Tool-using agents built on large language models (LLMs) are usually scored by comparing the final answer with a reference text, using a string heuristic or an LLM judge. This is unreliable when the references describe the task rather than its answer. We replace it with a deterministic validator that checks the tools called, the executed parameters, the resulting state, and whether the values in the answer are supported by the correct tool observations. We re-scored 33,583 traces from 13 agents on simulated calendar, weather, medical and retail tools. The simulator reproduced all 10,396 logged observations, and the gold observations supported the validator in 16 of 18 disagreements with a three-annotator majority (the annotators in none). The instrument changes conclusions: on the same 306 traces, Gemini-2.0-Flash scores 63.4% under the substring heuristic and 5.6% under the validator. In a paired design (same task and model, with and without one schema-valid wrong argument injected into the first tool call), the error rate rises from 52.4% to 74.5% (+22.0 percentage points, 95% CI 18.4-25.5; 1,085 pairs, 7 models), and by 44.7 points (38.4-51.0) where the corrupted call executed without errors. An LLM-judge ensemble estimates +2.5 points on the same pairs. Agents also report tool results without executing any call, in 0-42.5% of traces depending on the model (rule-based lower bounds), largely because the harness accepts an action and a final answer in one completion. A rule-based runtime monitor costs about 7 microseconds per trace but reduces wrong answers mainly by abstaining.
Defensive training methods such as positive preventative steering (PPS) and inoculation prompting (IP) offer surprising results through seemingly similar processes: both add trait-inducing objects to large language models (LLMs) during training, and both defend the LLM against acquiring the trait. The success of these methods raises the questions: how do they work, and are they doing the same thing? We provide behavioral and mechanistic analyses of PPS and IP, using the acquisition of an ``evil'' persona by Qwen2.5-7B-Instruct as an in-depth case study and replicating key measurements across six traits and four models. Behaviorally, neither method operates through a purely associative mechanism, and both methods prevent trait acquisition; but PPS can also substantially reduce pre-existing trait expression, whereas IP does so only modestly. Mechanistically, PPS induces a concentrated change in the activation gradient and shifts the signal along the trait direction from amplifying to attenuating, whereas IP changes the gradient more diffusely and shifts it close to zero along the trait direction. Gradient manipulations show that the direction of this component is causally relevant to trait acquisition, although it is one effect rather than a complete account of either defense. We then show that IP lowers the next-token prediction loss on trait-expressing data, whereas PPS can raise it at the intensities at which it defends; we show how this analysis raises the possibility that the loss changes and the gradient shift are two summaries of the same training signal rather than completely separate mechanisms. Together, our results show that PPS and IP are not interchangeable, and they highlight open questions about a complete mechanistic picture.
In data-sensitive domains such as finance, cross-silo federated learning (CFL) allows organizations to collaboratively train AI models without sharing raw data. However, many practical CFL deployments are coopetitive, where organizations cooperate during model training while competing in downstream markets. In such settings, training contributions can improve the global model yet inadvertently strengthen rivals, eroding one's competitive advantage. This dilemma is amplified by non-IID data, which leads to asymmetric learning gains and undermines sustained participation. While existing work reward organizations based on marginal training contributions, they fail to account for the implicit costs of strengthening competitors. In this paper, we introduce CoCoGen+, a coopetition-compatible data generation and incentivization framework that jointly models data heterogeneity and inter-organizational competition while endogenizing GenAI-based synthetic data generation as a strategic decision. Specifically, CoCoGen+ formulates each training round as a weighted potential game, where organizations strategically decide how much data to generate by balancing learning gains against computational costs and utility losses caused by competition. We establish the existence and uniqueness of the continuous Nash equilibrium under the adopted model and develop a provably convergent scalar bisection algorithm. We further analyze how bilateral payoff redistribution modifies residual competition and data-generation incentives, and empirically evaluate the social welfare and individual utilities of the resulting allocations. Experiments on FMNIST, CIFAR-10, and CIFAR-100 validate the feasibility of CoCoGen+. The results show how heterogeneity, competition intensity, and incentives shape organizational strategies and social welfare, while CoCoGen+ achieves higher social welfare than competition-aware baselines.
