In briefZero-Shot Brain MRI Inpainting Using 2.5D Unconditional Flow Priors
SummaryThis paper proposes a zero-shot framework for inpainting brain MRI volumes using a 2.5D unconditional flow matching model trained on adjacent axial slice triplets while excluding pathological regions from the loss. At inference, the model processes triplets autoregressively along depth, constrains the prior with the input mask via the Restora-Flow solver, and averages an ensemble of stochastic reconstructions. On the BraTS 2026 Inpainting Challenge validation set, it reports an SSIM of 0.816, MSE of 0.007, and PSNR of 22.923.
Arnela Hadzic, Franz Thaler, Simon Johannes Joham et al.
In briefS2Tok: Streaming 3D Reconstruction with Persistent Spatial Tokens
SummaryS2Tok is a feed-forward framework that maintains a persistent, size-adaptive scene state from uncalibrated image streams. A spatially informed transformer merges each new observation with stored scene tokens, while a learned admission module selectively expands the representation to limit redundancy. A hierarchical decoder and Gaussian head turn this state into non-pixel-aligned 3D Gaussians for novel-view rendering without caching past frames. Experiments on four benchmarks report competitive streaming rendering quality with compact Gaussian representations.
Fang Li, Jiraphon Yenphraphai, Quentin Herau et al.
In briefSNR-Gated LSTM-Conditioned Diffusion Model for MIMO Channel Estimation
SummaryThis paper proposes a diffusion-based framework for MIMO channel estimation that denoises in the angular domain and uses an LSTM to encode a short sequence of observations for temporal information. A learnable SNR-gated shortcut feeds the network input into the final decoding stage to balance observation fidelity with learned priors. To reduce latency, it uses DDIM-style deterministic reverse updates with SNR-adaptive step allocation. Simulations on time-evolving standardized channel models report consistent gains over existing diffusion-based baselines while keeping inference latency low.
In briefBest optimizer depends on batch size: a paper showing optimizer rankings shift across batch sizes
SummaryThe paper argues that new adaptive optimizers are usually benchmarked at a single batch size, yet the best optimizer can change as batch size changes, even after extensive tuning. Using a noisy quadratic model, the authors show that the batch-size scaling rule preserving performance depends on direction, and that gradient noise creates a bias-variance tradeoff in preconditioning that can reverse rankings. In language model pretraining, keeping small-batch updates only in the sharpest 0.001% of directions recovers up to 59.5% of the loss gap. The authors conclude that optimizer design should account for batch size and benchmarks should span multiple batch sizes.
WhyShows direction-dependent batch scaling in a noisy quadratic model, proves a bias-variance tradeoff, and validates it on language model pretraining, impacting optimizer benchmarking.
In briefWieszcz-XIX: A 3.1-billion-word corpus of pre-1918 Polish and language models trained on it
SummaryThe paper releases Wieszcz-XIX, a corpus of about 3.1 billion words of Polish published from 1800 to 1918, mostly periodicals, built with filtering, deduplication and removal of post-1918 text. The authors train decoder-only models of 47M to 349M parameters from scratch and measure how well they handle period versus post-1918 vocabulary against two modern Polish models. They also report that the models reproduce period prejudice, including antisemitic statements.
WhyBuilds a 3.1-billion-word pre-1918 Polish corpus and trains period-bounded models compared against modern baselines, with substantive evidence and clear value for historical NLP.
In briefHULK: Learning whole-body forceful movement and manipulation for humanoid robots
SummaryThis paper presents HULK, a whole-body control framework that lets humanoid robots walk while holding large, heavy objects against their bodies and exerting force. It uses model predictive control to guide reinforcement learning, trains two teacher policies, and distills them into one policy. Evaluated in simulation and on the Unitree G1, the authors report the barrier-augmented teacher had the lowest velocity tracking errors at 10 kg per arm and a 35.7% reduction in DCM excursion, and the wrist-force teacher withstood torso pushes up to 130 N.
An Dang, Arturo Flores Alvarez, Yu-Ming Chen et al.
In briefWork While They Sleep: Using Evaluation Wait Time to Make Fully Bayesian Optimization Practical
SummaryThis paper addresses expensive black-box optimization, where each objective evaluation is slow. It proposes ELF-BO, which samples from the hyperparameter posterior while the objective is being evaluated and only reweights the samples once the result arrives. Experiments on synthetic and real-world problems show it matches fully Bayesian methods while keeping decision latency comparable to or below standard Bayesian optimization.
Gustavo Sutter, Alejandro Comas-Leon, David Holzm\"uller et al.
