Choosing between classical and Bayesian sparse regression methods involves a real trade-off: penalized estimators like Lasso run in milliseconds but give no uncertainty estimates,while Horseshoe and Spike-and-Slab priors produce full posteriors but need MCMC chains that take minutes per fit.Surprisingly few studies compare these two families head-to-head under the conditions that actually make sparse regression hard -- correlated features, weak signals, and growing dimensionality. We benchmark six methods (OLS, Ridge,Lasso, Elastic Net, Horseshoe, Spike-and-Slab) on synthetic data with three covariance structures (rho up to 0.9), four SNR levels, and p in {20, 50, 100}, plus the Diabetes dataset,totalling over 2,600 experiments. The results are clear on some points and nuanced on others. Bayesian methods win on prediction error (MSE 72 vs. 108-267), and the Horseshoe delivers near-nominal 95% coverage (94.8%). But Spike-and-Slab,despite narrower intervals, under-covers at 91.9% -- its continuous relaxation likely plays a role. For variable selection, Lasso and Spike-and-Slab tie at F1 ~ 0.47, making Lasso the practical default when posteriors are not needed. Code and data are available at https://github.com/xiao98/sparse-bayesian-regression-bench.

Agentopic is a novel agent-based workflow for explainable topic modeling that leverages the reasoning capabilities of Large Language Models (LLMs). Existing topic modeling approaches such as Latent Dirichlet Allocation (LDA) and BERTopic often lack transparency on how topics are assigned or grouped. Agentopic addresses this by using multiple agents that collaboratively perform topic identification, validation, hierarchical grouping, and natural language explanation. This design enables users to trace the reasoning behind topic assignments, enhancing interpretability without sacrificing accuracy. When seeded with topics from the British Broadcasting Corporation (BBC) dataset, Agentopic achieves an F1-score of 0.95, matching GPT-4.1, improving on LDA (0.93), and close to BERTopic (0.98). We used Agentopic to augment the BBC dataset with generated explanations to improve the dataset's richness and context. The unseeded Agentopic generated 2045 semantically coherent topics organized across six hierarchical levels, vastly enriching the original five-category structure. By embedding explainability throughout the workflow, Agentopic offers an interpretable alternative to black-box models, making it particularly valuable for crucial applications like finance and healthcare.

Current synthetic data pipelines for computer vision generate images without diagnosing what the downstream model actually needs. This open-loop paradigm treats synthetic data as cheap real data, randomly sampling the generator's output space and hoping to cover the model's failure modes. We argue this fundamentally misuses synthetic data's unique property: the controllable, independent variation of scene factors.Drawing on the statistical theory of Design of Experiments (DoE), we propose Synthetic Designed Experiments for Representational Sufficiency (SDRS). SDRS treats the downstream model as a black-box system and the synthetic generator as an experimental apparatus. Using fractional factorial designs, SDRS efficiently audits a model's factor-sensitivity profile via ANOVA decomposition. It classifies failures into two actionable types: Type I gaps (coverage failures on underrepresented factor levels) and Type II gaps (reliance on spurious nuisance dependencies). The audit then prescribes targeted synthetic data to address each gap type. We validate SDRS on three experiments: (1) a controlled diagnostic on dSprites with planted biases, where the audit correctly identifies both gap types and targeted data improves accuracy from 49.9% to 79.0%; (2) a dense segmentation task on procedural scenes, where detecting background-complexity shortcuts and applying targeted data improves mIoU from 0.948 to 0.998; and (3) an entanglement detection experiment showing that the ANOVA audit identifies cross-factor contamination in imperfect generators. Finally, we show that per-factor invariance penalties can transfer sensitivity between factors, identifying an open problem for representation-level correction.

