorous quality guarantees from them. One of the most popular and advanced approaches to derive coresets for k-means is sensitivity sampling. However, sensitivity sampling requires to compute the importance of each input point with respect to the whole dataset over all possible choices of centers. Since the exact computation of such quantities is unfeasible, current approaches work by approximating the sensitivity values. Nevertheless, the runtime of such approaches is still impractical for large datasets. In this work, we propose to reduce the runtime of sensitivity-based approaches for k-means by leveraging predictions to approximate the importance of input points. We first formally prove that current theoretical results on coresets construction via sensitivity sampling hold for coarser approximations of sensitivities compared to the one required by existing approaches. This implies that even fairly noisy predictors can be leveraged for sensitivity-sampling approaches. We then propose a natural predictor, which applies to the common scenario where clustering is performed (over time) on a sequence of datasets from the same problem. We prove that when the datasets in the sequence come from the same (unknown) distribution, centers resulting in a low error on one dataset can be used as predictions for sensitivity sampling in subsequent datasets, with guarantees on their quality. We perform an extensive experimental evaluation showing that our approach significantly improves, in terms of clustering cost vs runtime, over uniform sampling and state-of-the-art sensitivity sampling approaches when applied to sequences of datasets.
作者Weichen Qin, Yufan Xie, Peihao Wang, Chia-Jui Chou, Minghui Du, Peng Xu, Ziren Luo, Yi Yang, Jingyi Yu, Bo Liang, Jiakai Zhang
Simulation-Based Inference (SBI) is critical for scientific discovery, with generative models offering a promising path toward efficient inference. However, existing methods struggle with effective multimodal modeling. They often rely on brute-force fusion strategies that ignore the structural disparities between parameters and observations, thus limiting estimation fidelity. In this work, we introduce FUSE (Feynman-Kac steered mUlti-modal flow matching for efficient Simulation-based posterior Estimation). Unlike prior work, FUSE employs a dual-track architecture that preserves the distinct features of multimodal inputs while facilitating dynamic interaction. Additionally, we propose an FK-steered sampling strategy that leverages intermediate observation likelihoods to guide the generative trajectories, effectively improving the sample quality during inference. Our approach outperforms state-of-the-art baselines on standard SBI benchmarks, producing posteriors that closely match ground-truth MCMC. Furthermore, in a real-world exoplanet orbital estimation task, FUSE successfully resolves complex parameter degeneracies that challenge existing methods, highlighting its potential to accelerate complex scientific discoveries in astrophysics and beyond.
展开完整摘要收起摘要↓
Simulation-Based Inference (SBI) is critical for scientific discovery, with generative models offering a promising path toward efficient inference. However, existing methods struggle with effective multimodal modeling. They often rely on brute-force fusion strategies that ignore the structural disparities between parameters and observations, thus limiting estimation fidelity. In this work, we introduce FUSE (Feynman-Kac steered mUlti-modal flow matching for efficient Simulation-based posterior Estimation). Unlike prior work, FUSE employs a dual-track architecture that preserves the distinct features of multimodal inputs while facilitating dynamic interaction. Additionally, we propose an FK-steered sampling strategy that leverages intermediate observation likelihoods to guide the generative trajectories, effectively improving the sample quality during inference. Our approach outperforms state-of-the-art baselines on standard SBI benchmarks, producing posteriors that closely match ground-truth MCMC. Furthermore, in a real-world exoplanet orbital estimation task, FUSE successfully resolves complex parameter degeneracies that challenge existing methods, highlighting its potential to accelerate complex scientific discoveries in astrophysics and beyond.
作者Stanislas Strasman, Gabriel Victorino Cardoso, Sylvain Le Corff, Vincent Lemaire, Antonio Ocello
Sequential Monte Carlo (SMC) methods are a natural tool for post-hoc conditioning of pretrained generative models, but in many applications the mutation kernels used by the particle system are biased approximations of an ideal Feynman--Kac flow. This paper develops a non-asymptotic error analysis for such SMC samplers. Under forward-smoothing forgetting conditions, we decompose the total error into a kernel bias, measuring the effect of replacing the ideal transition kernels by approximate ones, and a finite-particle Monte Carlo error. Our approach relies on extending local Doeblin-type conditions and Lyapunov drift arguments for Markov kernels to conditional distributions, thereby enabling a principled control of the bias. We then instantiate this general framework for conditional sampling with score-based diffusion models, and derive the first non-asymptotic error bound that jointly controls initialization error, time discretization, and score approximation in the reverse diffusion dynamics as well as finite-particle Monte Carlo error.