New multi-modal large language models (MLLMs) are continuously being trained and deployed, following rapid development cycles. This generative AI frenzy is driving steady increases in energy consumption, greenhouse gas emissions, and a plethora of other environmental impacts linked to datacenter construction and hardware manufacturing. Mitigating the environmental consequences of GenAI remains challenging due to an overall lack of transparency by the main actors in the field. Even when the environmental impacts of specific models are mentioned, they are typically restricted to the carbon footprint of the final training run, omitting the research and development stages. In this work, we explore the impact of GenAI research through a fine-grained analysis of the compute spent to create Moshi, a 7B-parameter speech-text foundation model for real-time dialogue developed by Kyutai, a leading privately funded open science AI lab. For the first time, our study dives into the anatomy of compute-intensive MLLM research, quantifying the GPU-time invested in specific model components and training phases, as well as early experimental stages, failed training runs, debugging, and ablation studies. Additionally, we assess the environmental impacts of creating Moshi from beginning to end using a life cycle assessment methodology: we quantify energy and water consumption, greenhouse gas emissions, and mineral resource depletion associated with the production and use of datacenter hardware. Although Moshi is a specific case study, its development involves research stages common to most GenAI R&D pipelines. Our analysis therefore supports general practical recommendations for reducing compute usage and environmental impacts, while paving the way toward more sustainable MLLM research.
AI is increasingly deployed in multi-agent systems; however, most research considers only the behavior of individual models. We experimentally show that multi-agent "AI organizations" are simultaneously more effective at achieving business goals, but less aligned, than individual AI agents. We examine 12 tasks across two practical settings: an AI consultancy providing solutions to business problems and an AI software team developing software products. Across all settings, AI Organizations composed of aligned models produce solutions with higher utility but greater misalignment compared to a single aligned model. Our work demonstrates the importance of considering interacting systems of AI agents when doing both capabilities and safety research.
Automatic sleep staging is a multimodal learning problem involving heterogeneous physiological signals such as EEG and EOG, which often suffer from domain shifts across institutions, devices, and populations. In practice, these data are also affected by noisy annotations, yet label-noise-robust multi-source domain generalization remains underexplored. We present the first benchmark for Noisy Labels in Multi-Source Domain-Generalized Sleep Staging (NL-DGSS) and show that existing noisy-label learning methods degrade substantially when domain shifts and label noise coexist. To address this challenge, we propose FF-TRUST, a domain-invariant multimodal sleep staging framework with Joint Time-Frequency Early Learning Regularization (JTF-ELR). By jointly exploiting temporal and spectral consistency together with confidence-diversity regularization, FF-TRUST improves robustness under noisy supervision. Experiments on five public datasets demonstrate consistent state-of-the-art performance under diverse symmetric and asymmetric noise settings.
Data collected from the physical world is always a combination of multiple sources: an underlying signal from the physical process of interest and a signal from measurement-dependent artifacts from the sensor or instrument. This secondary signal acts as a confounding factor, limiting our ability to extract information about the physics underlying the phenomena we observe. Furthermore, it complicates the combination of observations in heterogeneous or multi-instrument settings. We propose a deep learning framework that leverages overlapping observations, a dual-encoder architecture, and a counterfactual generation objective to disentangle these factors of variation. The resulting representations explicitly separate intrinsic signals from sensor-specific distortions and noise, and can be used for counterfactual view generation, parameter inference unconfounded by measurement distortions, and instrument-independent similarity search. We demonstrate the effectiveness of our approach on astrophysical galaxy images from the DESI Legacy Imaging Survey (Legacy) and the Hyper Suprime-Cam (HSC) Survey as a representative multi-instrument setting. This framework provides a general recipe for scientific and multi-modal self-supervised pretraining: construct training pairs from overlapping observations of the same physical system, treat sensor- or modality-specific effects as augmentations, and learn invariant representations through counterfactual generation.