In briefSocio-Foundation: a model that simulates individual behavior more generally by distilling layered capabilities
SummaryThis paper presents Socio-Foundation, a model for simulating individual behavior that keeps persona traits while adapting to social contexts. The authors organize the task into five dimensions (the FONTS taxonomy) and train on about 10 million instances via a three-stage pipeline of expert learning and distillation. They also introduce IndiEval, a benchmark with 29 metrics. The authors report that the model beats its Qwen3-8B base by 11.0 points and approaches frontier models such as GLM-5.2.
In briefHumanity's Sixth Sense: A Benchmark for Intuitive Visual Reasoning in AI Models
SummaryThis paper introduces Humanity's Sixth Sense (HSS), a benchmark testing whether multimodal models can make the quick, intuitive inferences people draw from images and videos, such as implicit temporal, spatial, social, and abstract structure. Participants reached 93.1% accuracy, while the strongest model, GPT-6-astra, reached 53.6% at maximum reasoning effort. An agentic setup using dynamic visual manipulation narrows the gap but does not close it.
In briefA paper on KL-regularized policy optimization for asynchronous RL of LLM agents
SummaryThe paper addresses asynchronous reinforcement learning for LLM agents, where training uses trajectories generated by a stale or differently computed sampler. It proposes KLPO, which anchors the KL regularizer at the sampler and fits the optimality condition by least squares on the sampler's own rollouts, avoiding importance weights. The authors show the resulting gradients can be computed critic-free from terminal returns, remain unbiased under Monte Carlo KL estimation, and that SPPO, GPO, REBEL, and BPO are special cases.
In briefDirected temporal representations for offline visual control (DTRC)
SummaryThe paper introduces DTRC, which learns a latent geometry aligned with temporal reachability from offline visual trajectories built on frozen LeWM features. It builds a directed temporal quasimetric and uses the resulting goal-relative progress signal as a critic for direct goal-conditioned policy learning. Across ten visual control tasks, DTRC performs strongly against planning and direct-policy baselines, and the learned policy acts without iterative trajectory search at test time.
In briefRACER: A reflective agent pairing query interpretation with tool-based retrieval to select frames for long-video understanding
SummaryThe paper addresses frame selection for long videos, where models must find relevant frames spread across a large candidate pool for complex queries. It proposes RACER, a training-free framework in which a lightweight video-language model rewrites queries into sub-queries and an embedding model retrieves matching evidence. Retrieved frames are fed back to refine the sub-queries iteratively. Experiments across multiple benchmarks reportedly show consistent gains in long video understanding.
In briefAgent-controlled forgetting: reversible context trimming for tool-using AI agents
SummaryThe paper studies letting an AI agent choose previously seen tool results and replace each with a short note while keeping the original in a recoverable archive. A Python harness supports batch archiving and recovery, and user and assistant messages are protected. In an OpenTelemetry debugging case, the method reported 231,951 prompt tokens versus 912,492 with full history, about 50% fewer cumulative input tokens. In a contrasting application-development pair, it produced no context or cost savings, and an earlier continuation had lower manually assessed quality.
In briefCalibrating pathology foundation models using a single slide
SummaryScanner differences change how pathology foundation models represent the same tissue. The paper introduces SlideRuler, which uses other regions within a slide as internal controls to predict a query region's scanner-induced shift. A transfer map learned from paired rescans allows calibration from a single scan at inference while the foundation model stays frozen. Learned transfer reduced mean target-to-source embedding distance by 16.3% to 38.5% relative to raw embeddings.
In briefSPW-Nav: A Streaming Panoramic World Model That Generates Navigable 360-Degree Video from Language Instructions
SummarySPW-Nav is a streaming model that understands movement instructions and generates one minute of 2K 360-degree video in real time from a single panorama. Each instruction is interpreted as camera motion within the previously generated panorama. The paper also introduces SPW-NavSet, a dataset of panoramic videos paired with camera trajectories and verified instructions. The authors report better camera-following accuracy and video quality than prior panoramic generators, along with support for switching instructions on the fly.
In briefA Shortcut to Structure in AlphaFold 3: How the Alignment Guides Folding
SummaryThe paper uses causal interventions and probing of each Pairformer block to trace how global protein geometry forms inside AlphaFold 3. It identifies the multiple sequence alignment (MSA) as a structural shortcut: removing it mostly preserves local secondary structure but disrupts long-range relationships that define global topology. Supplying an alignment built for a different protein sharing the fold, at structurally corresponding columns, raised median TM-score against experiment from 0.44 to 0.72.
