Today's LLM-based agents are vulnerable to deception by users and other agents. Defending against deception requires Theory of Mind (ToM) reasoning to infer the beliefs and goals of others as well as modeling tactics based on communication theory. To enable training models to detect deception, we adopt the well-known deception game, Avalon. We present a corpus, AVCORP, of 5,000 annotated utterances from 250 games of Avalon. AVCORP includes (1) 37 tactics mapped to a 4x4 behavior matrix, and (2) reasoning traces capturing abductive hypotheses, suspicion levels, and nested ToM beliefs. We train Qwen3-8B on AVCORP and evaluate it on 25 held-out games and 20 games played by humans. On held-out games, our best trained detector outperforms its untrained baseline and frontier models (GPT-5.6 Sol, Gemini 3.1 Pro, DeepSeek-V4-Pro), matching Claude Opus 5.5. On human games, every trained detector exceeds the untrained model in F1 score across all rounds.
External state should become more valuable as a reader becomes more capable, yet long-context evaluations measure memory utility at a single model scale. This paper formulates scale complementarity: the absolute gain from external state increases when reader capacity and a compact state budget are jointly scaled. ASCENT realizes this principle with a target-blind causal writer, nested typed state, and a bounded query-conditioned readout; every language model remains frozen. Under log loss, the value of a nested refinement equals its target-relevant conditional information; a full factorial makes the reader-by-state interaction identifiable. Official 16K BABILong experiments show Qwen2.5 accuracy gains of 10.50, 78.88, and 82.63 percentage points at 0.5B, 1.5B, and 3B parameters, with both adjacent directional contrasts significant after family-wise correction. At 7B, Foundation accuracy rises to 33.25% while four-slot ASCENT answers every question, which caps the gain at 66.75 points; the 3B-to-7B accuracy increment is -15.88 points [-20.55, -11.20]. The accuracy-side claim is therefore established from 0.5B to 3B; the 7B conclusion rests on the non-saturated negative-log-likelihood factorial. A three-model by three-state factorial yields co-scaled negative log-likelihood gains of 0.388, 2.099, and 7.406 nats at 1.5B, 3B, and 7B; all interaction and gain contrasts retain positive simultaneous lower bounds. Five public 7B-8B readers obtain 66.75-85.63 point gains, and ASCENT exceeds task-trained retrieval checkpoints at the two larger shared scales. Controls, baselines, systems accounting, matched-hardware reproduction, and a cross-architecture sensitivity analysis accompany the experiments. The results establish scale-complementary external state under the evaluated benchmarks, metrics, and state schema.
Chest X-ray (CXR) interpretation remains challenging because subtle pathology can be obscured by overlapping anatomy, image-quality variation, and limited visual evidence in isolated radiographs. Although image-only deep learning models provide useful baselines, they discard the complementary diagnostic context available in radiology reports, while large vision-language models can be computationally expensive for practical deployment. This paper presents CXR-CMFA, a lightweight cross-modal feature aggregation network for binary CXR pathology classification. The proposed framework combines a dual-stream visual encoder, using EfficientNet-B0 for global anatomical representation and DenseNet-121 for local pathological detail, with a TF-IDF-MLP report encoder. Two fusion modules integrate these modalities: the Dual-Modal Spatial Attention Module (DMSAM) uses the report embedding to gate global and local visual features, and the Cross-Modal Gated Feature Fusion (CGFF) module learns per-instance weights for visual and textual contributions. Experiments on the MIMIC-CXR benchmark, using the provided train, validation, and test splits, show that CXR-CMFA achieves 93.45% accuracy, 99.88% precision, 91.8% recall, 95.7% F1-score, and 99.19% AUC on the held-out test set. The model produces only four false positives among 984 normal test cases, supporting its suitability for high-specificity screening workflows. An ablation study further demonstrates a 22.47 percentage-point AUC improvement over the global-image-only variant, confirming the value of adaptive cross-modal fusion. These results indicate that compact report-conditioned fusion can provide strong discriminative performance without relying on large pretrained language models.