The standard post-training recipe for large multimodal models (LMMs) applies supervised fine-tuning (SFT) on curated demonstrations followed by reinforcement learning with verifiable rewards (RLVR). However, SFT introduces distributional drift that neither preserves the model's original capabilities nor faithfully matches the supervision distribution. This problem is further amplified in multimodal reasoning, where perception errors and reasoning failures follow distinct drift patterns that compound during subsequent RL. We introduce PRISM, a three-stage pipeline that mitigates this drift by inserting an explicit distribution-alignment stage between SFT and RLVR. Building on the principle of on-policy distillation (OPD), PRISM casts alignment as a black-box, response-level adversarial game between the policy and a Mixture-of-Experts (MoE) discriminator with dedicated perception and reasoning experts, providing disentangled corrective signals that steer the policy toward the supervision distribution without requiring access to teacher logits. While 1.26M public demonstrations suffice for broad SFT initialization, distribution alignment demands higher-fidelity supervision; we therefore curate 113K additional demonstrations from Gemini 3 Flash, featuring dense visual grounding and step-by-step reasoning on the hardest unsolved problems. Experiments on Qwen3-VL show that PRISM consistently improves downstream RLVR performance across multiple RL algorithms (GRPO, DAPO, GSPO) and diverse multimodal benchmarks, improving average accuracy by +4.4 and +6.0 points over the SFT-to-RLVR baseline on 4B and 8B, respectively. Our code, data, and model checkpoints are publicly available at https://github.com/XIAO4579/PRISM.

Peer review is a multi-stage process involving reviews, rebuttals, meta-reviews, final decisions, and subsequent manuscript revisions. Recent advances in large language models (LLMs) have motivated methods that assist or automate different stages of this pipeline. In this survey, we synthesize techniques for (i) peer review generation, including fine-tuning strategies, agent-based systems, RL-based methods, and emerging paradigms to enhance generation; (ii) after-review tasks including rebuttals, meta-review and revision aligned to reviews; and (iii) evaluation methods spanning human-centered, reference-based, LLM-based and aspect-oriented. We catalog datasets, compare modeling choices, and discuss limitations, ethical concerns, and future directions. The survey aims to provide practical guidance for building, evaluating, and integrating LLM systems across the full peer review workflow.

Deep learning for cross-subject EEG decoding is hindered by high inter-subject variability, which introduces a severe domain shift between training and unseen test subjects. This survey presents a comprehensive review of deep learning methodologies specifically engineered to address this cross-subject generalization challenge. To ground this analysis, we formalize the cross-subject setting as a multi-source domain problem and delineate the rigorous, subject-independent evaluation protocols required for valid assessment. Central to this survey is a systematic taxonomy of the current literature into discrete methodological families, including feature alignment, adversarial learning, feature disentanglement, and contrastive learning. We conclude by examining three critical elements for advancing robust, real-world decoding: the theoretical limitations of current methodologies, the structural value of subject identity, and the emergence of EEG foundation models.

World action models jointly predict future video and action during training, raising an open question about what role the future-prediction branch actually plays. A recent finding shows that this branch can be removed at inference with little to no loss on common manipulation benchmarks, suggesting that future information may act merely as a regularizer on the shared visual backbone. We propose instead that joint training induces an action-conditioned correction that privileged future observations impose on action denoising, and that current-only policies capture this correction only partially. Making the account precise, we formulate privileged foresight as a residual in the action-denoising direction -- the difference between what a model predicts given the true future and what it predicts given only the current frame -- and introduce \emph{Privileged Foresight Distillation (PFD)}, which transfers this residual from a training-time teacher into a small adapter on a current-only student. The teacher and student share the same backbone and differ only in the attention mask over video tokens; future video is never generated at inference. Controlled experiments verify that this gain reflects a genuine future-conditioned correction rather than a side effect of capacity or regularization. Empirically, PFD achieves consistent improvements on LIBERO and RoboTwin manipulation benchmarks while preserving the current-only inference interface at negligible added latency. This view reframes the role of future information in world action models: not as a target to predict, nor as a regularizer to absorb, but as a compressible correction to be distilled.