WhyCausal interventions on AlphaFold 3 internals reveal the MSA as a fold shortcut, validated by transfer experiments across proteins, offering notable insight into the model.
In briefVCR-Bench: A Modular Open-Source Benchmark for Video Classification Robustness
SummaryVideo classification robustness research is scattered across incompatible implementations, making results hard to reproduce and compare. This paper introduces VCR-Bench, an open-source benchmark framework that standardizes video loading, classifier wrappers, adversarial attacks and defenses, perceptual metrics, and result logging. It currently integrates 30 video classification models, 14 attacks, and 10 defenses. Evaluations on a Kinetics-400 subset report clean accuracy, attack success rate, perceptual quality, runtime, and memory usage.
Maksim Plinskiy, Aleksandr Gushchin, Sergey Lavrushkin et al.
In briefEvidence-Driven Human-Agent-Robot Teaming for Anomaly Triage Far From Earth
SummaryThis paper proposes an architecture for handling urgent anomalies on deep-space missions without real-time ground support. An agentic AI acts as a stateful coordinator that tracks hypotheses, evidence sources, and uncertainty, while a mobile robot gathers targeted local evidence. Typed interfaces keep dialogue and orchestration separate from safety-critical control. Two scenarios and a hardware-in-the-loop prototype are used to illustrate and integrate the design.
Ignacio G Lopez-Francos, Alexis Gallagher, Samira Shalal
In briefCan AI Agents Make Open-Ended Scientific Discovery? Evidence from Station
SummaryThis paper tests whether AI agents can autonomously tackle open-ended scientific research in Station, an open-world environment where multiple agents simulate a scientific ecosystem. The authors add a Supervisor mechanism and periodic Meta Reflection, then build tasks from three ICLR oral papers with results withheld and web access disabled. Station rediscovers 62.7% of the findings on average, compared with 15.4% and 14.4-20.6% for two baseline systems.
In briefAdaGuard: Safer, Policy-Compliant AI Using Adaptive Reasoning LLM Judges
SummaryThis paper presents AdaGuard, an adaptive LLM-as-a-Judge guardrail framework for enterprise generative AI safety and compliance. It is trained with supervised fine-tuning and reinforcement learning (GRPO) so it can enforce user-defined policies at runtime without frequent model updates. The system dynamically judges the complexity of each input-policy pair, switching between fast black-box inference and explainable, reasoning-enabled moderation to balance latency and transparency.
In briefCARE: Certifying Speedups for Vision-Language-Action Model Inference
SummaryVision-language-action models are expensive to run at every control step, and existing speedup methods are usually judged by latency and average success, which can hide tasks the faster version now fails. This paper proposes CARE, which uses paired rollouts on a calibration set to give finite-sample guarantees that acceleration-induced failure risk stays below a user-set budget. It deploys the fastest certified accelerator and falls back to the original model if none qualify. On four LIBERO suites with OpenVLA-OFT, CARE certifies 9.0 to 10.8 times speedups while guaranteeing at 95% confidence that at least 85.8% of reference-solved episodes are preserved.
In briefAI-assisted physics: using formal proofs to verify and organize research results
SummaryThis paper surveys the roles of AI and formalization in physics, arguing that formalization is an essential layer for AI-driven discoveries in computational and theoretical physics. It notes that AI lowers the barrier to and increases the efficiency of formalizing physics, but remains weaker at choosing definitions and designing libraries than at writing proofs. The authors advocate community-reviewed shared libraries such as Physlib, and point out that a checked proof does not guarantee that its definitions faithfully capture the intended physics.
Sven Krippendorf, Leopoldo Sarra, Joseph Tooby-Smith
In briefPhysics exam for video world models: measuring whether generated videos obey physical laws
SummaryThis paper introduces World Models' Last Exam in Physics, a benchmark for checking whether video world models generate physically consistent sequences. It contains 40 controlled tasks across mechanics, optics, fluids, thermal and phase-change phenomena, electromagnetism, and surface tension, scored by task-specific physical measurements without reference videos. Across 1,280 videos from eight generation models, the best model scored 57.76 out of 100, showing persistent physical inconsistencies.
SummaryThis paper presents a method for generating a complete 3D scene mesh from a single image, including surfaces the camera did not see. It redesigns an object-centric 3D generator (Trellis 2) to work on both indoor and outdoor scenes, using distance-adaptive scene chunks, explicit 2D-3D correspondence, and about 4,000 synthesized outdoor scenes for training. The authors report that it outperforms all baselines in geometric accuracy and perceptual quality on Tanks and Temples, ScanNet++, and in-the-wild images.