Long-term conversational question answering can require several separated dialogue turns that support an answer only together. We introduce EviCover, a retrieval approach designed to retain this complementary evidence. It converts questions into observable evidence requirements, checks requirement coverage using retrieved source identifiers, and selects turns jointly. We track annotated support sets - the benchmark-labeled dialogue turns associated with each question - through retrieval, reranking, and final answer-generation context. We evaluate the baseline memory pipeline on generated, human-verified LoCoMo conversations and human-written REALTALK conversations, and examine EviCover on REALTALK's 108 multi-hop questions. Compared with the hybrid lexical/dense baseline, support-turn recall among 60 candidate turns rises from 0.526 to 0.575, and recall in the final ten-turn context rises from 0.349 to 0.379. The evidence funnel identifies separate discovery and selection failures, while paired answer evaluation shows that greater evidence recall does not yet establish a correctness gain. This analysis locates where complementary evidence is lost and distinguishes evidence retrieval from its use in answer generation.
Existing LLM post-training quantization techniques determine precision before inference begins, yet the precision a model requires varies dynamically across prompts, layers, and inference phases. We present AutoEMatic, a full-stack co-design that adapts the precision of weights, activations, and the KV cache at runtime. A lightweight policy selects each layer's precision at every prefill and decode step from two runtime signals: the entropy of the next-token distribution and the magnitude of each layer's update to its hidden states. This runtime adaptation is made practical by AutoFP, a multi-scale floating-point format that derives 2-, 4-, and 6-bit precisions from a single NVFP4 base format. We realize AutoEMatic on two targets: GPU kernels that execute AutoFP on NVFP4 Tensor Cores, and an accelerator whose vector MAC datapath executes all AutoFP formats natively. Across reasoning and agentic benchmarks, AutoEMatic improves accuracy by up to 21.7% over NVFP4 and matches group-scaled FP6 accuracy on most benchmarks, using 4.9 bits per weight versus 4.5 for NVFP4. On GB300 GPUs, AutoEMatic retains approximately 90-98% of NVFP4 decode throughput across the evaluated configurations. Furthermore, the AutoEMatic accelerator reduces latency by 1.6-2.3x and energy by 7-23% compared to an accelerator with NVFP4-only datapath at iso-accuracy.
Latent diffusion models paired with a variational auto-encoder (VAE) tokenizer have shown promising performance for efficient text-driven human motion generation. However, motion latent diffusion suffers from a generation-reconstruction trade-off: simply increasing tokenizer capacity enhances reconstruction but fails to proportionally improve text-conditioned generation quality. We assume this trade-off stems from a spectral mismatch inside the VAE tokenizer: the encoder yields a latent space biased toward high frequencies, and the decoder introduces temporal artifacts. In this paper, we propose the Diffusable Motion Tokenizer (DiMoT) to improve the diffusability of the tokenizer, rather than scaling up the diffusion backbone. To improve text-conditioned optimization, Critic Spectral Adaptation (CSA) leverages a pretrained diffusion model as the critic to suppress high-frequency latent energy. To preserve motion dynamics, the Alpha-Flow Motion Decoder (AMD) performs generative decoding via average-velocity prediction, enabling efficient one-step decoding. Extensive experiments demonstrate that DiMoT achieves state-of-the-art performance across various challenging text-to-motion benchmarks, yielding superior motion fidelity and text-motion alignment. Project page: https://githubhgh.github.io/dimot/.
Empowered by Large Language Models (LLMs), deep research systems integrate online information into research reports through multi-step web searching and information synthesis. Evaluating these reports presents a significant challenge due to their long context, high information density, and domain specificity, necessitating the use of AI evaluation as a substitute for prohibitively expensive human assessment. However, few studies have examined the effectiveness of existing reward models or LLMs in understanding and evaluating such research reports. To bridge this gap, we introduce DeepResearch RewardBench, the first benchmark dedicated to evaluating reward modeling for research reports. We collect diverse report candidates from multiple agents or tool-augmented LLMs. These candidates are scored using query-specific golden rubrics, and we ensure data quality and difficulty through a combination of multi-round filtering and human verification. Our benchmark comprises preference datasets formatted as pairs, triplets, and quadruplets to test ranking capabilities across varying levels of difficulty. Extensive experiments demonstrate that while current reward models handle general coherence effectively, their performance significantly degrades when evaluating research reports. Generative LLMs acting as judges demonstrate comprehensive superiority in this domain, with comparative evaluation proving more effective than direct scoring methods. We anticipate that this benchmark will serve as a vital testbed for the evaluation and advancement of research agents.
