Structure--Residual Inductive Biases for Language Model Pretraining
Hai Huang, Bartosz Żuk, Rassul Magauin, Jiaxin Huang, Xue Liu
AI 速读 基于摘要
作者提出结构-残差归纳偏置,通过边界瓶颈概念和连续概念两种方法,在语言模型预训练中改进下一词预测。
cs.CL · cs.AI
查看原文摘要
Next-token prediction is relied upon to learn both predictive behavior and internal representations. We ask whether representation structure itself should be an explicit inductive bias. Motivated by the two-part coding view of Kolmogorov complexity, we study a structure--residual principle in which a representation separates structure from realization detail. We instantiate this principle in two complementary ways. Boundary Bottleneck Concept (BBC) routes information from token segments through boundary carriers while preserving local autoregressive computation. Continuous Concept (CC) regularizes representation geometry and induces a spectral decomposition into low-frequency structure and high-frequency residual. We derive coding bounds for both views and show that multiplicity-corrected boundary attention becomes exact as the within-segment path energy vanishes. Empirically, BBC and CC exhibit strikingly different depth preferences: BBC works best in early layers, whereas CC is most effective in later layers. At 300M training tokens, combining BBC and CC improves NLL by 1.44\% on GPT-2 Medium. These gains persist with training runs up to 20B tokens. Representation diagnostics show that BBC increases the relative predictive importance of boundary states, while CC reduces spectral residual energy. Together, these results suggest that structure--residual biases can improve language-model pretraining beyond next-token prediction. Code: https://github.com/haihuang-ml/structure-residual
WoRD: Diffusion-Based Posterior Inference for Online Goal Recognition
Almog Anschel, Sarah Keren
AI 速读 基于摘要
WoRD 用学习到的轨迹先验与扩散去噪链,从部分观测中推断智能体目标,并在模糊前缀下保持校准的多模态后验。
cs.AI · cs.LG · cs.RO
查看原文摘要
Goal recognition is the task of inferring an agent's intended goal from its observed behavior. Most existing work focuses on high-level symbolic recognition, abstracting away the continuous motion through which agents reach their goal. Methods that do reason over continuous trajectories rely on an explicit dynamics model that is hard to acquire and rarely transfers across settings. Neither approach explicitly learns how agents move toward each candidate goal, limiting accuracy on ambiguous trajectory prefixes where such knowledge is most needed. We present WoRD (Waypoint Recognition with Diffusion), which uses a learned trajectory prior to infer the agent's goal from partial observations. At each step, WoRD runs one warm-started denoising chain per candidate goal, steers each chain toward the observed trajectory through a closed-form log-likelihood gradient, and reweights the resulting samples by Bayes' rule into a calibrated posterior over goals. We evaluate WoRD on simulated robot settings, pedestrian trajectories, and vehicle trajectories using four metrics targeting distinct failure modes. Across all three settings, the same inference procedure preserves calibrated multimodal beliefs through ambiguous prefixes where filter- and classifier-based observers collapse to a single mode or grow overconfident.
Automated segmentation is routinely used as the first step in quantitative image-based clinical workflows, where predicted masks are subsequently reviewed and corrected by human operators. Automated correction of the residual structural errors is currently addressed one anatomy at a time, so the number of trained correctors scales with the number of segmentation targets. We present GRACE, a retrieval-augmented refiner that edits an input mask under context drawn from a shared knowledge base of verified prior cases. The refiner is trained once on a multi-organ corpus spanning compact organs, cavities, branching vessels, and tumour lesions. In the held-out dataset-transfer evaluation, three held-out datasets query a fixed knowledge base built from training data only. A separate leave-one-anatomy-out experiment tests parameter-free index extension: verified cases of an unseen solid organ enter the knowledge base, and the refiner stays frozen. A light shape-aware projection head on a frozen multimodal encoder maps embedding similarity to the shape agreement of the masks, so the knowledge base returns cases that the refiner can use as a shape prior. We evaluate GRACE on 18 public datasets spanning CT, MRI and ultrasound, using nine refinement configurations and seven upstream segmentors that span weak, mid, and strong operating regimes. We report each change with the DSC of its input mask. On the weak tier, the input DSC is 0.386 and GRACE adds up to +0.35 DSC. On a strong segmentor, the input DSC is 0.932 and the mean change is small. Transfer to an imaging modality that is absent from training and from the index is partial and depends on the anatomy, and remains an open research direction. Our results suggest retrieval-conditioned refinement is a practical alternative to anatomy-specific correction models. Code: https://github.com/robustml-eurecom/grace
Text-to-video diffusion models generate temporally coherent content from natural language, yet when a prompt describes an early scene that persists while a new event emerges on top of it (such as "a tall sandcastle standing on a beach where a wave rushes in and washes it away"), generation frequently fails to realize the late-segment event in the corresponding frames. We identify this failure as Temporal Prior Suppression (TPS), in which the dominant prior of the early segment captures the cross-attention trajectory across the temporal axis and suppresses the guidance signal needed for late-segment realization. We introduce Temporal Prior Decoupling (TPD), a training-free heuristic that strengthens the suppressed late-segment signal during diffusion sampling. TPD constructs a temporal counterfactual by conditioning on the early segment alone, and defines the discrepancy between the full-prompt and counterfactual predictions as a suppressed signal direction. Rather than removing a direction as in prior subtractive projection methods, TPD adds this direction back to the guidance update whenever the alignment of the update with it falls below a threshold. The threshold follows a hand-designed schedule over diffusion timestep and video frame, so the correction acts mainly on late frames and structure-forming steps and leaves early frames largely untouched. TPD runs entirely within standard diffusion sampling without retraining and is defined purely in classifier-free guidance space, so it applies to any backbone that exposes a CFG update. Experiments show that TPD substantially improves late-concept realization while preserving temporal coherence and visual fidelity, and that the targeted suppression recurs across distinct text-to-video backbones.