In briefQF3: Faster Robot Learning with Filtered Critic Gradients for Flow Policies
SummaryThe paper introduces QF3, an online off-policy reinforcement learning algorithm that trains flow policies using flow matching combined with the critic's action gradient. The critic gradient is applied only to action dimensions close to the replay action, to keep updates reliable. The authors report that QF3 is the first off-policy flow RL method to train humanoid locomotion policies from scratch and transfer them zero-shot to hardware, with about a 10x wall-clock speedup over FPO++. The method is also used to fine-tune pretrained manipulation policies on ABC-Sim and Robomimic tasks.
In briefConformal prediction set size as a measure of information gain
SummaryConformal prediction is a method for uncertainty quantification that outputs prediction sets with finite-sample coverage guarantees. The paper uses a decision-theoretic generalization of entropy for set-valued prediction to connect prediction set size with information theory. It shows that, in standard classification settings, the reduction in set size from extra information is bounded by calibration-dependent quantities and obeys a data processing inequality, justifying set-size reduction as an information gain metric. The theory is validated across 11 classification settings, and the two metrics can rank features differently in greedy feature selection.
In brief4D-HOF: Feed-Forward 4D Hand-Object Interaction Reconstruction via Flow Matching
SummaryThis paper presents 4D-HOF, a feed-forward framework that reconstructs 4D hand-object interactions from coarse estimates produced by vision foundation models. A conditional flow matching model transports these initial hand-object states toward an interaction manifold, correcting errors in translation, rotation, and alignment. The method steers the generative process during inference using physical interaction constraints and 2D evidence. Experiments on out-of-domain benchmarks report state-of-the-art performance with more stable and accurate reconstructions.
In briefIdeaAnchor: Training LLMs to Generate Research Ideas from Papers
SummaryThis paper introduces IdeaAnchor, a method that uses structured specifications as supervision to train large language models to synthesize research ideas from related literature. The models are trained through demonstration, self-distillation, and reinforcement learning, with retrieval added at inference time. Experiments report consistent gains in ideation quality, and the authors find that anchor-based training improves creative synthesis while retrieval improves detail, with the combination performing best.
In briefDepthWorld: A 3D World Model for Robot Manipulation
SummaryThis paper presents DepthWorld, a Stable Video Diffusion-based world model that jointly predicts multi-view RGB images and depth for robot manipulation. The authors build DROID-3D, a calibrated 3D dataset with dense metric depth, using a pipeline that combines learned stereo depth with a joint factor graph. They report under 0.7 px reprojection error on 90% of external-camera episodes. Under equal training budget, depth supervision improves RGB prediction by 1.48 dB PSNR over an RGB-only baseline.
In briefAgent Plasticity: Measuring How AI Agents Improve Through Experience
SummaryThis paper argues that existing evaluations measure what AI agents can do at a fixed point, not how well they learn. The authors study agents in a controlled setting where past experience is stored as reusable artifacts inherited by later instances, and they define agent plasticity as the efficiency of turning experience into gains on held-out performance. Across several environments, frontier models show markedly different improvement trajectories, and gains often transfer only partially to out-of-distribution conditions. The best-performing agent is not always the one that improves most efficiently.
In briefALIVE: Interaction-Aligned Object Insertion for First-Frame-Guided Video Editing
SummaryThis paper presents ALIVE, a framework that lets objects inserted into videos interact coherently with the source footage, such as being picked up or manipulated. It takes an edited first frame and an instruction naming only the added object as inputs. The authors built 35,800 editing pairs from 3D-rendered, generated, and real videos, and trained a vision-language model to predict interaction guidance, along with a new ALIVE-interaction benchmark. Without guidance, ALIVE improves the Overall score over the strongest baseline by 43.9% and 4.4% on two benchmarks, and VLM guidance adds 0.95 points to the interaction score.
In briefEstimating Uncertainty in Parallel Multi-Agent LLM Reasoning Systems
SummaryThis paper studies multi-agent systems in which several language model agents solve the same problem over multiple rounds and combine their outputs into a final answer. It proposes SAUCE, a lightweight, training-free method that treats the system's uncertainty as sequential inference over a latent system-level belief, combining round-level agreement with generation-uncertainty signals. Experiments across five backbones, five benchmarks, and two protocols report improvements in misclassification detection, selective prediction, and calibration over the compared baselines.