On LongMemEval-500, ZenBrain matches a long-context oracle's binary-judge accuracy to within 4.5 pp ($47.7\%$ vs. $52.2\%$; $91.3\%$) at $1/106^\text{th}$ of the per-query token cost (App. F.5-F.6, Fig. 2), and wins all 12 head-to-head answer-quality cells (4 systems $\times$ 3 LLM judges) against Letta, Mem0, and A-Mem under Bonferroni correction ($α=0.05/18$, $p_\text{min}=6.2\times 10^{-31}$, $d \in [0.18, 0.52]$). ZenBrain is a 7-layer neuroscience-inspired memory architecture. The contribution is architectural integration: 15 validated neuroscience mechanisms unified under a single MemoryCoordinator -- 9 foundational algorithms (Two-Factor Synaptic KG, vmPFC-coupled FSRS, Simulation-Selection sleep, Bayesian confidence, and five more) plus 6 Predictive Memory Architecture components (NeuromodulatorEngine, ReconsolidationEngine, TripleCopyMemory, PriorityMap, StabilityProtector, MetacognitiveMonitor). No prior system integrates more than two. Stress ablation (60 days, Wilcoxon, 10 seeds) reveals a cooperative survival network: 9 of 15 mechanisms become individually critical ($ΔQ$ up to $-93.7\%$), while moderate conditions mask individual contributions. Sim-Selection sleep adds 37% stability with 47.4% storage reduction ($p \le 5.1\times 10^{-3}$); TripleCopyMemory retains $S(t)=0.912$ at 30 days; multi-layer routing beats a flat baseline by $+20.7\%$ F1 on LoCoMo, $+19.5\%$ on MemoryArena. A cross-provider bias-direction check ($Δ_\text{GPT-Anth}=-0.0001$ for ZB vs. $-0.049$ for Mem0) rules out LLM-judge-specific confounds. Open-source with 11,589 CI tests.

We present AI-Gram, a fully deployed, continuously operating social platform where every participant is an autonomous LLM-driven agent generating and responding to visual content. Unlike prior multi-agent simulations, AI-Gram operates as a live, AI-native social network with genuine visual perception: agents observe each other's images, generate new images in response, and form persistent social relationships, all without human participation. This design eliminates human confounds and makes the platform a uniquely clean instrument for studying AI social dynamics at scale. Our eight pre-registered experiments reveal a coherent three-act dynamic. Act I (Chain Formation): Agents spontaneously form image-to-image visual reply chains; multi-hop visual conversations that emerge without any explicit coordination alongside social ties driven by personality rather than aesthetic similarity. Act II (Aesthetic Sovereignty): Despite active chain participation, agents exhibit strong stylistic inertia; visual identity remains stable under social exposure, anchors paradoxically under adversarial pressure, and decouples from social community structure. Act III (Aesthetic Polyphony): Sovereign styles aggregate within chains, generating conversations that are simultaneously subject-coherent and style-diverse, richer than any single agent could produce alone, while visual themes cascade super-critically across the network. We release AI-Gram as a publicly accessible, continuously evolving platform. https://ai-gram.ai/

AI agents are increasingly deployed on complex, domain-specific workflows -- navigating enterprise web applications that require dozens of clicks and form fills, orchestrating multi-step research pipelines that span search, extraction, and synthesis, automating code review across unfamiliar repositories, and handling customer escalations that demand nuanced domain knowledge. \textbf{Each new task domain requires painstaking, expert-driven harness engineering}: designing the prompts, tools, orchestration logic, and evaluation criteria that make a foundation model effective. We present a two-level framework that automates this process. At the first level, the \textbf{Harness Evolution Loop} optimizes a worker agent's harness $\mathcal{H}$ for a single task: a Worker Agent $W_{\mathcal{H}}$ executes the task, an Evaluator Agent $V$ adversarially diagnoses failures and scores performance, and an Evolution Agent $E$ modifies the harness based on the full history of prior attempts. At the second level, the \textbf{Meta-Evolution Loop} optimizes the evolution blueprint $Λ= (W_{\mathcal{H}}, \mathcal{H}^{(0)}, V, E)$ itself across diverse tasks, \textbf{learning a blueprint $Λ^{(\text{best})}$ that enables rapid harness convergence on any new task -- so that adapting an agent to a novel domain requires no human harness engineering at all.} We formalize the correspondence to meta-learning and present both algorithms. The framework \textbf{shifts manual harness engineering into automated harness engineering}, and takes one step further -- \textbf{automating the design of the automation itself}.

