Quantitative Bounds for Length Generalization in Transformers

/0 Comments/in /by

We study the problem of length generalization (LG) in transformers: the ability of a model trained on shorter sequences to maintain performance when evaluated on much longer, previously unseen inputs. Prior work by Huang et al. (2025) established that transformers eventually achieve length generalization once the training sequence length exceeds some finite threshold, but left open the question of how large it must be. In this work, we provide the first quantitative bounds on the required training length for length generalization to occur. Motivated by previous empirical and theoretical work, we analyze LG in several distinct problem settings: error control vs. average error control over an input distribution, infinite-precision softmax attention vs. finite-precision attention (which reduces to an argmax) in the transformer, and one- vs. two-layer transformers. In all scenarios, we prove that LG occurs when the internal behavior of the transformer on longer sequences can be “simulated” by its behavior on shorter sequences seen during training. Our bounds give qualitative estimates for the length of training data required for a transformer to generalize, and we verify these insights empirically. These results sharpen our theoretical understanding of the mechanisms underlying extrapolation in transformers, and formalize the intuition that richer training data is required for generalization on more complex tasks.

Read more

Sound Event Classification meets Data Assimilation with Distributed Fiber-Optic Sensing

/in /by

Distributed Fiber-Optic Sensing (DFOS) is a promising technique for large-scale acoustic monitoring. However, its wide variation in installation environments and sensor characteristics causes spatial heterogeneity. This heterogeneity makes it difficult to collect representative training data. It also degrades the generalization ability of learning-based models, such as fine-tuning methods, under a limited amount of training data. To address this, we formulate Sound Event Classification (SEC) as data assimilation in an embedding space. Instead of training models, we infer sound event classes by combining pretrained audio embeddings with simulated DFOS signals. Simulated DFOS signals are generated by applying various frequency responses and noise patterns to microphone data, which allows for diverse prior modeling of DFOS conditions. Our method achieves out-of-domain (OOD) robust classification without requiring model training. The proposed method achieved accuracy improvements of 6.42, 14.11, and 3.47 percentage points compared with conventional zero-shot and two types of fine-tune methods, respectively. By employing the simulator in the framework of data assimilation, the proposed method also enables precise estimation of physical parameters from observed DFOS signals.

Read more

SlideCraft: Context-aware Slides Generation Agent

/in /by

Creating effective slide presentations requires adapting both content and structure to match the communication context e.g. whether the presentation is for summarizing to executives, or reporting progress to research supervisors. In research and enterprise environments, this need for context-sensitive presentations often leads to repeated, manual reformatting of the same material to suit different audiences. Existing generative systems support slide creation but typically rely on structured inputs, assume a fixed format, and offer limited ability to iteratively refine outputs through natural language feedback. Moreover, they rarely accommodate organizational constraints such as formatting guidelines, domain-specific terminology, or branding requirements. We present SlideCraft, a context-aware generative agent that autonomously creates and edits slide presentations based on natural language instructions. SlideCraft infers the intended presentation context, such as an executive-facing or a project review summary for technical oversight, and selects the appropriate slide template. It then synthesizes content from input documents, enriches it with external knowledge and internal assets, assembles it into a structured intermediate representation, and generates a validated slide deck. SlideCraft supports both first-time slide creation and iterative updates, operating through familiar natural language interfaces like email or messaging tools. Our experiments demonstrate that SlideCraft consistently produces high-quality, context-aware presentations tailored to diverse communication settings, with minimal human input and reliable adherence to enterprise constraints.

