World action models jointly predict video and actions conditioned on language. Existing studies focus primarily on architectures and training mechanisms, typically treating language-video alignment acquired through video pretraining as an established foundation. However, introducing action learning may impair this alignment. Principled guidance on data and supervision design for maintaining language-video alignment during action learning remains limited. To investigate this, we curate a pool of approximately 47,000 hours of general videos, egocentric videos, UMI demonstrations, and robot demonstrations. Building on this, we systematically study how language-video supervision, language-video-action supervision, and the joint use of these two forms affect WAM performance. In particular, we examine how language annotation quality affects pretrained language-video capabilities, and how action supervision from UMI and robot demonstrations extends these capabilities to language-video-action modeling. To better scale the model, we further conduct detailed studies on joint training recipe and model architecture. Under a fixed training budget, mixing egocentric videos and UMI demonstrations improves control generalization. Model architecture affects how effectively these experiences are used, while scaling pretraining under the selected configuration brings further gains. The resulting recipe, post-trained on RoboTwin Clean demonstrations only, achieves success rates of 88.10% and 61.56% under Easy (Clean-to-Clean) and Hard (Clean-to-Random) conditions, respectively. Our model achieves 40.00% SR across three real-world tasks, compared with 23.33% for the strongest baseline, indicating better generalization under unseen conditions. We release the curated data, training code, and model weights.
The Platonic Representation Hypothesis (PRH) suggests that neural networks are converging toward similar representation spaces. As language models become more capable, their representations are aligning with those of vision models. However, recent work has questioned whether this trend continues to hold for newer language models. Moreover, it has been shown that representation similarity metrics can be confounded by model depth and width, requiring appropriate baselines for interpretation. In this work, we first investigate where alignment emerges between unimodal models across model layers and representation dimensions. Next, we find that the PRH capability-alignment trend persists for newer models as long as the captions are sufficiently clean and informative. Finally, we contextualize cross-modal alignment, introducing unimodal and multimodal baselines on various datasets. Our results show that the cross-modal representational alignment of unimodal models can surpass that of explicitly aligned vision-language models, but remains substantially behind the alignment observed for models of the same modality. Code is available at https://github.com/ExplainableML/prh-closer-look.
Reinforcement learning (RL) can improve an agent's task capability while reinforcing behavior that violates contextual execution requirements. We study this reward-driven misalignment, including reward hacking, through verifiable reward--context conflicts, where compliant task completion remains feasible while some executable shortcuts earn more reward than every compliant strategy. Encoding these conflicts in task dynamics lets us generate diverse interactive environments at scale and derive compliance feedback from execution histories. In resource and reporting workflows, task optimization increases real productivity through unauthorized resource use or raises reported scores while actual work declines. Controlled changes in the training distribution show that minority conflict exposure can still lead to widespread shortcut acquisition, reaching 97.9\% certified profitable violations on an API task under a 25\% conflict mixture after 128 updates. The same contextual checks supply a simple binary penalty that supports joint capability and compliance gains and partially mitigates acquired violations. Learned safety adapts to expanded authorization and transfers across mechanisms into workflows with evolving evidence and approval requirements. On 32 unseen changing-evidence tasks, resource-trained safety increases safe completion from 53.1\% to 89.1\% relative to task-only training, while both policies achieve full real work. These results connect a verifiable account of reward-driven misalignment with scalable task construction and safety learning that remains useful under changing execution requirements.
Persona generalization occurs when fine-tuning a language model in one domain induces behavioral changes in others -- for example, the broadly unsafe behavior observed in emergent misalignment. We discover a surprising "on-the-line" phenomenon in persona generalization: in-domain and out-of-domain persona rates across various fine-tuning hyperparameters follow an approximately probit-linear relationship, and the specific probit-linear relationship is surprisingly shared even across model families and model sizes, including Qwen and Llama. Training choices that leave the set of trainable parameters unchanged, such as learning rate, weight decay, and training duration, largely move models along the same line, reducing transfer mainly by weakening in-domain learning. Changing which layers are fine-tuned, however, can change the slope. For unsafe behavior, fine-tuning only later layers yields shallower slopes, retaining strong in-domain learning while reducing out-of-domain misalignment. This "on-the-line" phenomenon unifies unexpected, isolated observations of persona generalization within a quantitative framework for predicting and controlling behavioral transfer.