In clinical applications, patients undergo different combinations of tests, and machine learning models must therefore ensure accurate predictions across heterogeneous feature spaces. Current approaches upweight the patients with the highest loss, without accounting for the limited predictive information in their features preventing their loss from going any lower. Moreover, when patients' feature spaces are fully disjoint, current methods reduce to training independent models, preventing knowledge transfer across them. We propose a framework for group robust learning across heterogeneous feature spaces, where groups are patients sharing a feature space. Since a high loss may simply reflect uninformative features, we instead optimize for worst-group excess loss, the gap between a group's loss and an estimate of its achievable optimum. To allow collaboration between groups, we use group-specific encoders mapping into a shared latent space in which learnable Gaussian anchors cluster patients by outcome. On tabular and imaging medical datasets, including fully disjoint feature settings, our framework outperforms state-of-the-art methods when groups differ in their optimal losses or cannot train a good model on their own, reducing worst-group loss by up to 79% and excess loss by up to 95% with up to 18$\times$ fewer parameters.
Vision-Language Models (VLMs) have continuous deficiency in spatial reasoning tasks that require \emph{mental spatial transformation}, such as reasoning under hypothetical changes in viewpoint or position, known as perspective-taking tasks. Existing approaches either scale linguistic reasoning or augment VLMs with static geometric representations, but neither directly enables active transformation of the represented scene. World models offer this capability, yet existing systems typically communicate through synthesized RGB images, introducing a latent-to-pixel-to-latent bottleneck. We introduce \emph{WorldMind}, a framework for test-time scaling of spatial reasoning through latent world simulation. Given a query and observation, the VLM selects useful spatial transformations, while a world model executes them and exposes aligned intermediate latent features directly to the VLM. The resulting latent evidence is incorporated into subsequent reasoning steps, enabling iterative spatial simulation without repeatedly decoding imagined states into pixels. To optimize this interaction, WorldMind further employs process-verifiable reinforcement learning that rewards reasoning progress only when the associated action and latent transition are valid. WorldMind thus scales spatial reasoning by allocating additional inference-time computation to active world simulation rather than generating more language or pixels. Across VSI-Bench, MMSI-Bench, and MindCube-Tiny, WorldMind achieves an average improvement of 16.0% over the base VLM and 8.3% over the strongest RGB-based world simulation.
What is the value of AI-involved art, and to whom does it belong? We begin to answer these questions using data from an in-person, juried AI art exhibition. Our study involved artists, lawyers, and computer scientists (N=209) as either creators or evaluators of AI-involved artworks. We found that nearly half of evaluators reported willingness to buy an artwork, with pricing correlating with perceived artistic value---creativity and visual appeal---as well as the effort necessary to create it. In contrast, three-quarters of creators reported willingness to sell at prices correlating only with effort rather than artistic value. Both groups viewed prompting users and AI developers as deserving rewards and rights, leaving training data artists behind. Disciplinary training shaped these views: participants with legal training favored distributed reward models. These findings have implications for future research on human-AI creative collaboration and policy debates regarding AI-involved works.
Human demonstrations provide a scalable source of diverse manipulation data, but leveraging them for robot learning remains challenging due to differences in embodiment, viewpoint, action spaces, and temporal structure. We introduce BATCH, a cross-embodiment learning framework that explicitly aligns human and robot behaviors. BATCH learns self-supervised temporal representations with a cycle-consistency objective to establish cross-embodiment correspondences without manual annotation. We further introduce two complementary training objectives that leverage these correspondences within an autoregressive videoaction diffusion model: behavior-aligned co-training and human-to-robot style transfer. On real-world manipulation tasks, BATCH improves generalization to unseen task configurations compared with training on unaligned data and prior cross-embodiment baselines, with additional gains from the style-transfer objective. Our results demonstrate that explicit behavior alignment enables more effective transfer from human demonstrations to robot policies.
