Decompose-then-verify pipelines (FActScore-style fact-checkers, hallucination detectors) first split a passage into atomic claims and treat this step as neutral preprocessing. We ask whether a decomposer can substitute its own parametric belief for what the passage says, producing a claim that contradicts its source, and call this Decomposition-Induced Context-Memory Conflict (DI-CC). On biographies with a deliberately altered year, Qwen2.5-7B-Instruct does so in 4.3% of claims when the instruction licenses self-correction, but in only 0.2-0.5% under a neutral instruction, a FActScore-style few-shot prompt, a "use only the passage" instruction, a formatting instruction, or a knowledge-filling instruction (an upper bound: in the neutral and FActScore-style runs only 1 of 13 flagged claims is a real override); context-aware decoding removes only half. DI-CC here is therefore instruction-induced. We then test whether a linear probe trained only on classical context-memory conflict data (NQ-Swap) detects DI-CC zero-shot. Its AUC is high (up to 0.94 against legit claims, 0.98 under a position- and instruction-matched control), but random directions reach AUCs between about 0.1 and 0.9 on these contrasts, probes fit directly on the claims reach 0.96-1.00 everywhere (even on a control that separates only the text variant), and after multiplicity correction the NQ-Swap probe beats random directions only at the last layer, and only for two of three random-direction families. A direct NLI check between source and claim reaches AUC 0.998. High probe AUC on such small contrasts is not evidence of a shared mechanism without nulls, and two pitfalls, pair leakage in probe training and duplicated BOS tokens, can change such conclusions.
Detecting a traffic anomaly does not establish whether a video-language model can explain what happened, localize it in time, or identify its causes. We introduce TAR (Traffic Anomaly Reasoning) and TAR-Bench, paired resources for training and evaluating these complementary capabilities across 10 tasks spanning question answering, temporal reasoning, and scene understanding. TAR contains 44,040 automatically generated annotations with chain-of-thought traces for 3,670 CCTV videos from eight public datasets. TAR-Bench provides 960 human-curated annotations for 80 held-out clips from 17 public YouTube videos. Evaluation of eleven vision-language models and analysis of fine-tuned outputs reveal temporal, causal, and descriptive errors not captured by closed-form question-answering accuracy. Training on all 10 tasks raises mean benchmark scores from 34.3 to 55.7 for Cosmos-Reason2-8B and from 30.9 to 53.9 for Qwen3-VL-8B-Instruct. These results support joint training across complementary tasks, while highlighting persistent limitations in temporal precision and causal attribution. TAR and TAR-Bench serve as the official training and in-domain evaluation resources for AI City Challenge 2026 Track~3. The dataset is available under CC-BY-4.0 at https://huggingface.co/datasets/nvidia/PhysicalAI-Traffic-Anomaly-Reasoning.
All-in-one image restoration is a unified low-level vision task that aims to effectively recover high-quality images from inputs degraded by various types and levels of corruption using a single model. Recent works have achieved remarkable progress by learning degradation-adaptive prompts or network architectures. However, these methods typically apply a uniform restoration strategy across the entire image, neglecting the fact that different regions may suffer from distinct degradation types and varying degrees of severity. In contrast, we propose to perform restoration at the pixel level, thereby enabling more fine-grained and precise control over the restoration process. Specifically, we present MGN-AIR, a novel pixel-level restoration framework for all-in-one image restoration. Our approach first learns to estimate a pixel-level visual prompt. Then, it leverages both textual and visual prompts to provide global and local degradation cues, guiding the model on where to look and how to restore at each pixel. We conduct extensive experiments on multiple all-in-one image restoration benchmarks, covering a wide range of tasks including denoising, deraining, deblurring, dehazing, desnowing, and low-light enhancement. Experimental results demonstrate that our proposed method consistently and significantly outperforms existing approaches.