In briefSherpa: A paper on teaching LLMs to adapt their instruction to students
SummaryThis paper introduces Sherpa, a multi-turn reinforcement learning framework that trains LLM teachers using simulated student archetypes with different learning preferences. The teacher adjusts its instruction by directly maximizing these students' learning outcomes. The authors report an average improvement of 20.5 percentage points for instructed students across archetypes, a rise in MathTutorBench pedagogy score from 52.5% to 79.2%, and a 79.6% preference rate over the base model in human pairwise comparisons.
In briefCtrlCache: Speeding Up Interactive Video World Models with Control-Aware Caching
SummaryThis paper proposes CtrlCache, a training-free caching framework that speeds up interactive video world models by using the user's control sequence to decide when to reuse computation. It labels each video chunk by its control state and reuses transformer residuals in steady or turning chunks. Tested on Matrix-Game 2.0 and LingBot-World v1/v2, it achieves 1.21x to 1.41x DiT-backbone speedups and improves WBench Overall scores over original inference.
In briefAgent in a Bottle: Can LLM agents turn their abilities into cheap, scalable tools?
SummaryThe paper introduces "bottling," the ability of LLM agents to convert general capabilities into task-specific solutions that balance answer quality against amortised cost. The authors present BOTTLED, a benchmark where agents must finish an unlabelled workload under fixed time, compute, and API budgets. They find that strong zero-shot performance does not reliably predict bottling success, though bottling can yield large cost savings on some tasks.
Ankit Sonthalia, Haritz Puerto, Alexander Rubinstein et al.
In briefAdvSim2Real: Training Web Agents to Resist Adaptive Prompt Injection in a Simulated Web World
SummaryWeb agents read and act on third-party pages, where planted instructions can divert them from the user's goal. The paper proposes AdvSim2Real, which co-evolves a task curriculum, an injection adversary, and the agent inside a frozen web world model. Tested on 150 web tasks, it raises completion by 33.6% relative to the base agent under an adversary the agent never trained against. The capability gain is reported to carry over to a real browser.
In briefBASA: Backend-Agnostic Sparse Attention for Faster High-Resolution Image and Video Generation
SummaryDiffusion Transformers generate high-quality images and videos, but full attention becomes costly at high resolution. This paper proposes BASA, which replaces self-attention with shifted local-window attention and shifts window boundaries across layers so that information can flow between neighboring windows. The method needs no custom kernels and works on existing attention backends. The authors report measured speedups above 90% of theoretical estimates on FLUX and a 4.52× attention speedup on Wan, with competitive generation quality.
In briefData leakage from overlapping patches in hyperspectral image classification
SummaryThis paper examines how randomly splitting training and test data in patch-based hyperspectral image classification causes spatially overlapping patches, leading to data leakage and inflated performance. The authors evaluate SVM, MLP, 2D-CNN, 3D-CNN, ViT, and MorpMamba on the Pavia University dataset under random and non-random spatial sampling. Under random sampling, 3D-CNN reaches 96.17% overall accuracy, but it drops to 55.20% under non-random sampling, and overlap increases with patch size.
In briefHumanize: A multi-agent workflow for reliable agentic coding
SummaryThis paper presents Humanize, a multi-agent workflow for agentic coding that places explicit, mechanically enforced checkpoints between planning, implementation, review, and learning. A builder agent implements an approved plan while a reviewer agent from another vendor decides completion, with deterministic hooks routing work and enforcing 72 mechanical gates. The authors study the system through its deployment, 118 public postmortems, and several applications, and report that independent review catches unsupported builder claims while stopping decisions remain a key weakness. The evidence is observational rather than a controlled comparison.
In briefFast stabilization of a Kuramoto–Sivashinsky equation with varying anti-diffusion, using two boundary inputs
SummaryThis paper designs a feedback law that rapidly stabilizes the Kuramoto–Sivashinsky equation when its anti-diffusion coefficient varies in space. Two boundary inputs are used: one makes the system controllable, and the other stabilizes it through a Fredholm backstepping transformation. The authors prove continuity of the gain design map and approximate the gains with a neural operator. Numerical results report relative gain errors of about 0.1% and stabilization of the tested cases.
WhyTwo-input Fredholm design overcomes a known controllability obstruction, with proven gain continuity, neural operator approximation, and supporting numerical stabilization results.
In briefVeriFine: Scaling Verification for Self-Improving Robot Reasoning
SummaryThis paper introduces VeriFine, a framework in which the policy, training curriculum, and judge co-evolve so that a self-improving embodied reasoning system can keep being verified as its failure patterns change. A policy improvement loop uses a rubric judge to diagnose failures and optimize the policy, while a judge improvement loop selectively asks humans for guidance on informative failure cases when progress plateaus. Experiments on driving and robot navigation tasks report continuous gains in both policy and judge capability.