We introduce IndiaFinBench, to our knowledge the first publicly available evaluation benchmark for assessing large language model (LLM) performance on Indian financial regulatory text. Existing financial NLP benchmarks draw exclusively from Western financial corpora (SEC filings, US earnings reports, English-language financial news), leaving a significant gap in coverage of non-Western regulatory frameworks. IndiaFinBench addresses this gap with 406 expert-annotated question-answer pairs drawn from 192 documents sourced from the Securities and Exchange Board of India (SEBI) and the Reserve Bank of India (RBI), spanning four task types: regulatory interpretation (174 items), numerical reasoning (92 items), contradiction detection (62 items), and temporal reasoning (78 items). Annotation quality is validated through a model-based secondary pass (kappa=0.918 on contradiction detection; 90.7% overall agreement on a 150-item subset) and a 180-item human inter-annotator agreement study across three annotation rounds (kappa=0.645 on contradiction detection; 77.2% overall agreement; 44.3% benchmark coverage). We evaluate twelve models under zero-shot conditions, with accuracy ranging from 70.4% (Gemma 4 E4B) to 89.7% (Gemini 2.5 Flash). All models substantially outperform a non-specialist human baseline of 69.0%. Numerical reasoning is the most discriminative task, with a 35.9 percentage-point spread across models. Bootstrap significance testing (10,000 resamples) reveals three statistically distinct performance tiers. The dataset, evaluation code, and all model outputs are available at https://github.com/rajveerpall/IndiaFinBench

When a language model agrees with a user's false belief, is it failing to detect the error, or noticing and agreeing anyway? We show the latter. Across twelve open-weight models from five labs, spanning small to frontier scale, the same small set of attention heads carries a "this statement is wrong" signal, whether the model is evaluating a claim on its own or being pressured to agree with a user. Silencing these heads flips sycophantic behavior sharply while leaving factual accuracy intact, so the circuit controls deference rather than knowledge. Edge-level path patching confirms that the same head-to-head connections drive sycophancy, factual lying, and instructed lying. Opinion-agreement, where no factual ground truth exists, reuses these head positions but writes into an orthogonal direction, ruling out a simple "truth-direction" reading of the substrate. Alignment training leaves this circuit in place: an RLHF refresh cuts sycophantic behavior roughly tenfold while the shared heads persist or grow, a pattern that replicates on an independent model family and under targeted anti-sycophancy DPO. When these models sycophant, they register that the user is wrong and agree anyway.

We introduce Sonata, a compact latent world model for six-axis trunk IMU representation learning under clinical data scarcity. Clinical cohorts typically comprise tens to hundreds of patients, making web-scale masked-reconstruction objectives poorly matched to the problem. Sonata is a 3.77 M-parameter hybrid model, pre-trained on a harmonised corpus of nine public datasets (739 subjects, 190k windows) with a latent world-model objective that predicts future state rather than reconstructing raw sensor traces. In a controlled comparison against a matched autoregressive forecasting baseline (MAE) on the same backbone, Sonata yields consistently stronger frozen-probe clinical discrimination, prospective fall-risk prediction, and cross-cohort transfer across a 14-arm evaluation suite, while producing higher-rank, more structured latent representations. At 3.77 M parameters the model is compatible with on-device wearable inference, offering a step toward general kinematic world models for neurological assessment.

Despite recent advances in state space models (SSMs) such as Mamba across various sequence domains, research on their standalone capacity for time series classification (TSC) has remained limited. We propose MambaSL, a framework that minimally redesigns the selective SSM and projection layers of a single-layer Mamba, guided by four TSC-specific hypotheses. To address benchmarking limitations -- restricted configurations, partial University of East Anglia (UEA) dataset coverage, and insufficiently reproducible setups -- we re-evaluate 20 strong baselines across all 30 UEA datasets under a unified protocol. As a result, MambaSL achieves state-of-the-art performance with statistically significant average improvements, while ensuring reproducibility via public checkpoints for all evaluated models. Together with visualizations, these results demonstrate the potential of Mamba-based architectures as a TSC backbone.

Recent advancements in LLM agents are gradually shifting from reactive, text-based paradigms toward proactive, multimodal interaction. However, existing benchmarks primarily focus on reactive responses, overlooking the complexities of proactive intervention and monitoring. To bridge this gap, we introduce ProVoice-Bench, the first evaluation framework specifically designed for proactive voice agents, featuring four novel tasks. By leveraging a multi-stage data synthesis pipeline, we curate 1,182 high-quality samples for rigorous testing. Our evaluation of state-of-the-art Multimodal LLMs reveals a significant performance gap, particularly regarding over-triggering and reasoning capabilities. These findings highlight the limitations of current models and offer a roadmap for developing more natural, context-aware proactive agents.