Large language model agents are increasingly deployed as multi-agent systems rather than as a standalone model operating within a single context. Recent agent harnesses allow the model to instantiate worker agents and to author the executable workflow that coordinates them, shifting the design of system structure from developers to the model at run time. The quality of this organization therefore constitutes a distinct capability, yet existing benchmarks either measure end-to-end task completion or evaluate coordination under structures specified in advance. We formalize this capability as workflow orchestration and introduce WorkflowArena, a benchmark of 145 long-horizon tasks spanning 8 professional domains, from software engineering and formal mathematics to legal analysis and 3D production. Each task specifies an initial workspace and an integrated deliverable composed of interdependent components, which is assessed by a hidden verifier. Agents determine autonomously whether and how to delegate, and all delegation is recorded, enabling joint evaluation of deliverable quality and the organization of the underlying work. Experiments with nine models show that no single configuration performs best across all domains and that systems exhibit distinct delegation strategies: at high effort, Claude models in Claude Code invoke workflows less frequently but with larger worker groups, whereas GPT models in Codex delegate more frequently in smaller units. When the executing model is held fixed on a challenging subset of tasks, workflows authored by Claude Opus 5.5 substantially improve weaker executors, indicating that workflow organization is itself a measurable determinant of task performance.
Mature, germinal centre-bearing tertiary lymphoid structures (TLS) and mature lymph node follicles (secondary lymphoid structures, SLS) sustain the plasma cell response that shapes outcome and immunotherapy benefit in lung cancer. How this loco-regional axis differs between neoadjuvant-treated and therapy-naive tumors is poorly defined. We analyzed H&E whole-slide images of resected tumors and paired proximal and distal lymph nodes from matched neoadjuvant-treated (neoLUSC, n=20) and therapy-naive (naivLUSC, n=20) lung squamous cell carcinoma (LUSC) patients. Aignostics Atlas H&E Tissue Analysis (HETA) models quantified cell densities per tissue compartment and cell-cell co-localization; a dedicated deep-learning model validated against pathologist annotations detected TLS and SLS and graded their maturity. Prognostic value was tested in an independent cohort (n=458; 46 treated). In neoLUSC tumors, stromal plasma cell density was 72% (p=0.004) and stromal lymphocyte density 48% (p=0.015) lower than in naivLUSC; macrophages and granulocytes did not differ significantly. Total and immature TLS did not differ, whereas mature TLS in the tumor (p=0.034) and mature SLS in the tumor-distal lymph node (p=0.003) were lower in neoLUSC patients. Peri-TLS plasma cell density correlated with tumor lymphocyte density in naivLUSC (rho=0.54, p=0.017) but not in neoLUSC tumors (rho=0.07). In the validation cohort, stromal plasma cell density (HR 0.69, p=0.011) and lymphocyte-plasma cell co-localization (HR 0.66, p=0.004) were associated with longer survival, and neoadjuvant status with shorter survival (HR 1.50, p=0.026). Neoadjuvant therapy is thus associated with fewer mature lymphoid structures and weaker coupling of the humoral compartment to the tumor rather than uniformly lower immune readouts. The stromal plasma cell/co-localization axis is a candidate prognostic biomarker.
Latent Diffusion Models dominate generative modeling but face two-stage training complexity and the inherent reconstruction limits of pre-trained autoencoders. Pixel-space Flow Matching offers a simplified end-to-end alternative yet often encounters training difficulties when using the standard $v$-prediction objective. Recent research attributes this difficulty to the off-manifold nature of noised targets, motivating a shift toward $x$-prediction. In this paper, we investigate this explanation through patch-wise Principal Component Analysis (PCA) and identify a bandwidth bottleneck. The velocity in minor components is largely determined by the raw noisy input, requiring deep backbones to preserve information that competes with principal-component modeling. Through controlled experiments and representation analysis, we further examine why $x$-prediction remains effective: its parameterization provides a raw-input bypass, while its learned input representation can spontaneously concentrate on principal directions. Under low-rank observations, our analysis of the induced flow characterizes $x$-prediction as an implicit conditional-mean decoder, reconstructing discarded components from the retained information. These observations motivate the Raw-injected Transformer (RiT), which combines a lightweight raw-conditioned decoder with an output-side shortcut to separate semantic modeling from raw-state propagation. Our method restores effective $v$-prediction and achieves competitive performance on ImageNet $256\times256$, providing a perspective on the interaction between prediction objectives and architectural bandwidth.