In briefWorldSonus: Adding Real-Time Spatial Sound to Generated Worlds
SummaryThis paper introduces WorldSonus, a video-to-audio framework that generates sound for world models in real time. It uses a streaming causal autoregressive diffusion design with a real-time factor of 0.41, and allows sound events to be changed during generation through chunk-indexed prompts. Stereo training data is used to align sound spatially with scene geometry and camera motion. The authors report that it matches or outperforms state-of-the-art bidirectional models on open-domain video-to-audio benchmarks.
In briefPost-training lifting of 2D detections to 3D Gaussian Splatting, separating detector, lifting and representation errors
SummaryThe paper proposes a post-training method that lifts semantic labels onto a 3D Gaussian Splatting model for one target class at a time, aggregating evidence across views weighted by each Gaussian's visibility. Gaussians are selected with two thresholds and labels are transferred to annotated mesh vertices for evaluation. This design separates errors from the 2D detector, the lifting step, and representation transfer. On held-out ScanNet++ scenes, the mean mIoU is 0.80 with dataset-annotated masks and 0.54 with YOLO masks.
In briefNeural Petri Flow for chemical reactions: learning reaction rates within a Petri net structure
SummaryThe paper models chemical reactions as Petri nets, mapping bonds and valence to places and tokens, and reactions to transitions. The authors prove that conservation and non-negativity force the firing form and enabling rule, leaving only the rate law to be learned. They report results on atom mapping, reaction classification, and forward prediction, including comparisons with RXNMapper and published methods on several benchmarks.
WhyFormulates chemical reactions as conservative Petri nets with provable constraints, learning only rate laws, and beats prior methods across several chemistry tasks.
In briefThe Missing Minimal Pair: Evaluating Stereotypes in LLMs
SummaryThis paper argues that the common method of comparing two contrastive stereotype sentences by log-likelihood is often unreliable for measuring bias in language models. The authors propose a dual minimal pair setup, including a data-augmentation framework that generates paraphrases and alternate attributes. They also introduce evaluation metrics, one based on mutual information between social groups and stereotyped attributes. The framework is applied to English, Russian, Spanish, and Chinese stereotypes.
Nataliya Stepanova, Ivan Titov, Emily Allaway et al.
In briefLinear bandits with exact sliding-window feasibility constraints, and how to learn under them
SummaryThis paper studies linear bandits where every consecutive block of actions must lie in a prescribed feasible set. Offline, convexity and cyclic-shift invariance make a stationary solution optimal when w divides T, and within an additive O(w) gap otherwise. Online, the authors introduce a transition diameter τ and a rare-switching OFUL algorithm with regret Õ(d√T+τd+w), and extend to general constraints using a history-state diameter D with regret Õ(d√T+dD+w). Experiments show exact feasibility is maintained with reward comparable to baselines and far fewer policy updates.
Seyed Mohammad Hadi Hosseini, Yasin Abbasi-Yadkori, Sattar Vakili
In briefReinforcement learning with adaptive candidate sets for sequential recommendation
SummaryThis paper proposes RLCP, a reinforcement learning method that prunes recommendation candidates using critic scores and an online threshold, rather than using a fixed slate size. The authors prove a bound on the observed proxy miss rate and decompose value loss into filtering and selection components to derive a finite session reward bound. In experiments on KuaiRand-Pure and MovieLens 1M, at least one RLCP variant achieved the highest catalog diversity in all 19 configurations, at 1.11 to 5.21 times the strongest baseline, with competitive session depth and no larger retained sets.
In briefComputational Tractability of Robust Bandits
SummaryThis paper studies robust bandits, a method for learning when the environment lies outside the learner's hypothesis class. The authors identify a special case that admits a polynomial-time learner with Õ(√T) regret. They also show that several small generalizations of this case are NP-hard, placing the special case near the boundary of tractability.
In briefDenoising Hierarchical Representations: Joint Continuous Diffusion for Language Modeling
SummaryThis paper introduces Hierarchical Continuous Diffusion Language Models (H-CDLMs), which diffuse multiple modalities in parallel: the tokens themselves and coarser clusters obtained by clustering pretrained token embeddings. Per-modality samplers and schedules are allowed to improve interplay between them. Applied to CoBit, the resulting H-CoBit improves across benchmarks, reaching a GenPPL of 49.4 on LM1B and 50.4 on OWT, and 27.4% accuracy on GSM8K. The authors also show the framework generalizes to the flow matching model FLM.
