Jian Lan

Jiaqiao Zhang, Zhoujun Li, Raoyuan Zhao et al.

Large reasoning models (LRMs) achieve strong mathematical reasoning performance in English, but remain much less reliable in many low- and medium-resource languages. This gap is often explained as a failure to understand non-English problem statements. We show that this view is incomplete: even when the problem is given in English, controlling the model's reasoning language can substantially reduce accuracy, suggesting that language also affects reasoning execution itself. To study this effect, we introduce DATG, a Directed Acyclic Trace Graph framework that maps reasoning traces to language-independent mathematical anchors and dependencies. This allows us to align target-language traces with reference DAGs and measure whether they cover required mathematical nodes, respect dependency edges, and avoid harmful mathematical actions. Experiments on the Qwen3 series across 12 languages show that non-English reasoning often suffers from reduced anchor coverage and weaker dependency fidelity, especially in low-resource languages. Motivated by this diagnosis, we propose Loop-Retry and Formula-Retry, two simple test-time controls targeting DATG-exposed failure modes, and show that they consistently improve target-language reasoning performance in low-resource languages.

Jian Lan, Diego Frassinelli, Barbara Plank

Large vision-language models frequently struggle to accurately predict responses provided by multiple human annotators, particularly when those responses exhibit human uncertainty. In this study, we focus on the Visual Question Answering (VQA) task, and we comprehensively evaluate how well the state-of-the-art vision-language models correlate with the distribution of human responses. To do so, we categorize our samples based on their levels (low, medium, high) of human uncertainty in disagreement (HUD) and employ not only accuracy but also three new human-correlated metrics in VQA, to investigate the impact of HUD. To better align models with humans, we also verify the effect of common calibration and human calibration. Our results show that even BEiT3, currently the best model for this task, struggles to capture the multi-label distribution inherent in diverse human responses. Additionally, we observe that the commonly used accuracy-oriented calibration technique adversely affects BEiT3's ability to capture HUD, further widening the gap between model predictions and human distributions. In contrast, we show the benefits of calibrating models towards human distributions for VQA, better aligning model confidence with human uncertainty. Our findings highlight that for VQA, the consistent alignment between human responses and model predictions is understudied and should become the next crucial target of future studies.

Gengyuan Zhang, Tanveer Hannan, Hermine Kleiner et al.

An ideal vision-language agent serves as a bridge between the human users and their surrounding physical world in real-world applications like autonomous driving and embodied agents, and proactively provides accurate and timely responses given user intents. An intriguing challenge arises when agents interact with the world as a dynamic data stream and ad-hoc queries from users: supporting knowledge for queries, namely evidence, usually appears asynchronously with the arrival time of queries, and agents need to ground their responses in historical data, present observations, and even future streams. We frame this challenge as Query-Evidence Asynchrony, where user queries and their supporting evidence typically arrive asynchronously in the streaming setting. This setting requires not only strong reasoning capabilities but also the ability to retain past observations and respond to queries with temporal awareness. In this paper, we introduce a diagnostic benchmark that evaluates Multimodal Large Language Models (MLLMs) on their ability to handle interaction with streaming data. Further, we present AViLA, Asynchronous Video-Language Agent for streaming data interaction that can handle ad-hoc queries and give time-aware responses. For this purpose, AViLA consists of three key modules: comprehensive memory retention, evidence identification, and evidence-grounded trigger, that are designed to maintain a general-purpose memory and respond readily and timely to queries. Our experiments show that existing models often fail to respond at appropriate times, while AViLA significantly improves both accuracy and temporal awareness. Our code and dataset will be publicly available.

Yuhao Zhou, Yunpeng Zhu, Yang Zhou et al.

Vision-Language-Action (VLA) models hold great promise for general-purpose robotic intelligence, yet scaling up such models is severely bottlenecked by the high cost of acquiring annotated training data. Fortunately, vision-equipped robots deployed across various domains already produce abundant vision-action pairs that can be leveraged to scale up VLA training more efficiently. However, these raw data cannot be centrally aggregated due to various constraints and also exhibit severe heterogeneity. To address these challenges, in this paper, we propose ForgeVLA, a federated VLA training framework that learns VLA models from distributed vision-action pairs without centralizing raw data or requiring manual annotations. Specifically, each client in ForgeVLA is equipped with an embodied instruction classifier that maps vision-action pairs to a predefined instruction set, recovering the missing language modality and forming complete vision-language-action triplets. Beyond triplet construction, we also identify vision-language feature collapse as a critical challenge that has been largely overlooked in prior federated VLA research. To mitigate this issue, ForgeVLA combines a client-side contrastive planning loss with a server-side adaptive aggregation strategy to learn task-discriminative representations efficiently. Extensive experiments across multiple benchmarks show that ForgeVLA significantly outperforms other baselines, and ablation studies further validate the contribution of each component.

Jian Lan, Zhicheng Liu, Xinpeng Wang et al.