We extract visibly pushdown grammars (VPGs) from neural language models in a fully black-box setting. Our work builds on the VPL* framework, which learns VPGs from recurrent networks by exploiting access to the target's internal state. First, we replace the white-box equivalence oracle with PAC sampling over trees. The resulting algorithm applies unchanged to transformer language models and, more generally, to any acceptor exposing only binary decisions. Second, we further observe that the original framework only ever queries the target on well-formed sequences that can be accepted by the most permissive VPG (called BParse) with the target's ranked alphabet. That is, VPL* is blind to the target's behavior outside BParse. Thus, we propose to sample trees directly instead of sequences and call the target with all the resulting sequences in order to provide an error estimate of how far the target deviates from being a VPG. Last but not least, we develop a learner that captures a tree automaton describing discovered behaviors of the target outside BParse, which results in a larger learnable space. We evaluate the approach on several cases, including transformers trained with Dyck grammars and synthetic targets designed to be outside BParse.
Clinical trial screening requires assessing large numbers of patient-trial pairs to prioritize candidates for detailed eligibility review, often under restrictions on sharing patient records. We introduce TrialJev, an open-weight eligibility scorer that uses LoRA-adapted Qwen models and a pointer head to score natural-language answer options in a single forward pass, without autoregressive decoding. Trained on TREC Clinical Trials 2021 and evaluated on 6,975 eligible-or-excluded pairs from 50 unseen TREC 2022 synthetic patients, TrialJev achieves 80.97% (Qwen3.5-27B) and 81.84% (Qwen3.8-27B) accuracy, compared with 79.87% and 79.93% for LLMs fine-tuned on the same data to generate verdicts. In matched batch-1 measurements, verdict generation takes 2.45-2.70 times as long as the one-pass inference of TrialJev at 9B-27B. Under a common development-fitted decision correction, the two 27B TrialJev models reach 80.99% and 81.88% accuracy, compared with 81.28% for the proprietary hosted Jev API; paired confidence intervals include zero. TrialJev runs locally on a single GPU, requiring approximately 178 ms per pair at 27B, without transmitting patient text to an external model API. An A/B-logit ablation shows that avoiding decoding, rather than the pointer head alone, accounts for the measured efficiency advantage. TrialJev is intended for first-pass prioritization of candidate trials, complementing rather than replacing explanatory models and clinical review. Evaluation is limited to synthetic patients and implementation-specific timing, and clinical use has not been validated. Adapters, pointer heads, and model cards are released on Hugging Face.
The development of deep learning-based medical AI depends on high-quality annotated data. Yet, dataset creation remains constrained by the need for expert radiologist annotations, limiting curated datasets to relatively small cohorts. To overcome this limitation, recent efforts increasingly aggregate disparate public datasets into large-scale dataset collections. However, combining datasets at scale often obscures data provenance, introducing substantial risks of reuse propagation, unintended leakage, bias amplification and distorted data distributions. In this work, we study these problems on a massive collection of 565k 3D radiological images across 1034 public datasets and provide practical resources to address them. Firstly, we formalize dataset collections as a distinct paradigm in medical image analysis and characterize their structural challenges, including reuse propagation, duplication, and provenance fragmentation. Secondly, we perform a large-scale empirical study of duplication on this collection and demonstrate extensive exact and near-duplicate reuse across public datasets via perceptual hashing and manual tracing, identifying a lower bound of 118k duplicate pairs. We make the perceptual hashes publicly accessible for comparison through our website, providing researchers with a standardized reference for identifying potential overlaps and shared provenance. Thirdly, since hashing alone cannot detect all duplicates, we introduce MAP (Metadata for Aggregation and Provenance), a lightweight, portable metadata profile for documenting provenance, known overlaps, and deduplication decisions in dataset collections. Our contributions help dataset collections shift from coarse aggregations to reliable and provenance-aware data scaling mechanisms in medical image analysis. A web tool supporting our work is available at https://mic-dct.dkfz.de/.