Monocular 3D object detection spans two regimes: closed-set detectors operating within a fixed category vocabulary, and open-vocabulary detectors that localize arbitrary categories by leveraging depth foundation models for 3D geometry. We find that current depth foundation models, despite their strong zero-shot generalization, lack the object-level precision 3D detection demands: substituting a state-of-the-art depth foundation model for a strong detector's predicted depth degrades accuracy. Rather than pushing detectors or depth models to be more accurate end-to-end, we treat object-level depth refinement as a stand-alone task and present RefineAny3D, a vision-language model that corrects depth without ever predicting a numerical value. Our key insight is that depth error has a direct visual signature in image space: when projected onto the image, a correctly placed box tightly encloses the object, while a too-far box projects too small and a too-close box projects too large. Depth refinement thus reduces to a visual alignment problem rather than a metric regression problem, which we instantiate by extending the VLM's vocabulary with action tokens that replace numerical depth output with categorical decisions, and by supervising the model on a large-scale chain-of-thought dataset that grounds each decision in explicit visual evidence. Applied as a single post-hoc step, RefineAny3D delivers consistent gains across closed-set detectors, open-vocabulary detectors, and 3D auto-labeling tools, and generalizes to novel categories, scenes, and cameras without retraining. Code is available at https://github.com/alanzhangcs/RefineAny3D.
Image super-resolution (SR) with large generative models has recently achieved remarkable perceptual quality, yet maintaining fidelity to the LR observation remains challenging. In particular, we observe that diffusion transformers (DiTs) built on latent representations suffer from a critical limitation: the compression bottleneck of the VAE weakens fine-grained spatial information, leading to hallucinated details that are weakly grounded in the input image. In this work, we revisit generative SR from a representation perspective and propose a pixel-grounded super-resolution (PGSR) framework that preserves LR-observed pixel evidence before VAE compression and reuses it throughout restoration. Instead of relying solely on the compressed latent condition, PGSR extracts pre-VAE pixel evidence from the upsampled LR image and reuses it at two stages. First, Condition-Side Trajectory Guidance fuses LR-derived pixel evidence with the latent LR condition to guide the latent restoration trajectory. Second, Decoder-Side Pixel Grounding injects multi-scale pixel features into the frozen VAE decoder to ground the final rendering with LR-observed cues. To efficiently adapt large pretrained DiT models, we keep the latent autoencoder and main flow-matching backbone frozen, and train only lightweight restoration modules. We further study an efficient local-window attention variant for improved high-resolution efficiency and scalability. Extensive experiments demonstrate that PGSR improves the realism--fidelity trade-off and produces more faithful, visually convincing results than existing latent generative SR approaches.
Large Language Model (LLM) agents are increasingly deployed in multi-agent social settings where they must cooperate, negotiate, and adapt to other agents. Measuring these social skills is hard because most social interactions offer no objective ground truth. Evaluations often fall back on LLM judges, which are costly, subjective, and noisy. We introduce SocialGym, an environment of 22 multi-agent social games, from matrix games to hidden-role deduction, with verifiable outcomes determined by game rules. We benchmark ten LLMs across all games and roles using an Elo tournament. Our results show that no model is best in every game, and rankings shift across game families and sides. Secret keeping correlates more strongly with theory-of-mind and privacy benchmarks than with general knowledge benchmarks. Half of the models give away their side's secret in Chameleon and Spyfall, and the smallest models are not reliably better than chance at voting for hidden opponents in most hidden-role games. Finally, we examine whether models can improve their play by reflecting on self-play and writing strategy playbooks, finding mixed results across models and games. SocialGym provides verifiable measures for diagnosing LLM social capabilities across settings and roles.
A single model scale challenges the flexibility of a production retrieval system: some settings need it faster, others need a smaller index, and the right trade-off changes with the workload. In the context of information retrieval (IR), a transformer-based model can be made smaller in three ways --- using fewer layers, passing fewer tokens through the upper layers, or producing a shorter embedding --- and each way saves a different resource. These options have been studied in isolation, each as its own method with its own code and training setup, which makes them hard to combine or adapt to a new model. In this work, we present Tevatron-Elastic to bring all three under one unified abstraction: a single object names any size the model can run at, and a short schedule lists the sizes to train. Training then produces one checkpoint that serves all of those sizes, and at deployment the user has the liberty to pick any of them. The same abstraction covers both retrievers and rerankers and both encoder and decoder models, as it works through interfaces that Hugging Face transformers already expose; adding a new backbone can be a configuration change rather than a model rewrite. Prior methods --- Matryoshka embeddings, early exit, 2D Matryoshka (Starbucks), and layerwise token compression (LTC) --- become special cases of our unified abstraction. With Tevatron-Elastic, we also propose a new elastic algorithm Matryoshka~LTC (MLTC) to jointly train several token-compression ratios in one retriever checkpoint. To validate our framework, we train 20 checkpoints across three backbones and two tasks: the quality curves are smooth, full-point quality remains close to that of models trained for a single size, and a controlled study confirms the wall-clock speedups. We release the framework and all checkpoints to support elastic retrieval research.