展开完整摘要收起摘要↓
Sequential Monte Carlo (SMC) methods are a natural tool for post-hoc conditioning of pretrained generative models, but in many applications the mutation kernels used by the particle system are biased approximations of an ideal Feynman--Kac flow. This paper develops a non-asymptotic error analysis for such SMC samplers. Under forward-smoothing forgetting conditions, we decompose the total error into a kernel bias, measuring the effect of replacing the ideal transition kernels by approximate ones, and a finite-particle Monte Carlo error. Our approach relies on extending local Doeblin-type conditions and Lyapunov drift arguments for Markov kernels to conditional distributions, thereby enabling a principled control of the bias. We then instantiate this general framework for conditional sampling with score-based diffusion models, and derive the first non-asymptotic error bound that jointly controls initialization error, time discretization, and score approximation in the reverse diffusion dynamics as well as finite-particle Monte Carlo error.
作者Stanislas Strasman, Sobihan Surendran, Sylvain Le Corff
Score-based Generative Models (SGMs) have achieved impressive performance in data generation across a wide range of applications. While the statistical properties of their sampling procedures are increasingly well understood, the optimization dynamics underlying their training remain less explored. SGMs are typically trained by minimizing a weighted denoising scorematching objective, yet optimization guarantees with stochastic gradients remain limited. In this work, we study Stochastic Gradient Descent (SGD) for SGMs, contributing results in two complementary regimes. First, for general score parameterizations, we establish a non-convex convergence rate for SGD on the weighted denoising score-matching objective, with explicit dependence on the schedule-dependent weighting factors. Second, for overparameterized two-layer ReLU networks, we develop a Neural Tangent Kernel analysis tailored to diffusion training with stochastic gradients, yielding score-approximation error bounds along the SGD trajectory. Finally, our analysis quantifies the role of the reweighting factor in the score approximation error, providing theoretical guidance for weighting choices used in practice.
展开完整摘要收起摘要↓
Score-based Generative Models (SGMs) have achieved impressive performance in data generation across a wide range of applications. While the statistical properties of their sampling procedures are increasingly well understood, the optimization dynamics underlying their training remain less explored. SGMs are typically trained by minimizing a weighted denoising scorematching objective, yet optimization guarantees with stochastic gradients remain limited. In this work, we study Stochastic Gradient Descent (SGD) for SGMs, contributing results in two complementary regimes. First, for general score parameterizations, we establish a non-convex convergence rate for SGD on the weighted denoising score-matching objective, with explicit dependence on the schedule-dependent weighting factors. Second, for overparameterized two-layer ReLU networks, we develop a Neural Tangent Kernel analysis tailored to diffusion training with stochastic gradients, yielding score-approximation error bounds along the SGD trajectory. Finally, our analysis quantifies the role of the reweighting factor in the score approximation error, providing theoretical guidance for weighting choices used in practice.
作者Bonan Shen, Jiazhou Gao, Tao Ning, Wei-Jung Huang, Xin Liu
CI/CD workflows have become executable operational policy: they decide what gets built, tested, released, and deployed, and they mediate how maintainers interact with delivery infrastructure. That makes them an important measurement point for cyber-systems engineering. Recent large language model (LLM) work shows that workflow stages can be recognized directly from configuration files, but stage labels alone do not tell us whether a workflow is brittle, unusual for its ecosystem, or worth revising first. We present an LLM-based CI/CD analysis pipeline that combines repository enrichment, anti-pattern detection, stage mining, and recommendation generation over a large GitHub corpus. Starting from 59,550 repositories with at least 1,000 stars, we identify 34,225 projects with CI/CD and collect 127,559 configuration files. Across 75,201 analyzed workflows, the anti-pattern detector reports 434,769 findings, dominated by reliability and maintainability issues. Across 59,906 configurations, stage usage differs significantly by language ($χ^2 = 4168.88$, $p < 0.001$, Cramer's $V = 0.063$), and domain analysis shows distinct operational profiles, including higher release and cache usage in mobile projects. For repository-level recommendation generation, few-shot prompting performs best overall, averaging 8.25 recommendations per repository with 96.1% YAML-valid snippets. Taken together, the results argue for CI/CD observability that combines diagnosis, context, and human review rather than treating workflow mining as a stage-classification problem alone.