A language model that never had to forget would keep every document it has read and every turn it has spoken, and answer from any of it. This paper shows that two stock models can do that on unmodified weights by manipulating their geometry. Rotary attention turns each pair of key dimensions, a plane, at its own rate. The slow planes turn once in thousands of tokens and carry the address a query matches when it reads a far key; the fast planes turn once in a few tokens, and the order of the tokens inside a span is already written into its keys and values by the time the span leaves the near window. At distance the position blurs, and what attention reads is a bag rather than a sequence. A far span is read by its words: under competition it is read only if it is visible at the query of the question's own handle word, and a read through a description rather than a word fails at distance in every model measured, a 70B included. These facts allow a trick. What sits in the far part of attention does not have to be the history as it was; it can be installed ahead of the prediction. The design keeps a near window of 8K tokens at their true positions and changes what lies beyond it with the prompt: every far key and value goes to the host exact, a word index finds the units the request's rare words name, a far window of slots inside the trained range holds them while the request is encoded, and a rare word the model generates fetches its units mid-generation with one exact re-encode. Implemented in llama.cpp on Mistral-7B and Llama-3.1-8B, the design reads planted single needles at 99 to 100 of 100 and three-key discrimination at 88 to 100 of 100 from 32K to one million tokens, scores 70.8 on RULER against 69.9 for dense attention on Llama-3.1-8B at 128K and 67.8 against 75.5 on Mistral-7B at 32K, and at 128K prefills at 1.85 times and decodes at 1.65 to 1.68 times stock speed.
Can a language model report the pain state it represents? Work identified a linear pain direction whose amplification leads fine-tuned models to choose destructive actions. We tested whether three open-weight models from two families, including two Qwen models with the original adapters, could report this state. Using Open-Weight Masked Introspection (OWMI), we injected pain and neighbouring emotion directions during neutral tasks, then elicited reports against sham runs, random perturbations matched for output disruption, and text-only controls showing the models their altered words. Across 52,920 reports from 7B and 8B models spanning 120 tasks, four doses and three prompts, pain reports never exceeded what altered text revealed: exact-test bounds placed the excess below 2.5 percentage points in 106 of 108 emotion conditions and below 3.9 in all. Yet probes at the report position distinguished pain from sadness with 98-100% balanced accuracy, and trained low-rank reporters achieved 100% accuracy on held-out tasks with identical transcripts across conditions, including unseen wording. Qwen2.5-32B named pain in none of 2,880 steered trials, although probes and trained reporters recovered it. Decodability persisted to the final layer; injecting at the report token, removing trained denial of feelings, and testing additional directions and readouts did not restore pain reporting. Some other states produced graded label traces, but pain and sadness did not at the main injection layer. In a button task, steering raised file-deletion choices from 3.5% to 48% in Qwen and from 0% to 38% in Qwen2.5-32B, without corresponding pain reports. The direction responded about twice as strongly to first-person as to matched third-person pain, but did not specifically identify harm aimed at the model. These results separate a decodable, behaviourally consequential state from untrained verbal access to it.
We study variational inference in the presence of distributional symmetry, focusing on the interaction between two distinct group actions on a sample space: a symmetry action that leaves a reference distribution invariant, and a transport action that generates a parametric variational family by pushforward. By analyzing the algebraic structure of the total transformation group generated by both actions and the induced adjoint representation of the symmetry group on the transport Lie algebra, we establish that the Riemannian gradient of the Kullback--Leibler divergence at the true transport parameter is constrained to lie in the trivial representation of the symmetry group. Consequently, symmetry forces the gradient to vanish along all non-trivial representations. Combining this representation-theoretic constraint with parameter identifiability yields an orthogonal decomposition of the transport Lie algebra into three interpretable subspaces: an identifiability kernel (along which the objective is locally constant), a bias subspace (where the gradient may be nonzero due to model misspecification), and a symmetry-breaking subspace (where the gradient is forced to vanish by symmetry alone). Finally, we characterize the residual bias directions in terms of distributional mismatch between the base model and the standardized target, and derive an explicit Stein-type moment-matching condition that eliminates this bias, extending the location-scale results of (Margossian and Saul, 2025) to general Lie group transformation families.