We investigate how language models process negation in natural language inference (NLI). We propose the \textit{Polarity Composition Hypothesis}, a high-level causal model of this computation. Behavioral evaluations show that language-model predictions largely conform to the proposed hypothesis. Using Distributed Alignment Search, we identify causally effective subspaces aligned with the variables in the polarity-composition hypothesis. These representations transfer across negation forms and partially to a subset of MNLI. We release the dataset and code.
This paper proposes a measurement standardization framework that compresses expert-AI interactions into structured, comparable fields for prospective risk detection in deployed AI systems, without access to model internals. This concept paper defines the framework's scope, semantically and statistically, and specifies a protocol for its empirical testing. The population-level claims it is designed to support therefore belong to a staged research program rather than to results claimed here. Measurement standardization underpins three claims. The first is a reliability claim: under bounded conditions, large language models can produce reliable, standardized assessments of the evidential and policy alignment of expert-AI interactions. The second is a governance claim: alignment scores give experts an immediate signal during deployment and give institutions a basis for monitoring alignment patterns across mission types, models, and domains. The third is an outcome validation claim: once measurement standardization is established, aggregate alignment scores could be used to study associations with downstream outcomes in regulated professional settings. This introduces the possibility of an "AI epidemiology", a form of risk detection based on correlated variables instead of mechanistic analysis, inspired by epidemiological reasoning. A minimal application of the protocol to a published expert-AI corpus shows that the judge reproduces its policy and evidential alignment scores across two runs under the specified conditions. Judge reliability at scale remains to be validated in future work. The paper sets out a defined grammar of eight interaction fields, together with a statistical protocol based on paired bootstrap inference, DeLong's test for paired AUCs as a sensitivity check, a pre-specified one-sided non-inferiority margin of 0.05, and Holm-Bonferroni correction.
We present the first complete implementation of $β$-reduction in the untyped $λ$-calculus using exclusively spiking neural dynamics. Every step of the reduction---redex detection, subterm segmentation, variable-binding detection, De~Bruijn index arithmetic, token copying, and inter-step role maintenance---is performed by leaky integrate-and-fire (LIF) neurons with explicit membrane dynamics, exponential synaptic currents, axonal delays, and pair-based spike-timing-dependent plasticity (STDP). Each neural component maps onto a specific, experimentally documented mechanism: Jeffress-type delay-line coincidence for redex detection; a line-attractor balance counter for subterm segmentation; $θ$--$γ$ cross-frequency coupling for De~Bruijn variable binding; a phase-ring traveling wave for depth-index arithmetic; theta-cycle STDP for token copying; and recurrent persistent firing for role maintenance between reduction steps. All synaptic weights are derived analytically from measured EPSP amplitudes; no hand tuning is performed. The closed-loop system achieves 100\,\% agreement with a symbolic oracle across ten validation classes, including Church numeral arithmetic (\textsc{plus}, \textsc{mult}, \textsc{pred}), Boolean logic with divergent branches, Y-combinator steps, terms with up to six bound-variable occurrences, De~Bruijn indices up to~12, noise robustness, and determinism. To our knowledge this is the first Turing-complete symbolic computation performed entirely within spike dynamics using biologically grounded circuit primitives. We discuss implications for neuromorphic proof assistants and the neural basis of symbolic reasoning.