展开完整摘要收起摘要↓
CI/CD workflows have become executable operational policy: they decide what gets built, tested, released, and deployed, and they mediate how maintainers interact with delivery infrastructure. That makes them an important measurement point for cyber-systems engineering. Recent large language model (LLM) work shows that workflow stages can be recognized directly from configuration files, but stage labels alone do not tell us whether a workflow is brittle, unusual for its ecosystem, or worth revising first. We present an LLM-based CI/CD analysis pipeline that combines repository enrichment, anti-pattern detection, stage mining, and recommendation generation over a large GitHub corpus. Starting from 59,550 repositories with at least 1,000 stars, we identify 34,225 projects with CI/CD and collect 127,559 configuration files. Across 75,201 analyzed workflows, the anti-pattern detector reports 434,769 findings, dominated by reliability and maintainability issues. Across 59,906 configurations, stage usage differs significantly by language ($χ^2 = 4168.88$, $p < 0.001$, Cramer's $V = 0.063$), and domain analysis shows distinct operational profiles, including higher release and cache usage in mobile projects. For repository-level recommendation generation, few-shot prompting performs best overall, averaging 8.25 recommendations per repository with 96.1% YAML-valid snippets. Taken together, the results argue for CI/CD observability that combines diagnosis, context, and human review rather than treating workflow mining as a stage-classification problem alone.
Long-lived AI agents require continuity across interactions, but continuity cannot be obtained by simply extending the prompt window. An agent must preserve useful prior experience, retrieve it selectively, distinguish personal context from external evidence, and revise memory when the underlying situation changes. We propose an architectural memory substrate organized along two orthogonal axes: a representational axis spanning structured records, vector representations, and graph relations; and a temporal axis spanning short-term traces, medium-term abstractions, and long-term semantic commitments. Its key design constraint is synchronized structured-vector-graph memory: structured records govern eligibility, vector representations support recall, and graph relations adjudicate support, contradiction, and supersession before gated context projection. Its central claim is that reliable personalization is a memory design problem: useful memory is structured, selectively exposed, continuously consolidated, and epistemically labeled rather than stored as undifferentiated conversation history. Beyond the framework, we instantiate MRMS as a lightweight prototype implementing structured records, vector retrieval, temporal policies, and graph-based revision. The prototype exercises the core substrate mechanisms through pre-generation memory selection, revision, boundary enforcement, and evidence attribution under controlled long-lived interaction scenarios with explicit evidence requirements.
展开完整摘要收起摘要↓
Long-lived AI agents require continuity across interactions, but continuity cannot be obtained by simply extending the prompt window. An agent must preserve useful prior experience, retrieve it selectively, distinguish personal context from external evidence, and revise memory when the underlying situation changes. We propose an architectural memory substrate organized along two orthogonal axes: a representational axis spanning structured records, vector representations, and graph relations; and a temporal axis spanning short-term traces, medium-term abstractions, and long-term semantic commitments. Its key design constraint is synchronized structured-vector-graph memory: structured records govern eligibility, vector representations support recall, and graph relations adjudicate support, contradiction, and supersession before gated context projection. Its central claim is that reliable personalization is a memory design problem: useful memory is structured, selectively exposed, continuously consolidated, and epistemically labeled rather than stored as undifferentiated conversation history. Beyond the framework, we instantiate MRMS as a lightweight prototype implementing structured records, vector retrieval, temporal policies, and graph-based revision. The prototype exercises the core substrate mechanisms through pre-generation memory selection, revision, boundary enforcement, and evidence attribution under controlled long-lived interaction scenarios with explicit evidence requirements.