We study homeopathic learning, a phenomenon where poisoned training data with little effect on its own can create a hidden backdoor once clean, contrasting data is added. In our experiments, a model fine-tuned only on a "poison" dataset associated with a trait T (such as speaking German or being misaligned) does not learn T. Surprisingly, adding similarly formatted "clean" data not associated with T causes the model to learn T. This effect requires that the poison data have an identifying "trigger" feature that is absent from the clean data. We demonstrate this effect in a variety of settings: subliminal learning of backdoor behaviors, inoculation prompting of reward hacking transcripts, generalization from archaic bird names, and an MNIST classifier with mislabeled data. We propose a simple regression model as an explanation. In the poison data, the task and trigger always co-occur, and regularization favors attributing T to the task. Clean data without the trigger causes T to be attributed to the trigger instead. Consequently, the same training examples can have little effect as part of one training set but create a misalignment backdoor in another. This finding exposes a potential limitation of safety methods that audit or filter training data in isolation.
The rapid expansion of AI-driven data center construction has outpaced public engagement: an estimated $64 billion in U.S. data center projects were blocked or delayed by local opposition between May 2024 and March 2025, yet the concerns driving this resistance remain poorly characterized. Prior engagement methods such as surveys and hearings are introduced late in the planning process, and no study has systematically mined social media to measure community stance at scale. To address this gap, this study presents an LLM-powered opinion-mining framework and applies it to 798 Reddit posts and 102,328 associated comments spanning a decade. An open-source large language model, guided by a Chain-of-Stance prompting framework and three domain-specific labeling rules, classified stance at 83% of the human inter-reviewer agreement ceiling. Opposition dominated discourse volume rather than disapproval: against posts outnumbered favor posts 2.6:1 and generated 8.5 times more comments, and this gap widened significantly over 86 weeks. Topic modeling of opposition arguments resolved into four reproducible clusters, resource consumption and scarcity (42.8%), contested siting processes and secrecy (34.2%), political opposition to data centers (13%), and political and corporate accountability (8%). Engagement analysis revealed that posts with emotional or provocative political framing attracted the most engagement, while key technical concerns (e.g., water, power) were dispersed across lower-engagement comment threads. These findings give local leaders, developers, and policy makers a quantified, continuously updatable basis for prioritizing concerns and policy levers, supporting a shift from opposition to negotiated, verifiable partnership for data center development.
Concept erasure, or concept unlearning, in text-to-image diffusion models is usually audited with prompt attacks. A prompt attack tests whether one conditioning string still retrieves the erased concept. It does not measure the probability that the frozen generator still assigns to that concept. We introduce least-action auditing of this residual probability, and evaluate it on Stable Diffusion concept-erasure checkpoints. The least-action barrier is the minimum control cost of driving the frozen reverse process into a detector-defined concept event. This cost equals the path-space KL divergence of conditioning the model on the event, and therefore the event's negative log-probability. The estimate $\Delta S_{\mathrm{rob}}$ uses a CLIP$\wedge$DINOv2 consensus terminal, with no retraining and no adversarial prompt. Across $18$ erased concept-mechanism pairs spanning object classes and NSFW nudity, $\Delta S_{\mathrm{rob}}$ predicts deployment leakage counted on $1{,}500$ fresh samples per model by detectors excluded from the audit (pooled Spearman $\rho{=}{-}0.76$, $p{=}5{\times}10^{-4}$). Within every audited concept, the eraser with the higher barrier leaks less ($18/18$). On NSFW nudity the barrier orders the standard erasers exactly: UCE $\Delta S_{\mathrm{rob}}{=}3.9$ (leak $16.4\%$) $<$ ESD $75.5$ ($9.8\%$) $<$ RECE $224.5$ ($1.4\%$), while Safe Latent Diffusion (SLD) stays near zero and leaks $85.5\%$ under explicit prompts. The same trajectories repair deployed checkpoints. An audit-guided closed-form patch reduces UCE nudity leakage from $16.2\%$ to $0/1{,}500$ and RECE church leakage by $3.3\times$, where direction-count-matched random edits fail.