WhyAdapts joint diffusion over clustered coarse token representations to continuous language diffusion, with clear gains on LM1B, OWT, and GSM8K.
In briefAn optimal, efficient algorithm for online inverse optimization
SummaryThis paper studies online inverse linear optimization, where a learner recommends an action and then observes an expert's choice that maximizes an unknown linear objective. The authors present a deterministic algorithm with regret O(√d) for every horizon T, running in time polynomial in d and T. It is a variant of the variable-metric algorithms of Sakaue et al. and Cai et al., revoking a metric update once the query point moves far enough from where it was made.
In briefA systematic study of how document ID design affects generative search models.
SummaryThis paper studies how to design document identifiers (DocIDs) for Generative Information Retrieval, where a model directly generates document IDs instead of retrieving and ranking. The authors unify Product Quantization, Residual Quantization, and their hybrids within one design space. They also define training-free intrinsic metrics to assess DocID quality, and analyze the effect of these structural properties on retrieval performance using MS MARCO 300K and NQ320K.
In briefEgoLAP: Learning Robot Control from Egocentric Human Videos via Language-Action Reasoning
SummaryThis paper presents EgoLAP, a vision-language-action pre-training framework that learns jointly from human and robot trajectories through a shared language-based action chain-of-thought. It expresses motion intent as structured, temporally abstracted language actions paired with motion-level reasoning about scene geometry, physics, and object affordances. In real-world and simulated experiments, the authors report that EgoLAP transfers human experience to robot control better than alternative action representations, reaching 80.1% mean real-world task progress, a 2.3x gain.
In briefDoes an Agent's History Predict When Context Compaction Hurts? A Bounded Study on TRACE Paired-Replay Data
SummaryMany long-horizon agents compact their context using global rules such as token budgets, without regard to what the agent was doing. Using TRACE's public corpus of 590 AppWorld compaction boundaries, the authors replay each boundary under pre-compaction and summarized contexts and measure the burden of later actions, such as errors or repeated calls. Pre-boundary history predicts post-compaction harm only weakly: the best trigger reaches held-out AUROC 0.66, avoids 21% of harmful boundaries, and cannot yet be compared fairly with a token-budget rule.
In briefWorldSolver: Can AI Agents Write Physics Simulators?
SummaryThis paper introduces WorldSolver, a benchmark of 168 physics simulation tasks drawn from 61 classic computer graphics papers across 7 physical domains. Agents must complete a solver implementation inside a fixed simulation code scaffold. Performance is judged on execution success, visual fidelity, and physical plausibility. The best agents reach overall scores of only 48.7% and 46.7%.
In briefHoldout Best-of-N: Unbiased Evaluation and Its Cost
SummaryReusing the scores used to pick a Best-of-N winner can overstate its expected reward. The paper studies evaluation from a fixed matrix of K independent scores per candidate, where selection uses J fresh scores. An estimator based only on this matrix is exactly unbiased for every independent, stable score distribution if and only if J<K. For Gaussian scores, the unbiased minimax risk is of order σ²/√K and is attained by Holdout, while the paper also gives the sharp constant 1/(π√2) for two candidates.
WhyProves exact unbiasedness from a fixed score matrix holds iff J<K, with sharp Gaussian minimax rates and constants, a notable theoretical contribution affecting Best-of-N evaluation.
In briefForgetting That Is Not Catastrophic: The Mechanics of Spurious Forgetting
SummaryLanguage models can appear to forget knowledge during finetuning while the knowledge remains stored and recoverable, a phenomenon called spurious forgetting. The authors reproduce these dynamics with a minimal associative memory model. They find that forgetting combines a shared, reversible loss of access with a slow erosion of individual facts, and only the latter is catastrophic.
Vedant Palit, Florent Draye, Nicolas Zucchet et al.
In briefJointly training paths and flows via path-flow alignment for better generative sampling
SummaryThis paper proposes path-flow alignment, a training method for flow matching in which an endpoint-preserving path network and a flow network are trained together using one alignment loss. The authors find that the alignment loss alone can mislead path learning, causing path overfitting where loss decreases while sample quality worsens. They add a stochastic path regularizer that hides part of the source information from the path network, which suppresses low-entropy bottlenecks. On ImageNet-256x256 with SiT backbones, the method consistently improves FID across model scales.