Long-context decoding, whether over long reasoning traces, long videos, or robot episodes, is increasingly limited by the memory of the key-value (KV) cache. Page-retrieval sparse attention makes each decoding step cheaper by attending only to the top-scoring KV pages. However, it still stores every page exactly, and adds a small digest per page for scoring, so the cache ends up slightly \emph{larger} than the full cache. We propose OVAL, which stores each page as a pair of output-aware low-rank bases that act as both the retrieval index and the storage. A key basis scores the page and rebuilds its keys when the page is selected, and a PCA basis rebuilds its values. The exact pages can then be dropped, and nothing extra is added to the cache. The key idea is that the key basis should be output-aware. A low-rank error in the keys perturbs the attention output at first order, but perturbs the page mass that retrieval ranks only at second order in expectation. OVAL therefore builds the key basis from a mix of key geometry (key PCA) and the attention-output error, with a per-head mixing weight set in closed form. At the same attended budget, OVAL is best or tied-best in eight of nine long-reasoning settings. It matches FullKV on MATH500 with 66\% of its memory, is 1.25 points above the best baseline on AIME24 with 56\%, and is on par with FullKV on LongBench~v2. On \mbox{Video-MME} and LVBench, it is within 0.32 points of exact-page retrieval at the same read, while storing 53\% of the full cache instead of 103\%. Its savings also stack with quantization and token pruning: with 4-bit bases, it is within 0.25 points of FullKV at 41\% of its memory.
Say Less, Hallucinate Less? Diagnosing and Restoring Object Coverage in Preference-Optimized Vision-Language Models
Hansang Lee
AI 速读 基于摘要
作者发现偏好优化在降低物体幻觉的同时会减少物体类别覆盖,并提出无需训练的基础模型引导恢复方法。
cs.CV · cs.AI
查看原文摘要
Preference optimization (PO) is the standard way to reduce object hallucination in large vision-language models. We find that four public PO checkpoints of LLaVA-1.5-7B also lower object coverage. They name fewer object categories, and their lower hallucination rates bring no significant gain in set-level object F1 on CHAIR. We therefore ask how much of the PO gain the base model can reproduce by stopping earlier. A positive end-of-sequence (EOS) logit bias lets the base model reach the recall of each PO model, and this early stopping alone reproduces 58 to 86 percent of the CHAIRi gain of three checkpoints on two disjoint COCO sets. Writing longer does not bring the lost categories back. Under greedy decoding, a negative EOS bias recovers only 2 to 9 percent of the lost CHAIR recall of these checkpoints, and 94 to 96 percent of the added object mentions repeat categories already named. On CHAIR, the base caption of the same image still names about half of the missed object categories. We therefore propose base-guided object-category restoration, which appends high-confidence base-only categories to the PO caption without training. With thresholds fixed on a development set, it brings three checkpoints within one point of the base recall on a held-out COCO set with a lower distinct-object hallucination rate than the base model. Against the recall-matched base stopping curve, its gain holds for RLAIF-V near the base recall and for RLAIF-V and CHAIR-DPO at an intermediate recall. We will release the code of the recall-controlled evaluation.
A working-memory model must remember stimuli and use its memory flexibly to perform different tasks. Here we introduce a model of working memory for images that combines dynamic activity with temporary associations between image features. We train the model on five tasks, including remembering a sample while other images are presented, categorizing distractor images while remembering the sample, and determining whether an image is new or has been presented before. The same trained model performs all tasks, guided by explicit cues about which response is required. We use this flexible memory model to ask how delay, distraction, and image history affect performance. Delays have little effect on sample-memory performance for the delays we tested, while seeing more images leads to more false alarms even when the target image appeared equally recently in all cases. On the sample-memory-plus-categorization task, the probability of correct responses on all questions decreases as more questions are asked, despite robust sample-memory performance. Finally, the model maintains stable performance across 40 brief trials of each task under the tested conditions. Together, these results demonstrate how a single working-memory system can be flexibly used to perform different tasks, and how its errors depend on both previous images and task demands.
Empirical research is changing rapidly as AI coding agents are becoming increasingly proficient in every step of it, from writing code to manipulating data. We argue that while the potential of coding agents for science is vast, their use in hypothesis testing poses a substantial threat to classical inference, as these agents enable a supercharged form of selective inference. In particular, we argue that coding agents carrying out empirical research introduce a new type of post-selection inference that combines human-like, open-ended search over an implicit space of estimators with programmatic replayability of both the selection and the estimation process. Indeed, we find that on simulated data with zero effects, four frontier agents report significant p-values between 54% and 98% of the time when asked to produce a compelling finding. We propose a simple procedure for detecting and correcting these distortions. After an agent has completed its analysis, we reset and replay it on subsamples of the data to estimate the agent's relative probability of selecting that analysis when it looks statistically favorable. For agents with stable selection behavior, the resulting corrected test is asymptotically valid and consistent conditional on selection. Importantly, the correction is learned from the black-box agent's actual choices rather than a model of how it searches, so it remains valid even as AI technology continues to evolve rapidly. In simulations, the correction robustly lowers the high false rejection rates to near the nominal 5%. We illustrate the method on applications from two published papers.