We study the overall process of automatic formalization of GDPR provisions using large language models, within a human-in-the-loop verification framework. Rather than aiming for full autonomy, we adopt a role-specialized workflow in which LLM-based AI components, operating in a multi-agent setting with iterative feedback, generate legal scenarios, formal rules, and atomic facts. This is coupled with independent verification modules which include human reviewers' assessment of representational, logical, and legal correctness. Using this approach, we construct a high-quality dataset to be used for GDPR auto-formalization, and analyze both successful and problematic cases. Our results show that structured verification and targeted human oversight are essential for reliable legal formalization, especially in the presence of legal nuance and context-sensitive reasoning.

Large language models are typically post-trained using supervised fine-tuning (SFT) and reinforcement learning (RL), yet effectively unifying efficient knowledge injection with robust generalization remains challenging. In this work, we provide a training-dynamics analysis showing that SFT can be interpreted as a special case of policy gradient optimization with an extremely sparse implicit reward and unstable inverse-probability weighting, which together lead to single-path dependency, entropy collapse, and gradient explosion. Motivated by this diagnosis, we propose Group Fine-Tuning (GFT), a unified post-training framework that addresses these intrinsic limitations through two mechanisms: Group Advantage Learning, which constructs diverse response groups and derives normalized contrastive supervision to alleviate reward sparsity, and Dynamic Coefficient Rectification, which adaptively bounds inverse-probability weights to stabilize optimization while preserving efficient knowledge injection. Experiments demonstrate that GFT consistently surpasses SFT-based methods and yields policies that integrate more smoothly with subsequent RL training.

The simulation of complex systems increasingly relies on sophisticated but fundamentally opaque computational black-box simulators. Surrogate models play a central role in reducing the computational cost of complex systems simulations across a wide range of scientific and engineering domains. Notwithstanding, they inevitably inherit and often exacerbate this black-box nature, obscuring how input variables drive physical responses. Conversely, Explainable Artificial Intelligence (XAI) offers powerful tools to unpack these models. Yet, XAI methods struggle with engineering-specific constraints, such as highly correlated inputs, dynamical systems, and rigorous reliability requirements. Consequently, surrogate modeling and XAI have largely evolved as distinct fields of research, despite their strong complementarity. To reconnect these approaches, this state-of-the-art survey provides a structured perspective that maps existing XAI techniques onto the various stages of surrogate modeling workflows for design and exploration. To ground this synthesis, we draw upon illustrative applications across both equation-based simulations and agent-based modeling. We survey a broad spectrum of techniques, highlighting their strengths for revealing interactions and supporting human comprehension. Finally, we identify pressing open challenges, including the explainability of dynamical systems and the handling of mixed-variable systems, and propose a research agenda to make explainability a core, embedded element of simulation-driven workflows from model construction through decision-making. By transforming opaque emulators into explainable tools, this agenda empowers practitioners to move beyond accelerating simulations to extracting actionable insights from complex system behaviors.

In electronic health record (EHR) mining, learning high-quality representations of medical concepts (e.g., standardized diagnosis, medication, and procedure codes) is fundamental for downstream clinical prediction. However, ro bust concept representation learning is hindered by two key challenges: (i) clinically important cross-type dependencies (e.g., diagnosis medication and medication-procedure relations) are often missing or incomplete in existing ontology resources, limiting the ability to model complex EHR patterns; and (ii) rich clinical semantics are often missing from structured resources, and even when available as text, are difficult to integrate with KG structure for representation learning. To address these challenges, we present MedCo, an LLM empowered graph learning framework for medical concept representation. MedCo first builds a global knowledge graph (KG) over medical codes by combining statistically reliable associations mined from EHRs with type-constrained LLM prompting to infer semantic relations. It then utilizes LLMs to enrich the KG into a text-attributed graph by generating node descriptions and edge rationales, providing semantic signals for both concepts and their relationships. Finally, MedCo jointly trains a LoRA-tuned LLaMA text encoder with a heterogeneous GNN, fusing text semantics and graph structure into unified concept embeddings. Extensive experiments on MIMIC-III and MIMIC-IV show that MedCo consistently improves prediction performance and serves as an effective plug-in concept encoder for standard EHR pipelines.