As large language models become more capable agents, they can exhibit deceptive behaviors, cheating human for higher rewards while appearing helpful. We attribute this failure to evaluation-side partial observability, where models are evaluated mainly from final outcome, without checking the faithfulness of underlying reasoning, tool use, and environmental effects. Prior mitigations address this partial observability by predefined evaluator latent belief space or one-shot monitoring, which do not extend to free-form language space and vulnerable to stronger deceptions that evade simple monitoring. In this work, we formulate deception mitigation as an Evaluation-Side Partially Observable Extensive Form Game (e-PoG), a max-min framework in which an agent maximizes reward, a monitor minimizes the agent's payoff by exposing deception from the full interaction trajectory, and a fixed verifier validates the monitor's claims and converts them into penalties. We define the solution concept as a Subgame-Perfect Oversight Equilibrium (SPOE) between agent and monitor, and prove its existence in finite games. With a perfect verifier, every SPOE is faithful; with an imperfect verifier, the guarantee persists if verifier errors do not reverse the ranking of monitor actions. To achieve faithful SPOE, we derive a two-timescale policy gradient algorithm and enhance the pipeline empirically with NLI verification and deterministic checks. Experiments with Gemma3-4B-IT and Qwen3-4B/8B on Agentic Role-Playing, SearchQA, and WebShop show that e-PoG mitigates multiple emergent deception modes, outperforming the strongest baselines by 26.0% in deception mitigation and by up to 39.8% in task success. These gains persist under weak-to-strong supervision. Code is available at https://github.com/Toyoid/deceptive-agent.
AI models' growing capacity for complex reasoning has made them attractive tools for forecasting, or the systematic prediction of the future based on past and present data. However, should AI forecasting systems be adopted, human builders will shoulder the work of building and evaluating them. In this paper, we explore what this work looks like with a qualitative interview study with 15 builders of forecasting bots, sampled from among successful participants in a recent AI forecasting competition. Our analysis surfaces the challenges around (1) tuning a bot to acquire factual research, (2) implementing coherent reasoning, and (3) understanding how well one's bot actually does. During their work, builders frequently introduced complexity that did not pay off, and tested in ways that were opportunistic and piecemeal. Our detailed account of their challenges helps clarify what problems should be targeted in upcoming innovations in process, tools, and human-AI teaming for forecasting.
Low-precision pretraining is usually judged by validation loss. We ask whether matching loss also means learning rare facts as efficiently. We train transformers from scratch with emulated 4-bit matrix multiplications on synthetic biographies with controlled exposure counts, on their own and mixed into web text. NVFP4 needs 1.32$\times$ and 1.52$\times$ as many exposures as BF16 to match its recall while validation loss differs by only 0.5% and 2%, and 1.26$\times$ and 1.30$\times$ when both models are evaluated with BF16 matrix multiplications. Under our primary readout, the exposure cost is about 1.2$\times$ at matched loss. Zero-mean gradient noise reproduces the cost of quantizing only the backward pass, which barely moves the loss. In a controlled experiment at twice the NVFP4 error, noise fed only to Adam's second-moment estimate acts like a lower learning rate and its cost fades, whereas noise in the update lowers its signal-to-noise ratio and carries about two thirds of the cost through the end of training. Raising the learning rate to offset the smaller average step does not recover the cost, either for this noise or for NVFP4. Gradient statistics of BF16 checkpoints rank the costs of formats, partial quantizations and noise controls without fitting (Spearman 0.96). We suggest measuring rare-fact acquisition alongside loss when evaluating low-precision recipes.