Read more

TalentScout: Multimodal AI-Driven Expert Finding in Organizations

/in /by

Identifying subject-matter experts within organizations remains a challenging task due to the scale, heterogeneity, and unstructured nature of enterprise knowledge assets. We present TalentScout, an AI-driven expert identification system that constructs a unified, skill-centric knowledge graph by ingesting and analyzing diverse media, including research papers, reports, presentations, transcripts, and supervisor recommendations. TalentScout’s modular architecture integrates document parsing, audio/video transcription, metadata extraction, large language model-based skill extraction, multi-factor author disambiguation, and evidence-weighted skill attribution. At query time, TalentScout decomposes natural language queries into canonical skill requirements, traverses the constructed knowledge graph, and ranks experts based on aggregated skill weights, document quality, and endorsement signals, providing document-level justifications for each recommendation. We evaluate TalentScout on multiple public and internal enterprise datasets, including DBLP, TREC Enterprise, Tilburg, and ManConCorpus. Using standard information retrieval metrics such as Precision@ 5, Recall@5, nDCG@5, and Mean Reciprocal Rank (MRR), TalentScout consistently outperforms leading baselines, achieving up to 24% higher Precision@ 5 in early expert retrieval. The results highlight TalentScout’s scalability, transparency, and accuracy, establishing it as a practical solution for evidence-based expert discovery and organizational talent management.

Read more

AutoScape: Geometry-Consistent Long-Horizon Scene Generation

/in /by

This paper proposes AutoScape, a long-horizon driving scene generation framework. At its core is a novel RGB-D diffusion model that iteratively generates sparse, geometrically consistent keyframes, serving as reliable anchors for the scenes appearance and geometry. To maintain long-range geometric consistency, the model 1) jointly handles image and depth in a shared latent space, 2) explicitly conditions on the existing scene geometry (i.e., rendered point clouds) from previously generated keyframes, and 3) steers the sampling process with a warp-consistent guidance. Given high-quality RGB-D keyframes, a video diffusion model then interpolates between them to produce dense nd coherent video frames. AutoScape generates realistic and geometrically consistent driving videos of over 20 seconds, improving the long-horizon FID and FVD scores over the prior state-of-the-art by 48.6% and 43.0%, respectively.

Read more

DWIM: Towards Tool-aware Visual Reasoning via Discrepancy-aware Workflow Generation & Instruct-Masking Tuning

/in /by

Visual reasoning (VR), which is crucial in many fields for enabling human-like visual understanding, remains highly challenging. Recently, compositional visual reasoning approaches, which leverage the reasoning abilities of large language models (LLMs) with integrated tools to solve problems, have shown promise as more effective strategies than end-to-end VR methods. However, these approaches face limitations, as frozen LLMs lack tool awareness in VR, leading to performance bottlenecks. While leveraging LLMs for reasoning is widely used in other domains, they are not directly applicable to VR due to limited training data, imperfect tools that introduce errors and reduce data collection efficiency in VR, and challenges in fine-tuning on noisy workflows. To address these challenges, we propose DWIM: i) Discrepancy-aware training Workflow generation, which assesses tool usage and extracts more viable workflows for training; and ii) Instruct-Masking fine-tuning, which guides the model to only clone effective actions, enabling the generation of more practical solutions. Our experiments demonstrate that DWIM achieves state-of-the-art performance across various VR tasks, exhibiting strong generalization on multiple widely used datasets.

Read more

LANGTRAJ: Diffusion Model and Dataset for Language-Conditioned Trajectory Simulation

/in /by

Evaluating autonomous vehicles with controllability enables scalable testing in counterfactual or structured settings, enhancing both efficiency and safety. We introduce LangTraj, a language-conditioned scene-diffusion model that simulates the joint behavior of all agents in traffic scenarios. By conditioning on natural language inputs, LangTraj provides flexible and intuitive control over interactive behaviors, generating nuanced and realistic scenarios. Unlike prior approaches that depend on domain-specific guidance functions, LangTraj incorporates language conditioning during training, facilitating more intuitive traffic simulation control. We propose a novel closed-loop training strategy for diffusion models, explicitly tailored to enhance stability and realism during closed-loop simulation. To support language-conditioned simulation, we develop Inter-Drive, a large-scale dataset with diverse and interactive labels for training language-conditioned diffusion models. Our dataset is built upon a scalable pipeline for annotating agent-agent interactions and single-agent behaviors, ensuring rich and varied supervision. Validated on the Waymo Motion Dataset, LangTraj demonstrates strong performance in realism, language controllability, and language-conditioned safety-critical simulation, establishing a new paradigm for flexible and scalable autonomous vehicle testing. Project website: https://langtraj.github.io/.

