The choice of tokenizer can profoundly impact language model performance, yet accessible and reliable evaluations of tokenizer quality remain an open challenge. Inspired by scaling consistency, we show that smaller models can accurately predict significant differences in tokenizer impact on larger models at a fraction of the compute cost. By systematically evaluating both English-centric and multilingual tokenizers, we find that tokenizer choice has negligible effects on tasks in English but results in consistent performance differences in multilingual settings. We propose new intrinsic tokenizer metrics inspired by Zipf's law that correlate more strongly with downstream performance than text compression when modeling unseen languages. By combining several metrics to capture multiple aspects of tokenizer behavior, we develop a reliable framework for intrinsic tokenizer evaluations. Our work offers a more efficient path to informed tokenizer selection in future language model development.

In this paper, we present TalkingMachines -- an efficient framework that transforms pretrained video generation models into real-time, audio-driven character animators. TalkingMachines enables natural conversational experiences by integrating an audio large language model (LLM) with our video generation foundation model. Our primary contributions include: (1) We adapt a pretrained SOTA image-to-video DiT into an audio-driven avatar generation model of 18 billion parameters; (2) We enable infinite video streaming without error accumulation through asymmetric knowledge distillation from a bidirectional teacher model into a sparse causal, autoregressive student model; (3) We design a high-throughput, low-latency inference pipeline incorporating several key engineering optimizations such as: (a) disaggregation of the DiT and VAE decoder across separate devices, (b) efficient overlap of inter-device communication and computation using CUDA streams, (c) elimination of redundant recomputations to maximize frame-generation throughput. Please see demo videos here - https://aaxwaz.github.io/TalkingMachines/

Emerging embodied AI applications, such as wearable cameras and autonomous agents, have underscored the need for robust reasoning from first person video streams. We introduce EgoVLM, a vision-language model specifically designed to integrate visual comprehension and spatial-temporal reasoning within egocentric video contexts. EgoVLM is fine-tuned via Group Relative Policy Optimization (GRPO), a reinforcement learning method adapted to align model outputs with human-like reasoning steps. Following DeepSeek R1-Zero's approach, we directly tune using RL without any supervised fine-tuning phase on chain-of-thought (CoT) data. We evaluate EgoVLM on egocentric video question answering benchmarks and show that domain-specific training substantially improves performance over general-purpose VLMs. Our EgoVLM-3B, trained exclusively on non-CoT egocentric data, outperforms the base Qwen2.5-VL 3B and 7B models by 14.33 and 13.87 accuracy points on the EgoSchema benchmark, respectively. By explicitly generating reasoning traces, EgoVLM enhances interpretability, making it well-suited for downstream applications. Furthermore, we introduce a novel keyframe-based reward that incorporates salient frame selection to guide reinforcement learning optimization. This reward formulation opens a promising avenue for future exploration in temporally grounded egocentric reasoning.

Contrastive language-image pre-training aligns the features of text-image pairs in a common latent space via distinct encoders for each modality. While this approach achieves impressive performance in several zero-shot tasks, it cannot natively handle multimodal inputs, i.e., encoding image and text into a single feature vector. As a remedy, it is common practice to use additional modules to merge the features extracted by the unimodal encoders. In this work, we present FuseLIP, an alternative architecture for multimodal embedding. Leveraging recent progress in discrete image tokenizers, we propose to use a single transformer model which operates on an extended vocabulary of text and image tokens. This early fusion approach allows the different modalities to interact at each depth of encoding and obtain richer representations compared to common late fusion. We collect new datasets for multimodal pre-training and evaluation, designing challenging tasks for multimodal encoder models. We show that FuseLIP outperforms other approaches in multimodal embedding tasks such as VQA and text-guided image transformation retrieval, while being comparable to baselines on unimodal tasks.

Computer use agents (CUA) are systems that automatically interact with graphical user interfaces (GUIs) to complete tasks. CUA have made significant progress with the advent of large vision-language models (VLMs). However, these agents typically rely on cloud-based inference with substantial compute demands, raising critical privacy and scalability concerns, especially when operating on personal devices. In this work, we take a step toward privacy-preserving and resource-efficient agents by developing a lightweight vision-language model that runs entirely on local machines. T o train this compact agent, we introduce an LLM-as-Judge framework that automatically evaluates and filters synthetic interaction trajectories, producing high-quality data for reinforcement learning without human annotation. Experiments on the OS-W orld benchmark demonstrate that our fine-tuned local model outperforms existing baselines, highlighting a promising path toward private, efficient, and generalizable GUI agents.