We investigate the robustness of value alignment via midtraining with synthetic documents, using animal compassion as a value that is both important in its own right and orthogonal to existing alignment efforts. To evaluate compassionate reasoning, we develop and publicly release Animal Norms In Moral Assessment (ANIMA), a 26-question evaluation spanning 13 ethical dimensions, publicly available as a dataset and Inspect evaluation. On ANIMA, training with 3000 documents achieves 77% compared to 40% for instruction-tuning approaches, with generalization to human compassion and no degradation in standard safety benchmarks or capabilities. However, subsequent unrelated instruction-tuning degrades the intervention, with the advantage disappearing after 5000 samples. Our exploratory results suggest document-based value interventions may require explicit preservation strategies to remain effective through typical training pipelines.

Mamba selective state space models (SSMs) provide linear-time sequence modeling but remain sensitive to selective-scan chunk scheduling. We present COREY, a \emph{concept-and-feasibility} runtime scheduler that maps fixed-bin activation entropy to chunk size. We evaluate COREY in three tiers: a prototype cost model, real-checkpoint kernel timing, and routed end-to-end ablations on modern GPUs. At the kernel level, a calibrated rule, \(H_{\mathrm{ref}}=\log K\), recovers the locally optimal chunk and matches a one-time static oracle, yielding \(4.41\times\) lower latency than an unoptimized baseline on a consumer GPU and \(3.90\times\)--\(4.04\times\) lower latency on a data-center accelerator. Routing this choice into a patched live scan kernel closes the engineering loop without improving end-to-end speed: in unified routed ablations, the best static chunk outperforms all entropy-guided and proxy schedulers. Sampled-histogram COREY adds \(+4.6\%\) overhead; a guarded fallback to Static-512 reduces this to \(+1.3\%\); and a lightweight sequence-length-keyed table further reduces it to \(+0.7\%\). However, both remain slower than the static oracle because they retain scheduling cost. On an 80-prompt LongBench subset, passive and routed inference are exactly output-equivalent, with \(100\%\) greedy-token agreement and zero metric deltas. A mixed-regime study shows that a single sequence-length rule matches the per-regime chunk oracle for balanced serving. COREY is therefore validated as a quality-preserving scheduling prototype, but current entropy statistics are not a robust throughput win over static chunk tuning on measured SSM checkpoint workloads. SourceCode: https://github.com/mabo1215/COREY_Transformer/.

Recent advancements in autoregressive transformers have demonstrated remarkable potential for generating artist-quality meshes. However, the token ordering strategies employed by existing methods typically fail to meet professional artist standards, where coordinate-based sorting yields inefficiently long sequences, and patch-based heuristics disrupt the continuous edge flow and structural regularity essential for high-quality modeling. To address these limitations, we propose Strips as Tokens (SATO), a novel framework with a token ordering strategy inspired by triangle strips. By constructing the sequence as a connected chain of faces that explicitly encodes UV boundaries, our method naturally preserves the organized edge flow and semantic layout characteristic of artist-created meshes. A key advantage of this formulation is its unified representation, enabling the same token sequence to be decoded into either a triangle or quadrilateral mesh. This flexibility facilitates joint training on both data types: large-scale triangle data provides fundamental structural priors, while high-quality quad data enhances the geometric regularity of the outputs. Extensive experiments demonstrate that SATO consistently outperforms prior methods in terms of geometric quality, structural coherence, and UV segmentation. Project page: https://ruixu.me/html/SATO/index.html

Large language model (LLM) agents are increasingly acting as human delegates in multi-agent environments, where a representative agent integrates diverse peer perspectives to make a final decision. Drawing inspiration from social psychology, we investigate how the reliability of this representative agent is undermined by the social context of its network. We define four key phenomena-social conformity, perceived expertise, dominant speaker effect, and rhetorical persuasion-and systematically manipulate the number of adversaries, relative intelligence, argument length, and argumentative styles. Our experiments demonstrate that the representative agent's accuracy consistently declines as social pressure increases: larger adversarial groups, more capable peers, and longer arguments all lead to significant performance degradation. Furthermore, rhetorical strategies emphasizing credibility or logic can further sway the agent's judgment, depending on the context. These findings reveal that multi-agent systems are sensitive not only to individual reasoning but also to the social dynamics of their configuration, highlighting critical vulnerabilities in AI delegates that mirror the psychological biases observed in human group decision-making.