EmbeWebAgent: Embedding Web Agents into Custom Enterprise UIs
Chenyang Ma, Clyde Fare, Matthew Wilson, Dave Braines
AI 速读 基于摘要
提出一种把智能体直接嵌入企业自有前端界面的框架,借助轻量前端钩子与可复用后端工作流,实现比截图或原始 DOM 观察更稳健、动作表达更丰富的 Web 自动化。
cs.AI · cs.SE
查看原文摘要
Most web agents operate at the human interface level, observing screenshots or raw DOM trees without application-level access, which limits robustness and action expressiveness. In enterprise settings, however, explicit control of both the frontend and backend is available. We present EmbeWebAgent, a framework for embedding agents directly into existing UIs using lightweight frontend hooks (curated ARIA and URL-based observations, and a per-page function registry exposed via a WebSocket) and a reusable backend workflow that performs reasoning and takes actions. EmbeWebAgent is stack-agnostic (e.g., React or Angular), supports mixed-granularity actions ranging from GUI primitives to higher-level composites, and orchestrates navigation, manipulation, and domain-specific analytics via MCP tools. Our demo shows minimal retrofitting effort and multi-step behaviors grounded in a live UI setting. Demo Video: https://youtu.be/Cy06Ljee1JQ
In-context learning has recently been linked to implicit gradient descent in linear self-attention models, suggesting that context can induce a forward-pass update. Retrieval-augmented generation (RAG) also relies on context, but retrieved documents are usually treated as static evidence rather than signals for adaptation. We study RAG as an in-context optimization process. First, we show that one linear self-attention layer can implement one gradient-descent step on a unified linearized RAG objective covering both projection-based and dot-product retrieval interfaces. This gives an exact regime where retrieval-augmented prediction and in-context optimization coincide. We use this result not as a literal model of LLM computation, but as a guide for adapting the interaction between queries and retrieved evidence. We then test the boundary of this correspondence: it remains stable under controlled linear extensions, but becomes feature-distribution dependent under nonlinear architectures. Finally, we turn this view into a lightweight method for frozen RAG LLMs. The method keeps the retriever and backbone fixed, and predicts a context-conditioned update to a generator-side evidence-use interface. Across seven QA benchmarks, two retrievers, and two frozen LLM backbones, this forward-only update improves a shared-interface baseline, transfers to held-out tasks, and approaches test-time gradient adaptation at much lower per-query cost.
We consider estimating the one-step-ahead conditional distribution of a multivariate stochastic process. Many existing approaches rely on structural assumptions such as finite-range memory, sparsity, or additivity, which can be poorly suited to processes with long-range nonlinear interactions. However, without such structural assumptions, nonparametric estimation is known to be challenging due to the curse of dimensionality. To address this challenge, we introduce a new estimator based on the idea of learning representations of the \emph{predictive states} of a stochastic process. We theoretically analyze several aspects of this problem. As a bonus, we establish convergence rate guarantees for an instantiation of our method based on deep neural network estimators. Our results suggest that faster convergence rates are achievable when the predictive state space is lower-dimensional. That is, when the past history can be compressed into a lower-dimensional statistic that is sufficient for predicting the future. We implement our method in PyTorch and evaluate its performance on synthetic data and real \emph{C. elegans} data.
Retrieval is increasingly moving from one-shot matching toward interactive reasoning, where language agents iteratively inspect evidence, reformulate queries, and search again. Training such agents raises a credit-assignment challenge: executable actions such as queries or summaries can be directly evaluated by the retriever, while latent reasoning steps are not directly observable and only affect future executable actions. This asymmetry makes outcome-level reward assignment unreliable, as the same final reward may credit reasoning steps that did not actually shape retrieval success. We propose RICE-PO, a critic-free policy optimization framework that converts retrieval interactions into localized learning signals. RICE-PO selects high-uncertainty executable actions as anchors, evaluates local counterfactual branches using retrieval metrics, and propagates credit to latent reasoning steps only when reasoning-to-action influence is strong and future residual effects are stable. On BRIGHT and BEIR, RICE-PO consistently outperforms prompt-based agents and group-based RL baselines under the same retriever setting. These results show that the structure of agent-environment interaction itself can provide useful supervision for training reasoning-based retrieval agents.