Multimodal Large Language Models (MLLMs) have achieved strong performance on RGB image understanding, yet their ability to use spectral evidence beyond the visible range remains largely unexplored. Hyperspectral imagery (HSI) provides dense spectral measurements that reveal material and environmental cues unavailable in RGB, but current MLLMs cannot directly ingest high-dimensional HSI. To study this gap, we introduce HM-Bench, an evidence-grounded benchmark for hyperspectral image understanding with MLLMs. HM-Bench contains 19,337 question--answer pairs from 2,178 hyperspectral samples across 13 task categories, covering general perception, spectral reasoning, and spatial--spectral reasoning. To make HSI accessible to the native image--text interfaces of existing MLLMs, we further propose VSR$^{2}$, a training-free Visual--Spectral--Report Reasoning framework. VSR$^{2}$ represents each HSI sample with three aligned views: an RGB image for visual semantics and spatial context, a PCA-based image for dominant spectral variation, and a structured report for quantitative spectral--spatial evidence. Under a controlled RGB-only versus VSR$^{2}$ evaluation protocol, experiments on 17 representative MLLMs show that HSI information improves average accuracy from 38.90% to 40.95%, with gains varying substantially across models and task categories. These results indicate that HSI information (beyond RGB) is useful for current MLLMs, while robust hyperspectral reasoning remains an open challenge.
Fundus imaging such as CFP, OCT and SLO is crucial for the early detection of retinal anomalies and diseases. Fundus image understanding, due to its knowledge-intensive nature, poses a challenging vision-language task. An emerging approach to addressing the task is to post-train a generic multimodal large language model (MLLM), either by supervised finetuning (SFT) or by reinforcement learning with verifiable rewards (RLVR), on a considerable amount of in-house samples paired with high-quality clinical reports. However, these valuable samples are not publicly accessible, which not only hinders reproducibility but also practically limits research to few players. In this paper, we make a novel attempt to train a reasoning-enhanced fundus-reading MLLM, termed Fundus-R1, using exclusively public datasets, wherein over 94% of the data are annotated with image-level labels only. Our technical contributions are two-fold. First, we propose a RAG-based method for composing image-specific, knowledge-aware reasoning traces. Such auto-generated traces link visual findings identified by a generic MLLM to the image labels in terms of ophthalmic knowledge. Second, we enhance RLVR with a process reward that encourages self-consistency of the generated reasoning trace in each rollout. Extensive experiments on three fundus-reading benchmarks, i.e., FunBench, Omni-Fundus and GMAI-Fundus, show that Fundus-R1 clearly outperforms multiple baselines, including its generic counterpart (Qwen2.5-VL) and a stronger edition post-trained without using the generated traces. This work paves the way for training powerful fundus-reading MLLMs with publicly available data.
Interpreting the concepts encoded by individual neurons in deep neural networks is a crucial step towards understanding their complex decision-making processes and ensuring AI safety. Despite recent progress in neuron labeling, existing methods often limit the search space to predefined concept vocabularies or produce overly specific descriptions that fail to capture higher-order, global concepts. We introduce LINE, a novel, training-free iterative approach tailored for open-vocabulary concept labeling in vision models. Operating in a strictly gray-box setting, LINE leverages a large language model and a text-to-image generator to iteratively propose and refine concepts in a closed loop, guided by activation history. We demonstrate that LINE achieves state-of-the-art performance across multiple model architectures, yielding AUC improvements of up to 0.15 on ImageNet and 0.05 on Places365, while discovering, on average, 25% of new concepts missed by predefined vocabularies. Beyond identifying the top concept, LINE provides a complete generation history, enabling polysemanticity evaluation and producing supporting visual explanations that rival gradient-dependent activation maximization methods. The LINE source code is available at https://github.com/CredibleAI/LINE
There has been a recent surge in interest about the question of how we can ``delete'' specific data points or behaviours from a trained model, a goal referred to as ``machine unlearning''. We argue that the umbrella term ``unlearning'' actually spans two distinct problem formulations, but the distinction between them has not yet been observed in literature. This causes ambiguity around when an unlearning algorithm is expected to work, leads to the use of inappropriate metrics and baselines when comparing algorithms to one another, difficulty in interpreting results, and missed opportunities for pursuing critical research directions. In this paper, we argue for the position that a fundamental distinction must be made between two notions that we refer to as unlearning and untraining. On one hand, untraining aims to reverse the effect of having trained on a given ``forget set'', i.e. to remove the influence that that specific set of examples had on the model during training. On the other hand, unlearning aims not just to remove the influence of those given examples, but also of the entire underlying subdistribution from which they were sampled (and thus e.g. the concept or model behaviour that those examples represent). We discuss technical definitions of these problems and map problem settings studied in the literature to each notion. By disambiguating technical definitions, our work aims to accelerate progress in this important field.