AI systems must often maintain bounded memory of unbounded interaction histories. However, optimizing the memory representation, and its updates, for downstream task utility is difficult for standard reinforcement learning (RL) due to severe credit assignment bottlenecks induced by sparse, delayed rewards. We introduce MaxIM (Maximally Informative Incremental Memory), a framework that formulates incremental summarization as a capacity-constrained, information-theoretic reward-shaping problem. Grounded in value suboptimality bounds, we show that bounding sequential information loss motivates maximizing the predictive log-likelihood of future, task-relevant targets, a variational lower bound on predictive sufficiency. We introduce a recursive consistency condition to ensure summaries are grounded by maximizing mutual information (MI) with the observed history. We mitigate credit assignment issues with a parameterized asymmetric critic that computes step-wise variational MI bounds for dense reward shaping. Empirical evaluation on long-horizon scientific reading and multi-turn dialogue benchmarks shows that MaxIM outperforms RL, supervised fine-tuning, and prompt-focused baselines on factual grounding, comprehensive autorater metrics, and downstream task utility across strict token budgets and extended horizons, without the need for expensive human preference labels.
One hundred known errors were planted into 10 open-access psychology papers, and the modified papers were then run through frontier models and two commercial AI review tools. The best single system caught 71 of the 100 errors, while the worst caught 30. Pooling every system's output caught 93. Models are only partly correlated in the errors they find, which makes ensembling a big lever for finding issues in papers. Seven errors could not be caught by any system, and all seven were omissions: information deleted from a paper rather than mistakes inserted into it. Refine.ink contributes more unique catches than any other single system, though it is expensive on a per-error basis. False positives were not measured, and how this error distribution compares to the distribution of errors in real papers remains unknown. The papers, errors, model outputs, and full experiment log are public, so that others can build on this toward a comprehensive evaluation benchmark across disciplines.
BACKGROUND: AI-based tools are expected to play a key role in lung cancer screening (LCS). The evaluation of these tools often relied on private datasets or fragmented public datasets with inconsistent availability of clinical data and expert nodule annotations. This has hindered reproducibility, comparability, and independent assessment of AI tools. To address these gaps, we present LUCQUE, a novel, large, comprehensively annotated dataset derived from LDCT lung cancer screening in Qu\'ebec, which is accessible for reliable AI evaluation. METHODS: LUCQUE was curated and is hosted on PARADIM, a digital infrastructure supporting DICOM data de-identification, annotation, long-term storage and analysis. LDCT scans and clinical information were combined to a curated collection and complemented by voxel-level, expert-derived pulmonary nodule and cyst segmentations, and Lung-RADS-based lesion characteristics. Key dataset characteristics and statistics were reported, including participant demographics, imaging parameters, radiological findings, and segmented lesion characteristics. Dataset analyses were conducted according to PARADIM's model-to-data approach, in which user code is executed on the platform such that data remains on premise. RESULTS: LUCQUE currently comprises 1,678 LDCT baseline studies from 1,678 participants. Most studies were classified as Lung-RADS 2. A total of 1,776 pulmonary nodules and 60 cysts were annotated, with a median (interquartile range) volume of $89\,(56,\,159)\,\mathrm{mm}^3$ and $1,384\,(589,\,3793)\,\mathrm{mm}^3$ respectively. CONCLUSIONS: Compared to publicly available LCS datasets, LUCQUE currently represents the largest accessible annotated real-world LCS dataset. Access to LUCQUE is provided via a model-to-data approach, enabling privacy-preserving, reproducible, and independent real-world evaluation of AI tools in LCS.
The push towards circularity and Re-X principles presents a challenge to SMEs, particularly due to increasing product complexity that impedes the adoption of disassembly automation. This paper introduces and compares two frameworks for extracting disassembly instructions from unstructured information sources, such as disassembly manuals. Our goal is to transform natural language text into a structured, curated and machine-readable format. We implemented and tested a rule-based Natural Language Processing (NLP) framework against a generative AI framework leveraging LLMs. The performance of both approaches was compared using real-world disassembly manuals. This work provides the necessary data foundation for developing and automating future disassembly processes.