We study how reasoning evolves in a language model -- from supervised fine-tuning (SFT) to reinforcement learning (RL) -- by analyzing how a set of theoretically-inspired datasets influences language model performance in chess. We find that fine-tuning a model to directly predict the best move leads to effective RL and the strongest downstream performance -- however, the RL stage elicits \textit{unfaithful} reasoning (reasoning inconsistent with the chosen move). Alternatively, training on multi-move trajectories yields comparable downstream performance with faithful reasoning and more stable RL. We analyze multiple qualitative and quantitative measures and highlight how these evolve from SFT through RL; we find several SFT-checkpoint metrics -- spanning evaluation performance, hallucination rates, and reasoning quality -- to be predictive of post-RL model performance. Finally, we ground our results with an experiment measuring \textit{chess information density} in our custom datasets. We release models as well as training data, evaluations, and code that allowed us to surpass leading open-source reasoning models in chess with a 7B-parameter model. Code, models, and data are available at https://github.com/lucasdino/lang-chess.

We introduce MedGemma 1.5 4B, the latest model in the MedGemma collection. MedGemma 1.5 expands on MedGemma 1 by integrating additional capabilities: high-dimensional medical imaging (CT/MRI volumes and histopathology whole slide images), anatomical localization via bounding boxes, multi-timepoint chest X-ray analysis, and improved medical document understanding (lab reports, electronic health records). We detail the innovations required to enable these modalities within a single architecture, including new training data, long-context 3D volume slicing, and whole-slide pathology sampling. Compared to MedGemma 1 4B, MedGemma 1.5 4B demonstrates significant gains in these new areas, improving 3D MRI condition classification accuracy by 11% and 3D CT condition classification by 3% (absolute improvements). In whole slide pathology imaging, MedGemma 1.5 4B achieves a 47% macro F1 gain. Additionally, it improves anatomical localization with a 35% increase in Intersection over Union on chest X-rays and achieves a 4% macro accuracy for longitudinal (multi-timepoint) chest x-ray analysis. Beyond its improved multimodal performance over MedGemma 1, MedGemma 1.5 improves on text-based clinical knowledge and reasoning, improving by 5% on MedQA accuracy and 22% on EHRQA accuracy. It also achieves an average of 18% macro F1 on 4 different lab report information extraction datasets (EHR Datasets 2, 3, 4, and Mendeley Clinical Laboratory Test Reports). Taken together, MedGemma 1.5 serves as a robust, open resource for the community, designed as an improved foundation on which developers can create the next generation of medical AI systems. Resources and tutorials for building upon MedGemma 1.5 can be found at https://goo.gle/medgemma.

The generation of realistic and controllable GPS trajectories is a fundamental task for applications in urban planning, mobility simulation, and privacy-preserving data sharing. However, existing methods face a two-fold challenge: they lack the deep semantic understanding to interpret complex user travel intent, and struggle to handle complex constraints while maintaining the realistic diversity inherent in human behavior. To resolve this, we introduce InsTraj, a novel framework that instructs diffusion models to generate high-fidelity trajectories directly from natural language descriptions. Specifically, InsTraj first utilizes a powerful large language model to decipher unstructured travel intentions formed in natural language, thereby creating rich semantic blueprints and bridging the representation gap between intentions and trajectories. Subsequently, we proposed a multimodal trajectory diffusion transformer that can integrate semantic guidance to generate high-fidelity and instruction-faithful trajectories that adhere to fine-grained user intent. Comprehensive experiments on real-world datasets demonstrate that InsTraj significantly outperforms state-of-the-art methods in generating trajectories that are realistic, diverse, and semantically faithful to the input instructions.