Dynamic urban environments are often captured by cameras placed at spatially separated locations with little or no view overlap. However, most existing 4D reconstruction methods assume densely overlapping views and struggle under sparse multi-location observations, producing unstable reconstructions in unobserved intermediate regions. To address this practical yet underexplored setting, we propose Stitch4D, a unified 4D reconstruction framework that compensates for missing spatial coverage in sparsely observed urban environments. Stitch4D synthesizes intermediate \emph{bridge views} between distant camera locations and jointly optimizes real and synthesized observations in a unified coordinate frame with inter-location consistency constraints. By recovering intermediate spatial coverage before optimization, Stitch4D mitigates geometric collapse and improves reconstruction stability in sparse regions. To evaluate this setting, we introduce Urban Sparse 4D (U-S4D), a controlled CARLA-based benchmark for free-viewpoint reconstruction under sparse multi-location configurations. Experiments on U-S4D show that Stitch4D consistently outperforms representative 4D reconstruction baselines in image-quality metrics. These results suggest that recovering intermediate spatial coverage is an effective strategy for stabilizing 4D reconstruction in sparse urban environments. The project page is provided in https://stitch4d-project-page.vercel.app/.
Self-attention in Transformers generates dynamic operands that force conventional Compute-in-Memory (CIM) accelerators into costly non-volatile memory (NVM) reprogramming cycles, degrading throughput and stressing device endurance. Existing solutions either reduce but retain NVM writes through matrix decomposition or sparsity, or move attention computation to digital CMOS at the expense of NVM density. We present TrilinearCIM, a Double-Gate FeFET (DG-FeFET)-based architecture that uses back-gate modulation to realize a three-operand multiply-accumulate primitive for in-memory attention computation without dynamic ferroelectric reprogramming. Evaluated on BERT-base (GLUE) and ViT-base (ImageNet and CIFAR), TrilinearCIM outperforms conventional CIM on seven of nine GLUE tasks while achieving up to 46.6\% energy reduction and 20.4\% latency improvement over conventional FeFET CIM at 37.3\% area overhead. To our knowledge, this is the first architecture to perform complete Transformer attention computation exclusively in NVM cores without runtime reprogramming.
Cross-city transfer uses labeled data from well-instrumented cities to improve prediction in label-scarce ones. The task is difficult when cities have incompatible region partitions and no ground-truth correspondence. Existing methods either rely on heuristic anchors, which can create hubness, or align global distributions, which can over-mix heterogeneous regions. We propose SCOT (Soft Correspondence via Optimal Transport), an optimal transport-based framework for learning explicit, mass-controlled soft correspondences between unequal region sets. SCOT combines an optimal transport-weighted contrastive loss, a one-sided cycle regularizer, and a shared prototype hub for multi-source transfer. Across real-world cities and tasks, SCOT consistently outperforms strong baselines, reducing MAE and MAPE by 5-50%. Additional analyses show that the learned correspondences are selective, region-specific, and robust to heterogeneous settings.