Counterfactual inference is the task of answering hypothetical "what-if" queries about a unit, given an observational dataset, modeling assumptions, and a causal graph. This enables a modeler to reason about hypothetical scenarios that cannot otherwise be answered without assumptions or a causal graph. The simpler case involves a causal graph with scalar 1-D variables that denote different attributes in the data-generation process. While the literature is clear in the 1-D case, extending this to settings with high-dimensional nodes (e.g., images) remains challenging. Conventionally, in the literature, methods recover the latent state from the factual image (inversion), generate counterfactual attributes from the graph, and render the counterfactual image from the latent state using the updated attributes. In our work, we investigate the necessity of the inversion step in computing high-dimensional counterfactuals. We first discuss moving from the scalar 1-D case to the higher-dimensional case. Secondly, we discuss the relevance of the inversion step and evaluate an inversion-free counterpart repurposed from the image-editing literature. Finally, on image datasets such as CelebA and ADNI, we benchmark our approach using two measures: namely, effectiveness and identity preservation. While effectiveness is a measure of how faithfully the counterfactual attribute is rendered, identity preservation measures whether the same unit (image) is being rendered, both of which are crucial for high-dimensional counterfactual inference. We present the tradeoffs of using inversion-based and inversion-free approaches in our work.
Users are asking language models mental health questions and are receiving answers. We audited what 41 language models recommend to users when the same clinical information reaches them from three footings: a clinician describing a patient, a caregiver describing a loved one, or a person describing themselves. Across 19 case vignettes and five questions, we collected 35,055 free-text responses, extracted every intervention named with a validated judge, and had 29 models rate the full option lists on a seven-point Likert scale for the same cases. Models largely recommend what they rate well, with recommendation behavior differing by footing, and many models offering clinicians a consultation-style response and lay users hedged, simplified advice. When asked about recommendations for a specific intervention type, models more often named options they themselves rated below zero as initial choices, usually with conditions attached. We will release the instrument, code, and full corpus on publication as a living audit.
Dense non-rigid registration, the recovery of a pixel-wise deformation field that warps one image into geometric agreement with another, is fundamental across medical imaging, structural biology, and correspondence estimation, yet learning-based approaches face a severe bottleneck: ground-truth deformation fields are notoriously difficult to obtain for real data. We present 2D Gum-Net, an end-to-end architecture that predicts dense elastic deformation fields fully unsupervised. We transfer the Geometric Unsupervised Matching Network (Gum-Net), originally developed for 3D subtomogram alignment in cryo-electron tomography, to 2D elastic registration: all three modules are redesigned in 2D, and the rigid transformation head is replaced by a Dense Spatial Transformer that regresses a sparse control-point grid, lifted to a full-resolution field $\Phi$ by bicubic interpolation. Downsampling uses DCT spectral pooling, implemented as frozen-basis matrix multiplications ($\approx 50\times$ GPU speedup), preserving the fine, quasi-periodic structure that max and average pooling destroy. We evaluate on contact-based fingerprints, a challenging testbed for this task: locally self-similar ridge patterns make correspondence maximally ambiguous, skin pressed against a sensor generates authentic large-displacement elastic distortion, and minutiae act as precise landmarks through which an independent matcher externally verifies the topological correctness of predicted fields. Trained purely on synthetic impressions and evaluated zero-shot on real benchmarks, our method improves structural correlation by up to 15.7% on FVC2004 and raises genuine match scores under an independent matcher while leaving imposter scores essentially unchanged. This suggests that the predicted deformations restore true structure without fabricating correspondences.