Networks that predict angles, gaze directions, and rotations do not output these quantities directly: they emit raw coordinates that a fixed decoder, such as normalization, maps to the final prediction. Attribution methods are usually applied to the raw coordinates, so they can credit inputs for changes the decoder discards, such as rescaling a vector when only its direction matters. The resulting explanation can highlight inputs that have no effect on the prediction. We instead attribute through the decoder differential, which for normalization decoders reduces to a scaled tangent projection. We characterize invariance to output rescaling, show that every fixed smooth scalar target on a sphere has points of zero tangential sensitivity, and show that an endpoint-projected variant of Integrated Gradients can still attribute to scale-only inputs, whereas direct angular Integrated Gradients cannot. On synthetic tasks with analytically known relevance, ambient attribution assigns up to 72% of its mass to an irrelevant region. On MPIIGaze validation data, tangent Grad-CAM produces 44% greater prediction displacement under 10% pixel deletion. On PASCAL3D+ with a 6D rotation decoder, the full differential improves mean deletion displacement by 14.1 degrees over the ambient coordinate sum and by 13.8 degrees over a decoder-invariant scalar. Results for a two-polarity tangent LayerCAM variant are mixed: it exceeds ambient-sum LayerCAM on circular regression but no difference is detected on quaternion pose. These results separate two failure sources, radial contamination and scalarization loss. The approach uses existing attribution methods with decoder-aware targets, although some variants also require changes to attribution aggregation. An invariant attribution can still miss inputs that drive the prediction, so invariance and deletion performance should each be checked.
MAGIC-SSCIL: Manifold Anchoring and Geometric Incremental Calibration for Semi-Supervised Class Incremental Learning
Yousef Abdi, Mohammad Asadpour, Yousef Seyfari
AI 速读 基于摘要
MAGIC 通过在图传播伪标签并加权类别统计,提升无样本半监督类增量学习在极少标签下的性能。
cs.CV · cs.LG
查看原文摘要
Exemplar-free semi-supervised class-incremental learning (SSCIL) asks a model to learn new classes from a few labeled and many unlabeled images without storing past data. A common recipe replays old classes from stored Gaussian statistics and pseudo-labels unlabeled data with a confidence threshold. With one or two labels per class, however, the model's confidence is a poor guide to which pseudo-labels are right, so the statistics that replay depends on rest on few or noisy samples. We propose MAGIC (Manifold Anchoring and Geometric Incremental Calibration), which propagates labels over a k-nearest-neighbor graph of frozen features and uses the propagated confidences as weights in the unlabeled loss and, through Soft-Weighted Geometry Calibration (SWGC), in the class statistics from which phantom features replay old classes. With a frozen ResNet-18 on CIFAR-100, CUB-200, and ImageNet-R at 1%, 5%, and 10% labels, we compare MAGIC with nine supervised class-incremental methods equipped with FixMatch and two native SSCIL methods. MAGIC gains from unlabeled data in all nine settings, most at 1%, and leads on fine-grained CUB-200 at every ratio; at 1% it is level with the best on ImageNet-R and 2.2 points behind the native SSCIL method TACLE on CIFAR-100. Experiments inside MAGIC show that most of this gain flows through the weighted statistics, not the loss, that at 1% labels graph pseudo-labels are more accurate than the model's own at equal coverage, and that a fixed threshold of 0.95 removes most of the gain. At higher ratios, baselines that adapt the backbone or model full covariances overtake MAGIC on CIFAR-100 and ImageNet-R.
Recent advances in text-to-image generation have substantially improved image quality, particularly through diffusion-based methods that require multiple inference steps. In contrast, Sphere Encoder offers an efficient generative framework capable of producing images with fewer forward steps. However, extending it to free-form text is nontrivial, because its class-conditioned design lacks cross-modal alignment. To address this challenge, we propose TextSphere, an efficient text-to-image generative framework that extends Sphere Encoder with language conditioning and cross-modal alignment. We incorporate pretrained CLIP text embeddings through adaptive layer normalization (AdaLN) to guide image generation while retaining the original spherical sampling mechanism. To bridge the cross-modal representation gap, we introduce a Spherical Semantic Alignment and Uniformity (SSAU) objective that jointly promotes semantic alignment and representation uniformity on the hypersphere. Experiments on CUB-200-2011, MS-COCO-2014, and Flickr30k demonstrate that TextSphere consistently reduces generation FID (gFID) over the text-conditioned baseline while achieving reconstruction FID (rFID) below 1.0. With only 4 and 6 sampling steps, TextSphere achieves lower gFID than the compared diffusion- and GAN-based models trained from scratch on CUB-200-2011 and MS-COCO-2014, respectively. We further observe that improved reconstruction fidelity does not necessarily yield better generation quality. In addition, we use Partial Least Squares Singular Value Decomposition (PLS-SVD) to analyze cross-modal alignment and modality gap, offering insight into the effectiveness of our approach.