How well do modern long-context language models understand literary fiction? We explore this question via the task of literary evidence retrieval, repurposing the RELiC dataset of That et al. (2022) to construct a benchmark where the entire text of a primary source (e.g., The Great Gatsby) is provided to an LLM alongside literary criticism with a missing quotation from that work. This setting, in which the model must generate the missing quotation, mirrors the human process of literary analysis by requiring models to perform both global narrative reasoning and close textual examination. We curate a high-quality subset of 292 examples through extensive filtering and human verification. Our experiments show that recent reasoning models, such as Gemini Pro 2.5 can exceed human expert performance (62.5% vs. 50% accuracy). In contrast, the best open-weight model achieves only 29.1% accuracy, highlighting a wide gap in interpretive reasoning between open and closed-weight models. Despite their speed and apparent accuracy, even the strongest models struggle with nuanced literary signals and overgeneration, signaling open challenges for applying LLMs to literary analysis. We release our dataset and evaluation code to encourage future work in this direction.

Understanding user queries is fundamental in many applications, such as home assistants, booking systems, or recommendations. Accordingly, it is crucial to develop accurate Spoken Language Understanding (SLU) approaches to ensure the reliability of the considered system. Current State-of-the-Art SLU techniques rely on large amounts of training data; however, only limited annotated examples are available for specific tasks or languages. In the meantime, instruction-tuned large language models (LLMs) have shown exceptional performance on unseen tasks in a few-shot setting when provided with adequate prompts. In this work, we propose to explore example selection by leveraging Information retrieval (IR) approaches to build an enhanced prompt that is applied to an SLU task. We evaluate the effectiveness of the proposed method on several SLU benchmarks. Experimental results show that lexical IR methods significantly enhance performance without increasing prompt length.

Accurately assessing internal human states is key to understanding preferences, offering personalized services, and identifying challenges in real-world applications. Originating from psychometrics, adaptive testing has become the mainstream method for human measurement and has now been widely applied in education, healthcare, sports, and sociology. It customizes assessments by selecting the fewest test questions . However, current adaptive testing methods face several challenges. The mechanized nature of most algorithms leads to guessing behavior and difficulties with open-ended questions. Additionally, subjective assessments suffer from noisy response data and coarse-grained test outputs, further limiting their effectiveness. To move closer to an ideal adaptive testing process, we propose TestAgent, a large language model (LLM)-powered agent designed to enhance adaptive testing through interactive engagement. This is the first application of LLMs in adaptive testing. TestAgent supports personalized question selection, captures test-takers' responses and anomalies, and provides precise outcomes through dynamic, conversational interactions. Experiments on psychological, educational, and lifestyle assessments show our approach achieves more accurate results with 20% fewer questions than state-of-the-art baselines, and testers preferred it in speed, smoothness, and other dimensions.

The assessment of legal problems requires the consideration of a specific legal system and its levels of abstraction, from constitutional law to statutory law to case law. The extent to which Large Language Models (LLMs) internalize such legal systems is unknown. In this paper, we propose and investigate different approaches to condition LLMs at different levels of abstraction in legal systems. This paper examines different approaches to conditioning LLMs at multiple levels of abstraction in legal systems to detect potentially punishable hate speech. We focus on the task of classifying whether a specific social media posts falls under the criminal offense of incitement to hatred as prescribed by the German Criminal Code. The results show that there is still a significant performance gap between models and legal experts in the legal assessment of hate speech, regardless of the level of abstraction with which the models were conditioned. Our analysis revealed, that models conditioned on abstract legal knowledge lacked deep task understanding, often contradicting themselves and hallucinating answers, while models using concrete legal knowledge performed reasonably well in identifying relevant target groups, but struggled with classifying target conducts.

Privacy policies inform users about data collection and usage, yet their complexity limits accessibility for diverse populations. Existing Privacy Policy Question Answering (QA) systems exhibit performance disparities across English dialects, disadvantaging speakers of non-standard varieties. We propose a novel multi-agent framework inspired by human-centered design principles to mitigate dialectal biases. Our approach integrates a Dialect Agent, which translates queries into Standard American English (SAE) while preserving dialectal intent, and a Privacy Policy Agent, which refines predictions using domain expertise. Unlike prior approaches, our method does not require retraining or dialect-specific fine-tuning, making it broadly applicable across models and domains. Evaluated on PrivacyQA and PolicyQA, our framework improves GPT-4o-mini's zero-shot accuracy from 0.394 to 0.601 on PrivacyQA and from 0.352 to 0.464 on PolicyQA, surpassing or matching few-shot baselines without additional training data. These results highlight the effectiveness of structured agent collaboration in mitigating dialect biases and underscore the importance of designing NLP systems that account for linguistic diversity to ensure equitable access to privacy information.