Interpretability is often treated as a tax on capability: language models are trained as opaque systems, then explained after the fact, with methods whose reliability is difficult to establish. In this work, we challenge this premise. Rather than reverse-engineering a model, we make interpretability a constraint of the training pipeline, optimized alongside the language modeling objective. Across three orders of magnitude of compute, on both autoregressive and diffusion language models, interpretability scales with capability rather than against it. Surprisingly, model representations become more disentangled and aligned with human-understandable concepts with scale. We instantiate the training-time recipe with Steerling-8B, a diffusion language model with a causal attention mask. For any group of generated tokens, Steerling-8B attributes the output to relevant input tokens, human-understandable concepts, and training data. This enables closed-loop intervention: diagnose an output through its concept or feature attribution, retrieve similar training data, and correct the behavior through concept steering without retraining. Steerling-8B remains competitive with open peer models trained on substantially 2-16x more compute, suggesting a different scaling paradigm: interpretability can be designed into training, and it improves with scale.
Large language models (LLMs) are increasingly involved in scientific discovery, yet it remains unclear whether they can support complex real laboratory science. Here we introduce Science Edge Evaluation (SEE), a multimodal benchmark of expert-curated questions grounded in peer-reviewed literature and experimental practice in chemistry, biology, and materials science. Evaluation of 28 multimodal large language models (MLLMs) shows that even the best-performing model reaches only 54.9% accuracy. Moreover, general-purpose models outperform science-specialized models on average. In the visual-agent evaluation, the use of tools increases the best accuracy to 56.1%. Tool use can expand the information available to models, but more information does not necessarily lead to reliable scientific reasoning. The key challenge is whether models can manage tool-derived information within the boundaries of the original experimental evidence. Together, these findings reveal that current MLLMs still cannot reliably make justified and evidence-bounded inferences from experimental results, which is an essential capability in real scientific discovery. Bridging this gap requires MLLMs to transition from explaining established scientific concepts to deriving novel and evidence-based insights from experimental data.
In overparameterised classification, training data can be linearly separable even when the underlying distribution is not. In this setting, gradient descent (GD) on the logistic loss diverges in norm while converging in direction to a max-margin interpolating classifier, whose implicit bias can be statistically suboptimal. In this work, we show that early stopping can overcome this suboptimality: in a Gaussian mixture model with label-flipping noise, GD stopped at an appropriate oracle time achieves minimax-optimal excess zero-one risk for covariance spectra with fast and continuous decay, including polynomial and exponential spectral decays. Our analysis combines a sharp upper bound for the early-stopped iterate with a matching statistical lower bound over arbitrary classifiers, yielding optimal rates that are validated by experiments. A central technical contribution is a new calibration result that converts excess logistic risk into excess zero-one risk; it handles the model misspecification induced by the label-flipping noise, and removes the square-root rate in standard bounds. We also establish a lower bound for linear interpolators, showing that interpolation can require exponentially more samples than early stopping to achieve the same excess risk.