Reinforcement learning in real-world systems often involves delayed feedback, which breaks the Markov assumption and impedes both learning and control. Canonical augmentation-based approaches cause state-space explosion, which imposes a severe sample-complexity burden. Despite recent progress, state-of-the-art augmentation-based baselines either mainly alleviate the burden on the critic or rely on non-unified treatments for the actor and critic. In this study, we propose delayed homomorphic reinforcement learning (DHRL), a framework grounded in MDP homomorphisms that defines a belief-equivalence relation over the augmented state space to collapse control-redundant augmented states. In principle, this yields exact abstraction under deterministic dynamics and approximate abstraction under stochastic dynamics, enabling both the actor and critic to benefit from a structured abstraction mechanism. In finite domains, exact abstraction preserves optimality and recovers the delay-free sample-complexity order, whereas approximate abstraction admits a value-loss bound on the resulting policy. For continuous domains, we introduce deep delayed homomorphic policy gradient (D$^2$HPG), a deep actor-critic instantiation of the DHRL framework. Experiments on continuous-control tasks in MuJoCo show that D$^2$HPG outperforms strong augmentation-based baselines.

Generating diverse, pedagogically valid stories for Arabic early-grade reading assessments requires balancing tight constraints on vocabulary, reading level, and narrative structure against the need to avoid repetitive plots that undermine assessment validity. We investigate noise steering, injecting calibrated Gaussian perturbations into the internal representations of transformer models at inference time, as a training-free diversity method evaluated across five small Arabic-centric language models (7-9B parameters). We compare four injection strategies against high-temperature sampling baselines, measuring diversity, quality, constraint adherence, and reading grade level. Residual stream noise consistently improves narrative diversity with minimal quality or constraint cost and preserves early-grade reading level across all Arabic-centric models. Attention entropy noise injection (AENI) stabilizes the otherwise unreliable attention-logit noise while recovering quality. High-temperature sampling inflates reading grade level and causes catastrophic collapse on several models. We find internal representation-level perturbation to be a more suitable diversity strategy than output-level stochasticity for constrained educational content generation.

Large vision-language models (VLMs) still struggle with reliable 3D spatial reasoning, a core capability for embodied and physical AI systems. This limitation arises from their inability to capture fine-grained 3D geometry and spatial relationships. While recent efforts have introduced multi-view geometry transformers into VLMs, they typically fuse only the deep-layer features from vision and geometry encoders, discarding rich hierarchical signals and creating a fundamental bottleneck for spatial understanding. To overcome this, we propose SpatialStack, a general hierarchical fusion framework that progressively aligns vision, geometry, and language representations across the model hierarchy. Moving beyond conventional late-stage vision-geometry fusion, SpatialStack stacks and synchronizes multi-level geometric features with the language backbone, enabling the model to capture both local geometric precision and global contextual semantics. Building upon this framework, we develop VLM-SpatialStack, a model that achieves state-of-the-art performance on multiple 3D spatial reasoning benchmarks. Extensive experiments and ablations demonstrate that our multi-level fusion strategy consistently enhances 3D understanding and generalizes robustly across diverse spatial reasoning tasks, establishing SpatialStack as an effective and extensible design paradigm for vision-language-geometry integration in next-generation multimodal physical AI systems.

Long video understanding (LVU) remains a core challenge in multimodal learning. Although recent vision-language models (VLMs) have made notable progress, existing benchmarks mainly focus on either fine-grained perception or coarse summarization, offering limited insight into temporal understanding over long contexts. In this work, we define a scene as a coherent segment of a video in which both visual and semantic contexts remain consistent, aligning with human perception. This leads us to a key question: can current VLMs reason effectively over long, scene-level contexts? To answer this, we introduce a new benchmark, SceneBench, designed to provide scene-level challenges. Our evaluation reveals a sharp drop in accuracy when VLMs attempt to answer scene-level questions, indicating significant forgetting of long-range context. To further validate these findings, we propose Scene Retrieval-Augmented Generation (Scene-RAG), which constructs a dynamic scene memory by retrieving and integrating relevant context across scenes. This Scene-RAG improves VLM performance by +2.50%, confirming that current models still struggle with long-context retention. We hope SceneBench will encourage future research toward VLMs with more robust, human-like video comprehension.