Existing hallucination detection methods for large language models (LLMs) rely on external verification at inference time, requiring gold answers, retrieval systems, or auxiliary judge models. We ask whether this external supervision can instead be distilled into the model's own representations during training, enabling hallucination detection from internal activations alone at inference time. We introduce a weak supervision framework that combines three complementary grounding signals: substring matching, sentence embedding similarity, and an LLM as a judge verdict to label generated responses as grounded or hallucinated without human annotation. Using this framework, we construct a 15000-sample dataset from SQuAD v2 (10500 train/development samples and a separate 5000-sample test set), where each example pairs a LLaMA-2-7B generated answer with its full per-layer hidden states and structured hallucination labels. We then train five probing classifiers: ProbeMLP (M0), LayerWiseMLP (M1), CrossLayerTransformer (M2), HierarchicalTransformer (M3), and CrossLayerAttentionTransformerV2 (M4), directly on these hidden states, treating external grounding signals as training-time supervision only. Our central hypothesis is that hallucination detection signals can be distilled into transformer representations, enabling internal detection without any external verification at inference time. Results support this hypothesis. Transformer-based probes achieve the strongest discrimination, with M2 performing best on 5-fold average AUC/F1, and M3 performing best on both single-fold validation and held-out test evaluation. We also benchmark inference efficiency: probe latency ranges from 0.15 to 5.62 ms (batched) and 1.55 to 6.66 ms (single sample), while end-to-end generation plus probe throughput remains approximately 0.231 queries per second, indicating negligible practical overhead.
Large vision-language models can produce object hallucinations in image descriptions, highlighting the need for effective detection and mitigation strategies. Prior work commonly relies on the model's attention weights on visual tokens as a detection signal. We reveal that coarse-grained attention-based analysis is unreliable due to hidden confounders, specifically token position and object repetition in a description. This leads to Simpson's paradox: the attention trends reverse or disappear when statistics are aggregated. Based on this observation, we introduce HaloProbe, a Bayesian framework that factorizes external description statistics and internal decoding signals to estimate token-level hallucination probabilities. HaloProbe uses balanced training to encourage learning from internal evidence and combines it with a learned prior over external features to recover the true posterior. While intervention-based mitigation methods often degrade utility or fluency by modifying models' internals, we use HaloProbe as an external scoring signal for non-invasive mitigation. Our experiments show that HaloProbe-guided decoding reduces hallucinations more effectively than state-of-the-art intervention-based methods while preserving caption utility and linguistic fluency.
Legal information retrieval in Portuguese remains difficult to evaluate systematically because available datasets differ widely in document type, query style, and relevance definition. We present JU\'A, a public benchmark for Brazilian legal retrieval designed to support more reproducible and comparable evaluation across heterogeneous legal collections. More broadly, JU\'A is intended not only as a benchmark, but as a continuous evaluation infrastructure for Brazilian legal IR, combining shared protocols, common ranking metrics, fixed splits when applicable, and a public leaderboard. The benchmark covers jurisprudence retrieval as well as broader legislative, regulatory, and question-driven legal search. We evaluate lexical, dense, and BM25-based reranking pipelines, including a domain-adapted Qwen embedding model fine-tuned on JU\'A-aligned supervision. Results show that the benchmark is sufficiently heterogeneous to distinguish retrieval paradigms and reveal substantial cross-dataset trade-offs. Domain adaptation yields its clearest gains on the supervision-aligned JU\'A-Juris subset, while BM25 remains highly competitive on other collections, especially in settings with strong lexical and institutional phrasing cues. Overall, JU\'A provides a practical evaluation framework for studying legal retrieval across multiple Brazilian legal domains under a common benchmark design.