The ability to derive precise spatial and physical insights is a cornerstone of vision-language models (VLMs), yet their poor performances in related spatial intelligence tasks such as physical reasoning remain a fundamental barrier. The community critically lacks a scientific analysis revealing whether VLMs faithfully reach answers or plausibly make guesses. This work aims to provide a fundamental understanding of how VLMs perceive the physical world, and utilize physical laws, while assessing the reliability of model confidence. We propose NICE and FACT, a dual-diagnostic paradigm that explicitly decomposes quantitative reasoning for kinematic physics: FACT diagnoses visual fidelity, physical law comprehension, and temporal grounding. NICE studies our novel neighborhood-informed calibration method and novel metrics to evaluate and calibrate confidence reliability. Evaluated across 6 latest state-of-the-art VLMs, we uncover that models fail to identify visual preconditions or utilize necessary physical laws to reach answers. This work highlights and establishes a standardized diagnostic paradigm to guide the development of faithful, physically-grounded VLMs.

Udo Schlegel, Franziska Weeber, Jian Lan et al.

Contrastive Language-Image Pre-training (CLIP) is a widely used multimodal model that aligns text and image representations through large-scale training. While it performs strongly on zero-shot and few-shot tasks, its robustness to linguistic variation, particularly paraphrasing, remains underexplored. Paraphrase robustness is essential for reliable deployment, especially in socially sensitive contexts where inconsistent representations can amplify demographic biases. In this paper, we introduce the Paraphrase Ranking Stability Metric (PRSM), a novel measure for quantifying CLIP's sensitivity to paraphrased queries. Using the Social Counterfactuals dataset, a benchmark designed to reveal social and demographic biases, we empirically assess CLIP's stability under paraphrastic variation, examine the interaction between paraphrase robustness and gender, and discuss implications for fairness and equitable deployment of multimodal systems. Our analysis reveals that robustness varies across paraphrasing strategies, with subtle yet consistent differences observed between male- and female-associated queries.

Haokun Chen, Yueqi Zhang, Yuan Bi et al. · deepmind, oxford

In recent years, Large Language Models (LLMs) have achieved remarkable advancements, drawing significant attention from the research community. Their capabilities are largely attributed to large-scale architectures, which require extensive training on massive datasets. However, such datasets often contain sensitive or copyrighted content sourced from the public internet, raising concerns about data privacy and ownership. Regulatory frameworks, such as the General Data Protection Regulation (GDPR), grant individuals the right to request the removal of such sensitive information. This has motivated the development of machine unlearning algorithms that aim to remove specific knowledge from models without the need for costly retraining. Despite these advancements, evaluating the efficacy of unlearning algorithms remains a challenge due to the inherent complexity and generative nature of LLMs. In this work, we introduce a comprehensive auditing framework for unlearning evaluation, comprising three benchmark datasets, six unlearning algorithms, and five prompt-based auditing methods. By using various auditing algorithms, we evaluate the effectiveness and robustness of different unlearning strategies. To explore alternatives beyond prompt-based auditing, we propose a novel technique that leverages intermediate activation perturbations, addressing the limitations of auditing methods that rely solely on model inputs and outputs.

Jian Lan, Zhicheng Liu, Udo Schlegel et al.

Large vision-language models (VLMs) achieve strong performance in Visual Question Answering but still rely heavily on supervised fine-tuning (SFT) with massive labeled datasets, which is costly due to human annotations. Crucially, real-world datasets often exhibit human uncertainty (HU) -- variation in human confidence across annotations -- but standard SFT simply optimizes toward the most frequent label, disregarding HU distributions. This leaves two open questions: How does HU affect SFT, and how can HU be effectively leveraged in training? In this work, we first conduct a systematic evaluation of VLMs across varying HU levels. We have two key findings: (i) surprisingly, high-HU samples contribute little or even degrade model performance, and (ii) naively training on the full dataset yields under-calibrated models that fail to capture HU distributions. Motivated by these findings, we introduce HaDola, a human uncertainty-aware data selection and automatic labeling framework. HaDola operates in four stages -- discriminate, self-annotate, error trigger, and training -- to iteratively identify harmful samples, prioritize informative ones, and bootstrap from a small seed set (5\% of data). Our approach substantially reduces reliance on costly HU annotations and makes VLMs more accurate and better calibrated. Extensive experiments on VQAv2 and VizWiz datasets demonstrate that HaDola consistently matches or outperforms state-of-the-art baselines with less training data. Our work highlights the importance of explicitly modeling HU in SFT, suggesting that better utilization of HU is more effective than merely scaling up dataset size.

Han Yang, Jian Lan, Yihong Liu et al.

Autoregressive language models are vulnerable to orthographic attacks, where input text is perturbed with characters from multilingual alphabets, leading to substantial performance degradation. This vulnerability primarily stems from the out-of-vocabulary issue inherent in subword tokenizers and their embeddings. To address this limitation, we propose a pixel-based generative language model that replaces the text-based embeddings with pixel-based representations by rendering words as individual images. This design provides stronger robustness to noisy inputs, while an extension of compatibility to multilingual text across diverse writing systems. We evaluate the proposed method on the multilingual LAMBADA dataset, WMT24 dataset and the SST-2 benchmark, demonstrating both its resilience to orthographic noise and its effectiveness in multilingual settings.