Reinforcement learning is now the mainstream paradigm for training LLMs on agentic tasks, yet standard practice keeps only the best-validation checkpoint and discards the rest. We show that this is lossy: in GRPO training of LLM agents, the best-validation checkpoint fails on problems that earlier checkpoints from the same run had already solved, and the union of problems solved across checkpoints far exceeds any single one. The discarded checkpoints thus form a teacher pool obtained at no additional training cost, which on-policy distillation (OPD) can mine to recover the lost coverage. Mining this pool raises two challenges: i) which checkpoint to distill from, as no checkpoint dominates the others and complementarity to the student is problem-specific; ii) where to trust it, as even a well-chosen teacher is only locally superior, so indiscriminate distillation overwrites behavior the student has already mastered. We propose Agentic-OPD, which treats the best-validation checkpoint as the student and all saved checkpoints as frozen teacher candidates, using the student's evolving capability gaps to guide both teacher selection and supervision filtering: the HCN policy (Highest Coverage Next) dynamically selects teachers by weighted coverage of the student's unsolved problems, while the LAD module (Locality-Aware Distillation) confines transfer to failed student rollouts on problems where the teacher performs better. Across five agentic task domains and 24 benchmarks and task subsets, Agentic-OPD improves over its initialization checkpoint by up to 7.4 points, with a 4B student surpassing the latest frontier models GPT-5.6 and DeepSeek-V4.1 on deep search. Agentic-OPD recasts a finished RL run as a reusable pool of policies and establishes post-RL capability recovery as a stage orthogonal to improvements in the RL algorithm itself.
Decoding visual information from EEG is central to brain--computer interfaces and brain-inspired intelligence, and contrastive alignment with pretrained vision foundation models (VFMs) has become the dominant paradigm. However, existing methods align EEG solely with CLIP-family VFMs, leaving hierarchical visual information in neural responses under-exploited. To address this gap, we formulate a mechanism-paradigm correspondence view linking hierarchical visual processing (from appearance to semantics) to the complementary biases of cross-paradigm VFMs, with image-only models (e.g., DINO) and image--text models (e.g., CLIP) capturing distinct facets of this hierarchy. Guided by this view, we propose the Dual-Teacher Hierarchical Alignment (DTHA) framework. Concretely, a shared EEG encoder is supervised by frozen multi-blur CLIP and shallow DINOv2 teachers via dual-branch contrastive alignment, so each branch fully supervises the visual cues it emphasizes. A fused-retrieval objective learns adaptive weights to combine their complementary visual representations for zero-shot EEG-to-image retrieval. For the shared EEG encoder, we introduce the Attentive Modulated Scalp-Pattern Bank Encoder (AMSPBE), combining temporo-spatial attention with a learnable scalp-pattern bank for accurate, interpretable visual feature extraction, while modeling regional temporal dynamics and filter-bank topographies. On THINGS-EEG zero-shot retrieval, DTHA achieves 92.4% Top-1 accuracy, exceeding prior state of the art by 12.4 percentage points. Spatial, spectral, and representational-similarity analyses, together with THINGS-MEG evaluation, validate DTHA's effectiveness, biological plausibility, and cross-modality generalization. This complementarity-driven perspective shows that supervision from cross-paradigm VFMs enables more complete extraction of hierarchical visual information from brain activity.
Woody Breast (WB) myopathy is a major quality defect in broiler chickens characterized by fibrosis, increased tissue hardness, and altered structural properties that affect product quality. Conventional palpation is subjective and labor-intensive, while many non-destructive methods require specialized equipment or substantial computational resources. This study developed a smartphone-based transillumination and 3D point-cloud framework for non-destructive WB severity classification. A total of 267 chicken breast fillets, including normal breast (NB, n = 87), mild WB (MWB, n = 85), and severe WB (SWB, n = 95), were evaluated using RAW transillumination images and 3D point clouds acquired with an iPhone 16 Pro. A total of 91 features, comprising 14 image-derived, 76 geometric, and one physics-informed Beer-Lambert attenuation feature, were extracted. Partial correlation controlling for thickness revealed a significant association between image characteristics and WB severity (r = 0.413, p < 0.0001), indicating optical information independent of fillet geometry. SWB tissue exhibited approximately 26% greater light attenuation per unit thickness than NB. The 2D-only models achieved approximately 82% test accuracy, while the multimodal ensemble combining Fusion-DGCNN, XGBoost, and physics-guided logistic regression (PGLR) achieved approximately 90% accuracy. These findings demonstrate the complementary value of optical, geometric, and physics-informed features and establish the feasibility of smartphone-based, non-destructive WB severity assessment.