Inference-time scaling improves the performance of large language models by sampling multiple responses and evaluating their collective utility. However, many existing approaches obtain these samples by repeatedly sampling from a single policy. In this work, we study if the sample budget at inference time can instead be allocated across a population of policies. We introduce stratified policies, where each of $k$ learned policies contributes one response to a $k$-sample inference-time objective. We realize this policy population using separate LoRA adapters on a shared language model backbone and optimize it for pass@$k$. We theoretically motivate the approach by analyzing the linear setting; a population of LoRAs can achieve zero best-of-$k$ approximation error in settings where every single LoRA has strictly positive prediction error. Empirically, we demonstrate that stratified policies consistently improve pass@$k$ over a single-policy baseline across multiple reasoning tasks. Overall, our results highlight how distributing samples across complementary policies can be a more effective use of sample budget than repeated i.i.d. sampling from a single policy at inference time.
Video question answering under tight frame budgets requires preserving localized motion and event order, yet current video multimodal large language models (MLLMs) reduce comprehension to static multi-image reasoning by discarding unselected frames and encoding keyframes independently: (1) Spatially, whole-frame pooling dilutes small moving targets into static backgrounds; (2) Temporally, dropping intermediate frames or averaging tokens erases inter-frame motion, confusing direction, speed, and order. We propose VIVID (Visual Inter-frame Variations and Individual Dynamics), which moves beyond static snapshots like human vision: keyframes capture core states while inter-frame Patch Diffs preserve neighboring changes within $K_f$ keyframes. Autonomous Delta-Driven Keyframe Planning (ADKP) selects anchors from a $3\text{ fps}$ pool and feeds surrounding patch differences into Dual-Road Change Memory (DRCM), forming Dynamic Tokens at zero extra anchor tokens alongside a camera-stabilized Active Motion Region of Interest (ROI) and chronological Entity State Chains. Adding $0.22\%$ trainable parameters to a frozen $27\text{B}$ backbone, VIVID-Small ($50$ examples, $\sim 1.5$ GPU hours) beats all eleven baselines in overall accuracy across $8,083$ questions and trails only VIVID-Large ($3,600$ examples), which leads on MotionBench, Video-MME, and TempCompass.
Answer continuation often returns a diagnosis already stated in unfinished clinical reasoning. We study this behavior in MamaReason, adapted from dense Gemma 4 31B, on MamaBench's 217 paired maternal and child-health presentations. An explicit-commitment extractor identifies a diagnosis in 75 of 94 failed RL first passes and 75 of 93 failed SFT first passes. The continuation matches that diagnosis in 67/75 RL cases and 58/75 SFT cases, using diagnosis strings alone. Reading these commitments without additional generation recovers 30 of the 40 RL pairs restored by continuation under the archived correctness labels. Recurrence controls distinguish routine mentions from repeated answer selection. An offline second-commitment trigger measures the tradeoff between earlier readout and paired accuracy. The resulting decomposition separates diagnostic content already present in a trace, completion of an accepted answer, and paired diagnostic correctness. For clinical language-model evaluation, these are distinct outcomes that become visible when the same reasoning trace is examined through multiple readouts.
The agents in a heterogeneous team can cooperate to accomplish tasks that exceed their individual capabilities. The team trajectories for a cooperative task can be generated by a generative model, provided that the model has been trained on demonstrations of the agents cooperating. Yet, demonstrations of cooperation are costly and may not exist for tasks that the team has never performed. The conditions under which the agents can combine their capabilities, however, are physical properties of the team and are known before any demonstration is collected. We encode the conditions as constraints on the team trajectories, and we enforce the constraints while sampling from generative models trained on demonstrations of each agent alone. We prove that the proposed method can generate every team trajectory that satisfies the constraints, although the demonstrations contain no cooperation. Hence the constraints determine which team trajectories can be generated, and the demonstrations determine how likely each team trajectory is. Numerical results on two cooperative tasks show the effectiveness of the method, which generates cooperative team trajectories that satisfy the constraints and requires no retraining when agents join or leave the team. Finally, we implement the method on a multi-robot testbed.