Large Language Models (LLMs) are integrated into software systems and AI services, making efficient LLM serving a concern for software engineering. Serving LLMs is challenging because inference requires computation, memory, GPU resources, and execution while maintaining latency and throughput. Although prior research has proposed LLM inference, optimization, and serving techniques and frameworks, little is known about how they are adopted in practice. In this study, we investigate the use of LLM serving frameworks and serving methods in open-source software systems. We identify and analyze five LLM-specific frameworks: vLLM, SGLang, TensorRT-LLM, LMDeploy, and FlashInfer. We examine how these frameworks and techniques are adopted individually and in combination, how adoption varies across categories of LLMs, and how repositories differ in intent, focus, use case, and architectural design. Our results show that vLLM is the most visible framework in popularity and adoption, while parallel computation, memory management, and network pruning are the most frequently used serving-method categories. Multi-framework usage is limited, suggesting that developers rely on a single serving framework; however, combined frameworks connect complementary capabilities across the serving stack. Framework adoption varies across model families, modalities, model sizes, domain specializations, and deployment settings. Repository-level analysis shows that LLM serving frameworks support applications and architectures, including Reinforcement Learning (RL)-based reasoning, multimodal generation and understanding, microservices, and cloud infrastructure. Overall, this study provides a large-scale empirical characterization of LLM serving framework adoption in practice and offers insights for researchers, framework maintainers, and practitioners working on LLM systems.
Time series forecasting is fundamental to temporal data mining, yet numerical models often struggle to account for changing contextual factors. A key challenge is determining when contextual evidence should guide forecast revisions while maintaining consistency with temporal patterns and domain knowledge. To address this, we propose CastFSR, a fast--slow--reflect agentic framework for context-aware time series forecasting. Fast-thinking forecasting analyzes numerical patterns, selects a suitable forecaster, and constructs a data-driven forecast prior. Slow deliberative reasoning incorporates relevant contextual evidence to selectively refine this prior, while reflective evaluation checks the revised forecast against temporal patterns, contextual evidence, and domain constraints. A shared state maintains forecasting priors, contextual evidence, intermediate forecasts, and feedback across stages. CastFSR supports both training-free LLM coordination and two-stage post-training for compact LLM coordinators. Workflow-aware trajectory distillation (WATD) learns state-conditioned tool use and forecasting decisions from teacher trajectories, while trajectory-level policy optimization (TLPO) further improves the coordinator through reinforcement learning guided by forecasting accuracy and workflow feedback. Experiments on multiple real-world datasets under an offline evaluation protocol using historical context demonstrate competitive performance against representative baselines. Our code is available at https://github.com/Xiaoyu-Tao/CastFSR.
The widespread adoption of proprietary Large Language Models (LLMs) via closed APIs poses an interpretability challenge. While white-box and gray-box (logit-dependent) attribution methods require internal model access, post-hoc attribution under a strictly black-box, text-only generative regime remains largely unexplored. To address this, we propose a model-agnostic, post-hoc attribution interpreter operating at the sentence level. We first train a transformer-based Energy-Based Model (EBM) as a surrogate to score the consistency between prompt and response sentences based on observed LLM generations. To prevent energy collapse and avoid a focus on superficial artifacts, we design contrastive negative samplers that evaluate partial counterfactual dependencies across prompt and response segments. This learned energy landscape then guides the training of a lightweight interpreter network. After model-specific training, the interpreter operates as a standalone tool; it estimates the influence of prompt sentences on a user-specified output sentence without requiring runtime API queries to the target LLM. Comprehensive experiments demonstrate that our EBM surrogate captures nuanced generative dependencies beyond surface-level heuristics, while the interpreter achieves competitive attribution quality with parameter-efficient, $\mathcal{O}(1)$ inference.