Read more

Mapillary Vistas Validation for Fine-Grained Traffic Signs: A Benchmark Revealing Vision-Language Model Limitations

/in /by

Obtaining high-quality fine-grained annotations for traffic signs is critical for accurate and safe decision-making in autonomous driving. Widely used datasets, such as Mapillary, often provide only coarse-grained labels without distinguishing semantically important types such as stop signs or speed limit signs. To this end, we present a new validation set for traffic signs derived from the Mapillary dataset called Mapillary Vistas Validation for Traffic Signs (MVV), where we decompose composite traffic signs into granular, semantically meaningful categories. The dataset includes pixel-level instance masks and has been manually annotated by expert annotators to ensure label fidelity. Further, we benchmark several state-of-the-art VLMs against the self-supervised DINOv2 model on this dataset and show that DINOv2 consistently outperforms all VLM baselines not only on traffic sign recognition, but also on heavily represented categories like vehicles and humans. Our analysis reveals significant limitations in current vision-language models for fine-grained visual understanding and establishes DINOv2 as a strong baseline for dense semantic matching in autonomous driving scenarios. This dataset and evaluation framework pave the way for more reliable, interpretable, and scalable perception systems. Code and data are available at: https://github.com/nec-labs-ma/relabeling

Read more

THAT: Token-wise High-frequency Augmentation Transformer for Hyperspectral Pansharpening

/in /by

Transformer-based methods have demonstrated strong potential in hyperspectral pansharpening by modeling long-range dependencies. However, their effectiveness is often limited by redundant token representations and a lack of multiscale feature modeling. Hyperspectral images exhibit intrinsic spectral priors (e.g., abundance sparsity) and spatial priors(e.g., non-local similarity), which are critical for accurate reconstruction. From a spectral–spatial perspective, Vision Transformers (ViTs) face two major limitations: they struggle to preserve high-frequency components—such as material edges and texture transitions, and suffer from attention dispersion across redundant tokens. These issues stem from the global self-attention mechanism, which tends to dilute high-frequency signals and overlook localized details. To address these challenges, we propose the Token-wise High-frequency AugmentationTransformer (THAT), a novel framework designed to enhance hyperspectral pansharpening through improved high-frequency feature representation and token selection. Specifically, THAT introduces: (1) Pivotal Token Selective Attention (PTSA) to prioritize informative tokens and suppress redundancy; (2) a Multi-level Variance-aware Feed-forward Network (MVFN) to enhance high-frequency detail learning. Experiments on standard benchmarks show that THAT achieves state-of-the-art performance with improved reconstruction quality and efficiency.

Read more

EditGRPO: Reinforcement Learning with Post-Rollout Edits for Clinically Accurate Chest X-Ray Report Generation

/in /by

Radiology report generation requires advanced medical image analysis, effective temporal reasoning, and accurate text generation. Although recent innovations, particularly multimodal large language models (MLLMs), have shown improved performance, their supervised fine-tuning (SFT) objective is not explicitly aligned with clinical efficacy. In this work, we introduce EditGRPO, a mixed-policy reinforcement learning (RL) algorithm designed specifically to optimize the generation through clinically motivated rewards. EditGRPO integrates on-policy exploration with off-policy guidance by injecting sentence-level detailed corrections during training rollouts. This mixed-policy approach addresses the exploration dilemma and sampling efficiency issues typically encountered in RL. Applied to a Qwen2.5-VL-3B MLLM initialized with supervised fine-tuning (SFT), EditGRPO outperforms both SFT and vanilla GRPO baselines, achieving an average improvement of 3.4% in CheXbert, GREEN, Radgraph, and RATEScore metrics across four major chest X-ray report generation datasets. Notably, EditGRPO also demonstrates superior out-of-domain generalization, with an average performance gain of 5.9% on unseen datasets.

Read more