Software is increasingly written by AI agents that produce an implementation from a natural language specification, a practice known as agentic coding. Code is now produced faster than it can be reviewed, and has been shown to exhibit its own characteristic failure patterns. Property-based testing (PBT), a long-standing technique in which invariant properties of a program are checked against many generated inputs, has recently been revisited with LLM agents authoring the properties, thereby removing the high-effort derivation of invariants that has long hindered its adoption. Existing methods, however, derive these properties from the implementation alone (Code-driven PBT). At the same time, a thorough specification such as a product requirements document has become routine within agentic coding, yet its role in testing the resulting code remains underexplored. We study the value of the specification as a source of properties for testing agent-written code. Our key insight is that the specification, unlike the implementation, records what the code was meant to do, and invariants derived from it are not biased by particular implementation choices. We propose Spec-Driven PBT, an agent that first expresses invariants in natural language from the specification and only then inspects the implementation to translate them into concrete tests. Across 25 applications from ViBench, Spec-Driven PBT finds a novel set of bugs that Code-driven PBT misses, while writing 2.4x fewer property tests. Though the two approaches surface a roughly equivalent number of bugs, there is little overlap: nearly 60% of the 169 total bugs are found by only one. Qualitatively, we find that the bugs unique to Spec-Driven PBT tend to violate documented product workflows whereas those to Code-driven PBT instead concentrate on implementation robustness, offering complementary value for bug discovery in agent-written code.
In graph-level learning, graph neural networks are typically applied with a fixed computational template: the same depth, local message-passing operator, and global attention mechanism are used for every input graph. However, different graphs within the same dataset can vary substantially in prediction difficulty and may require different mixtures of local and non-local computation. We study this phenomenon as graph-level heterogeneity in both classification and regression tasks. To address it, we propose SiGMA (Sigmoid Gated Message-Passing and Attention), a gated hybrid graph neural network architecture that combines message-passing and attention heads within each layer. Inspired by the growing use of gated attention and gated computation in large language models and vision transformers, each head in SiGMA is modulated by a learned node-wise sigmoid gate, allowing the model to adaptively regulate which local or global mechanisms are active for different nodes, layers, and input graphs. The resulting architecture can emulate expressive message-passing backbones, admits near-identity computation paths that are not forced to over-smooth, and enables direct non-local communication beyond message-passing neighborhoods to mitigate over-squashing. Empirically, SiGMA achieves strong performance across graph-level benchmarks, consistently matching or improving over competitive graph transformer and hybrid baselines. Our results suggest that adapting computation to graph-level heterogeneity is a promising direction for building more flexible and effective graph learning architectures.
$π^4$ is a unified foundation model for 4D scene perception. From diverse, uncalibrated RGB observations in the wild---a single image, an unordered collection, or a video sequence---it recovers dense geometry, camera poses, and metric scale, together with 2D/3D point tracks and moving-object masks for multi-image inputs. It couples a geometry branch with a geometry-conditioned motion branch, which predicts image-space correspondences that are lifted into 3D using reconstructed target-view geometry. Across diverse benchmarks, $π^4$ achieves the best overall average rank in geometry evaluation, alongside competitive performance in single- and multiview 2D/3D point tracking and moving-object segmentation. With a shared backbone, the full geometry-and-motion system contains approximately 0.5B parameters---less than half the size of most baselines. It produces joint reconstructions and tracks in one forward pass, including large-scale scenes with thousands of input images despite training with contexts of at most 64 views. We further present the data and training recipe underlying these capabilities, together with empirical studies of key design choices. Code and models are available at https://github.com/yyfz/Pi4.
Robot manipulation videos provide a rich visual experience, yet recovering robot action and gripper-state annotations from RGB observations still remains challenging. Using the embodiment-agnostic tool-center-point (TCP) pose as the action space, we present TCP4D, a unified model in which a shared 4D scene representation both tracks observed TCP motion and generates future TCP trajectories. Given a monocular video and initial 2D TCP queries, the tracker directly recovers metric TCP poses in SE(3) and gripper states alongside scene geometry and camera parameters. The generator samples local features from the same 4D representation along the recovered TCP tracks, then combines them with robot observations and language instructions to predict future TCP states. Our new cross-embodiment augmentation pipeline converts existing RGB videos into demonstrations for alternative embodiments by retargeting the recovered motion and synthesizing the corresponding observations, without recorded source-robot states. Experiments show that the tracker achieves absolute trajectory errors of 7.60 and 16.07 mm on DROID and RoboLab and adapts to a new embodiment with only ten demonstrations. The generator achieves 69.19% mean success on RoboTwin and outperforms well-established policy models on two real-world tasks, while joint training on source and augmented demonstrations improves target-robot success by 41.5% over augmented-only training.