Tool-using large language model (LLM) agents produce long, multi-turn trajectories, making gradient-based post-training memory-intensive. Evolution strategies (ES) enable memory-efficient full-parameter post-training without backpropagation and can eventually match the performance of gradient-based reinforcement learning (RL). However, resource-constrained settings typically offer only a few GPUs, so the high GPU-hour requirements of ES translate into prohibitively long training times. To address this, we introduce Cooperative Parameter-subspace Evolution Strategy (CoPES), a cooperative coevolutionary method that decomposes the full parameter space into lower-dimensional subspaces and searches over them cooperatively to improve optimization efficiency. On Qwen3.5-4B math, CoPES recovers 92% of full-parameter GRPO's validation-accuracy gain within the GPU-hour budget of GRPO's best validation checkpoint, versus 67% for standard ES. Its theoretical GPU memory requirement is less than one-eighth that of full-parameter GRPO. It consistently outperforms standard ES and LoRA-based GRPO on all evaluated pass@k metrics across five math benchmarks. CoPES's gains extend to Qwen3.5-9B with unchanged optimization hyperparameters and to multi-hop QA. These results demonstrate an improved trade-off between memory requirements and training time for agentic LLM post-training under resource constraints. The code is open-sourced in https://github.com/MetaronWang/CoPES
Radio resource management in dense networks is subject to budgets that must hold on every realized trajectory, carried by different agent groups: an energy budget at each base station (BS) and a handover budget at each user, coupled through the association decisions that consume both. Constrained multi-agent reinforcement learning (MARL) handles such budgets using Lagrangian multipliers updated across episodes or Lyapunov virtual queues updated per slot, but their relative behavior under coupled two-sided constraints remains unclear. We formulate this problem and introduce a windowed proportional-fairness objective that captures the temporal distribution of service beyond the classical horizon-average formulation. We characterize the distinct budget-regulation timescales of the two mechanisms through complementary theoretical bounds and identify queue-price buildup as a key mechanism governing finite-horizon budget pacing. Using a common heterogeneous MARL architecture, we compare three utility surrogates and both constraint-handling mechanisms. The results show that utility design primarily determines throughput and fairness, whereas constraint-price dynamics govern training efficiency and budget pacing. Under the default setting, the queue-based PF policy reaches near-final performance after a single training episode, but its unscaled penalty can lead to delayed pacing under tight budgets. Guided by the analysis, we show that queue-price calibration reduces the empirical drift constant by an order of magnitude and substantially narrows the performance gap to Lagrangian methods, which adapt constraint prices across episodes at higher training cost. These findings provide practical design guidelines for constrained MARL.
We propose an instantiation of flow matching that relies on a time-independent velocity field---an autonomous flow---to exactly map between two distributions when the target is supported on a sufficiently lower-dimensional data manifold. We also show that the associated one-step generative map is constant in the direction of the autonomous velocity field and therefore satisfies a simple conservation equation, which motivates learning the map directly from samples. These autonomous flows and maps give a dynamical meaning to the flux constraint of Beckmann's transportation problem, and we show that, under suitable conditions, the construction extends beyond flow matching to other probability fluxes satisfying the same constraint. The resulting framework includes, for instance, a closed-form Coulomb transport closely related to Poisson Flow and a flow-matching-consistent version of Equilibrium Matching. We illustrate these distinctions and demonstrate the effectiveness of the autonomous flow and the one-step map on ImageNet $256\times256$.
Diffusion Transformers (DiTs) have emerged as a core architecture in generative modeling due to their scalability and adaptability to multimodal tasks. DiTs comprise isotropic transformer blocks, and learn representations progressively across depth, where the denoising objective drives later layers to focus on fine-detail reconstruction. This results in degraded representation quality and an imbalanced encoder-decoder behavior. Prior approaches such as representation alignment (REPA) mitigate this by encouraging stronger early representations via training regularization. Alternatively, U-Net-style DiT architectures introduce explicit multi-scale encoder-decoder structures for improved convergence. But they build on standard U-Net wisdom via learnable operators for spatial downsampling, which are not well-suited to transformer architectures, introducing inefficiencies and compatibility issues with components such as cross-attention and representation regularization. In this work, we propose UDT, a U-Net diffusion transformer that combines the representation power of DiTs with the encoding-decoding benefits of U-Nets, through data-adaptive token merging for downsampling and upsampling, while preserving the DiT token dimension. Our baseline UDT architecture outperforms existing U-Net DiTs and achieves performance comparable to REPA across all model sizes. Furthermore, using architectural optimization and REPA, UDT outperforms SiT's 7.9 FID at 1400 epochs (w/o CFG) within 40 epochs (~ 40x faster convergence) for XL model size on 256x256 ImageNet. Finally, it achieves strong image generation performance with CFG, reaching FID of 1.38 (320 epochs) with SD-VAE and 1.35 (500 epochs) with VA-VAE, providing a new backbone for DiTs with strong empirical benefits. Code is available at https://github.com/JN-Yun/UDT