We study the problem of trajectory optimization in settings where the system dynamics are unknown and the controller cannot simulate candidate trajectories through a model at test time. When an offline dataset of trajectories is available, an agent could directly learn a trajectory generator by distribution matching. However, this approach only recovers the behavior distribution in the dataset, and does not in general produce a model that minimizes a desired cost criterion. In this work, we propose Drifting Receding Horizon Control (RHC), an offline trajectory optimization framework that combines drifting generative models with receding-horizon planning under unknown dynamics. The goal of Drifting RHC is to learn, from an offline dataset of trajectories, a conditional distribution over trajectories that is both supported by the data and biased toward optimal plans. We show that the target distribution of Drifting RHC is the unique solution of an objective that trades off optimality with closeness to the offline prior. We further bound the gap between the predicted and the executed cost of a generated plan by its consistency with the dynamics, and introduce a rollout-consistency loss that distills a learned model into the generator at training time only. On a mass-spring-damper, a cart-pole and a planar quadrotor, Drifting RHC attains the lowest regret among generative and model-based planners, outperforms DeePC on the nonlinear systems, and plans $9$ to $83$ times faster than iLQR, model-based and diffusion-based planners.
The Sinkhorn--Knopp (SK) algorithm is a cornerstone method for matrix scaling and entropically regularized optimal transport (EOT). Despite its empirical efficiency, existing bounds for achieving marginal accuracy $\varepsilon$ deteriorate severely in the presence of extreme entries, bottlenecked either by the global maximum regularized cost $\eta\|C\|_\infty$ (where $\eta$ is the regularization parameter and $C$ the cost matrix) or the matrix's minimum-to-maximum entry ratio $\nu$. These bounds may be overly sensitive to rare extreme entries that do not dominate the overall matrix structure. We identify a broad structural regime in which the dependence on the magnitude of global extreme entries can be circumvented. For EOT, we introduce well-boundedness, a local mass property that captures the row- and column-wise bulk structure of the underlying matrix while maintaining robustness against scattered entrywise outliers. We prove that, governed by this structural condition, SK converges to the target transport plan for a problem of dimension $n$ in $O(\log n + \log(1/\varepsilon))$ iterations. This complexity bound is independent of the maximum regularized cost $\eta\|C\|_\infty$. A computationally negligible pre-scaling step eliminates this dimensional dependence, reducing the iteration complexity to a strictly dimension-free $O(\log(1/\varepsilon))$. Beyond EOT, we establish a sharp phase transition for general $(\boldsymbol{u},\boldsymbol{v})$-scaling governed by a critical matrix density threshold. We prove that when a matrix's density exceeds this threshold, the iteration complexity is strictly independent of $\nu$. Conversely, when the density falls below this threshold, the dependence on $\nu$ becomes unavoidable; in this sub-critical regime, we construct instances where SK requires $\Omega(n/\varepsilon)$ iterations.
Large Vision-Language Models (LVLMs) have achieved impressive progress in multimodal reasoning, yet they remain prone to object hallucinations, generating descriptions of objects that are not present in the input image. In this work, we investigate hallucination from the perspective of attention-value dynamics inside LVLM vision encoders. We identify a consistent three-phase structure of visual processing---diffusion, focus, and rediffusion---and show that the focus phase is where attention most clearly separates strongly and weakly supported visual tokens. However, low attention does not necessarily imply negligible downstream influence: low-attention tokens in the focus phase can still exert non-negligible value-side influence on the attention output relative to their small attention mass. Through controlled phase-wise value interventions, we find that hallucination behavior is particularly sensitive to the value content of low-attention tokens during this phase. Replacing or neutralizing these values reduces hallucination metrics while largely preserving grounded object evidence. A token-level teacher-forcing analysis further shows that the intervention reduces the probability of hallucinated object tokens with much smaller effects on ground-truth object tokens. In addition, Visual Attention Ratio (VAR) analysis shows that focus-phase intervention is accompanied by increased attention to visual tokens during decoding. Based on these observations, we instantiate a simple training-free inference-time intervention that replaces focus-phase low-attention values with an image-level mean value vector using statistics from a single forward pass. Experiments across multiple LVLM backbones demonstrate that this analysis-derived intervention reduces object hallucination with negligible additional runtime and remains compatible with existing inference-time mitigation methods.