The "AI revolution" has enabled teams across the technology industry to deliver software features at unprecedented velocity. While code evolves faster than ever, customer-facing documentation often struggles to stay up to speed. As users increasingly rely on LLM-based agents that consume documentation as their primary interface, the impact of this documentation gap worsens: stale information propagates and is treated as authoritative. Signals for identifying where documentation has fallen behind already exist in abundance, both explicit (e.g., resolved support tickets revealing what documentation lacks) and implicit (e.g., code-documentation discrepancies detectable from automated loops). We deploy a multi-agent system, built around a triage agent we call "TriAgent," that continuously extracts these signals and repairs documentation drift. Over 43 days of shadow-mode operation, our system triaged 499 tickets and surfaced 92 gaps in the documentation. Using a panel of LLM judges from independent model families, 81.1% of TriAgent's verifiable assertions were correct (245 of the 302 assertions the human resolution addresses; majority of three judges); on length-comparable pairs the panel preferred TriAgent's response in 70.5% of pairs and the human engineer's in 23.2% (the rest had no judge majority). Importantly, agent- and human-sourced replies tended to be complementary, with the human often surfacing insights that were either buried within the existing documentation or missing from it entirely; our system turns each such insight into a documentation gap and, after human review, a fix. Repairing a gap measurably improved the agent's later answer: when the agent re-answered the same ticket after the repair, blinded judges preferred the post-repair answer over its own pre-repair answer 71.4% of the time (n=35). A blinded human audit confirms strong alignment on self-improvement.
In the first version of this paper I reported that conservative Reddit users began writing more alike after ChatGPT was released while progressive users did not. That result came from a similarity measure that rises and falls with comment length. I redo every analysis of the original paper on all 6,056,593 comments with a sentence encoder and a dispersion measure computed within length strata, and the conservative break goes away. The two camps show no reliable difference in how within-group similarity changed after late 2022, and the ChatGPT date behaves like any of the more than two thousand other cutoffs I test. The matched flair comparison, the cumulative AI index, the stayer panel and the entropy signature each lose their support once length is held fixed. The one change that survives is older. Left and right comments drew closer to each other between 2019 and 2022, and that convergence was complete before ChatGPT arrived. I end with what this implies for studies that measure homogenization with embeddings.
Visual foundation models are central to physical intelligence, yet their development remains constrained by fragmented training datasets, inconsistent annotation quality, and limited multimodal supervision. To address these challenges, we introduce GameRealm, a large-scale, richly annotated dataset that grounds visual observations in synchronized, game-native supervision. GameRealm encompasses over 120 million RGB frames and 1,000 hours of gameplay footage across all views. Beyond its unprecedented scale, GameRealm provides rich multimodal supervision, including metric depth, camera parameters, synchronized audio, control inputs, and 3D point trajectories. Its diversity further extends to flexible view configurations, spanning single-view, stereo, multi-view, and panoramic recordings. To enable scalable acquisition across diverse game environments, we develop a unified data collection pipeline that combines human-operated interaction with agentic exploration. Extensive experiments across diverse visual learning tasks demonstrate that GameRealm consistently strengthens visual foundation models in both visual generation and geometric learning. These results establish GameRealm as a versatile and valuable data resource, unlocking new opportunities for advancing visual intelligence.
Diffusion models generate realistic visual content, yet often fail to produce rare but plausible compositions. When prompted with combinations that are valid but underrepresented in training data (such as a snowy beach or a rainbow at night), the generation process frequently collapses toward more common alternatives. We identify this failure mode as default completion bias, where denoising trajectories are implicitly attracted toward high-frequency semantic configurations. Existing guidance mechanisms do not explicitly model this competing tendency and therefore struggle to prevent such collapse. We introduce Default Completion Repulsion (DCR), a training-free framework that explicitly models and suppresses default completion behavior. DCR constructs a counterfactual attractor by relaxing the rare compositional factor while preserving surrounding semantics, inducing an alternative denoising trajectory reflecting the model's preferred completion. We define the discrepancy between target and attractor trajectories as a counterfactual drift, and propose a projection-based repulsion mechanism that removes guidance components aligned with this drift direction. This suppresses undesired frequent completions while preserving other semantic components. DCR operates entirely within the standard diffusion sampling process without retraining or architectural modification. Experiments on rare compositional prompts show that DCR improves compositional fidelity while maintaining visual quality. Our analysis further shows that generations from the attractor prompt alone closely match the baseline outputs, confirming that it captures the model's default completion.