Large language model agents have shown strong capabilities in generating coherent and contextually appropriate responses, yet robust long-horizon dialogue remains limited by the lack of external memory that is traceable, updatable, and diagnostically transparent. Existing memory-augmented agents often store memories as isolated records or overwritable states, making it difficult to preserve how information originates, evolves, conflicts, or becomes obsolete over time. We propose \textbf{TrajWiki}, a trajectory-based memory framework for long-horizon conversational agents. Instead of treating memory as static entries, TrajWiki represents each memory as a source-grounded evolution trajectory, maintained through immutable episodic snapshots and claim-level operations such as \texttt{ADD}, \texttt{REVISE}, and \texttt{DEPRECATE}. To reduce fragmentation and retrieval cost, TrajWiki further introduces \textbf{Memory Wiki}, a persistent intermediate layer that incrementally compiles dialogue history into structured and interlinked wiki pages capturing salient entities, events, quantities, topics, and conflicts. At inference time, queries are routed hierarchically from relevant wiki pages to linked memory trajectories, then to corresponding snapshots and source messages for evidence-grounded answer synthesis. Experiments on LoCoMo and MedMT show that TrajWiki improves long-horizon dialogue performance across both open-source and closed-source LLM backbones, while providing greater interpretability and diagnostic visibility into memory evolution, retrieval failures, and answer generation.
Training large dense and mixture-of-experts (MoE) rerankers is constrained by GPU memory and distributed-execution efficiency, which has kept academic reranker research largely on dense models. Tevatron trains rerankers through the Hugging Face Trainer with DeepSpeed and PyTorch FSDP1, a stack that stalls on three fronts: DeepSpeed ZeRO-2 has a known gradient-gathering failure on this path, FSDP1 full sharding is memory-safe but throughput-bound, and neither offers the expert parallelism (EP) that MoE training needs. We present Tevatron 3.0, which integrates Megatron-Core as a reranker training backend while preserving Tevatron's data processing, evaluation pipeline, and Hugging Face checkpoint format. For full-parameter dense 8B training, Megatron matches FSDP1 on BEIR-15 NDCG@10 with approximately 4% lower epoch time and twice the peak GPU memory under matched data-parallel and recomputation settings, and approximately 22% lower epoch time in its recommended configuration; it supports both LoRA and full-parameter fine-tuning. Expert parallelism makes it possible to train a 30B-parameter MoE reranker (Qwen3-30B-A3B) on two nodes. A controlled study of architecture, LoRA, and distillation on BEIR-15, with three first-stage retrievers and Hugging Face and vLLM serving, shows that the MoE reranker matches dense-8B quality at less than half the activated parameters and at higher inference throughput. We open-source the framework and trained checkpoints.
Large language models have demonstrated strong mathematical problem-solving capabilities, yet reliably verifying their candidate answers remains challenging. Existing representative methods mainly revise outputs through natural-language reflection or assist verification by directly generating verification programs; the former may not reliably support exact computation, whereas the latter prematurely couples mathematical modeling with low-level implementation. We propose AMTFV (Agentic Mathematical Tool-Flow Verification). By introducing Mathematical Tool Flow (MTF) as an interrupt--execute--resume interface, AMTFV decouples verification modeling from concrete execution and supports exact computation through a mathematical toolbox. Specifically, the verification agent first constructs a verification workflow, encodes the mathematical objects and computational intent requiring reliable execution in an MTF request, and sends it to the mathematical toolbox agent. The latter parses the request, generates executable calls, and dispatches them to the backend for exact computation. Tool outputs then support candidate-answer adjudication, answer revision, and verification-workflow revision. We evaluate AMTFV on five challenging mathematical reasoning datasets with seven model configurations from DeepSeek, GPT, and Gemini. Experimental results show that AMTFV outperforms the representative baselines evaluated in this study overall; under an individual model configuration, it improves average accuracy over the strongest baseline by up to 8.3 percentage points, with larger gains on samples of medium and high verification complexity.