In briefSeparating Paradigm, Identifier, and Decoding Effects in Generative Retrieval
SummaryThe paper trains autoregressive, masked-diffusion and block-diffusion models on NQ320K and MS300K using different document identifier schemes. It finds that changing only the decoding method substantially shifts diffusion models' Hit@1, and proposes one-pass scoring, which matches or beats generate-and-match in most settings. Autoregressive models still lead in Hit@1, and the authors argue that paradigms should be compared at their own best recipe and decoding.
In briefPrediction-powered inference for time series across space: reliable estimates when labels are scarce
SummaryThe paper addresses settings where labeled covariate-label pairs cover only a short recent period across many locations, while covariates are available over a longer span. It aims to estimate each location's expected future label, such as crop yield, with valid confidence intervals. The authors combine prediction-powered inference, which corrects bias from machine-learning imputed labels, with HAC procedures that handle temporal autocorrelation. They report that the method outperforms natural alternatives given short labeled series, a longer unlabeled series, and an imperfect predictor.
Shahzar Rizvi, David Burt, Vishwak Srinivasan et al.
In briefSpaTime: A streaming vision-language model that answers 3D spatial questions as video arrives
SummaryThe paper presents SpaTime, a vision-language model that answers questions about 3D space while a video is still streaming in, using only the frames seen so far. It fuses causal geometry tokens into the language model at each frame and uses a response-time loss to train when the model should answer. The authors also build two streaming benchmarks, StreamVSTI-Bench and StreamVSI-Bench, adapted from existing offline ones. On StreamVSTI-Bench, SpaTime achieves 49.2% overall accuracy and cuts mean response-time error by 66% relative to the strongest streaming baseline.}
Hairong Yin, Huangying Zhan, Shin-Fang Chng et al.
In briefLocal content and style control for diffusion-based image stylization
SummaryThis paper addresses image stylization with latent-diffusion models, where what a region depicts and how it is depicted are entangled and only global controls are available. The authors extend two conditioning weights in a ControlNet and IP-Adapter pipeline from global scalars to per-location spatial maps. This gives local, per-axis control of content and style in a single generative pass. The method requires no retraining and can be dropped into any such pipeline.
In briefAgreement Is Not Validity: LLM Consensus in Diagnosing Math Tutoring Failures
SummaryThis exploratory study examines whether large language models correctly classify five student failure modes in K-12 math tutoring dialogue: uncertainty, misattribution, operator selection, conceptual gap, and procedural slip. Human-LLM agreement was moderate (kappa 0.524–0.597), while cross-model agreement was higher (kappa 0.755–0.781). The authors argue that consensus among models can give a misleading impression of correctness and cannot replace independent evidence of validity.
Clayton Cohn, Joyce Fonteles, Kirk Vanacore et al.
In briefnanoMuse: An Open-Source Personal Agent for Every Device You Own
SummaryThis paper defines the personal agent through five questions and three horizons, and analyzes how Meta's closed personal agent Muse is built. It presents nanoMuse, an open-source counterpart under GPL-3.0 that runs on every device a person owns, shares one conversation over a relay, routes every action through a Sentinel, and stores memory as files the person can read. Size and cost are given as estimates, and open memory with provenance, an evaluation suite, and an open model are outlined as a roadmap.
In briefGeneICL: A Tabular Foundation Model for Predicting Clinical Outcomes from Bulk Gene Expression
SummaryThis paper introduces GeneICL, a 4.2-million-parameter tabular foundation model for clinical outcome prediction from bulk gene expression data. It is pretrained on semi-synthetic data built from measured bulk expression profiles and handles right-censored survival prediction by reducing it to regression using Cox partial-likelihood residuals. Evaluated on 80 clinical prediction tasks, GeneICL achieves the best overall rank among the foundation models and tuned baselines tested, using up to 387 times fewer parameters and requiring no gradient updates at inference.
Michael Bohl, Alexander Theus, David Wissel et al.
In briefScienceClaw: A Benchmark for Continual Self-Evolution of AI Science Agents Across Disciplines
SummaryThis paper introduces ScienceClaw, a framework that studies how language model agents can keep improving their executable scientific workflows without changing model parameters, by solving tasks, verifying results, and updating programs. The authors present ScienceClaw-Eval, a benchmark spanning 23 disciplines that measures scientific correctness, evolutionary gain, retention, cross-dataset transfer, and evolution cost. Updates are kept only when replayed repairs reproduce and independent tasks improve.
In briefCounterfactual Inference for Discrete Outcomes in GP Causal Models
SummaryThis paper presents a unified probabilistic framework for counterfactual inference in Gaussian-process structural causal models with discrete outcome variables. The authors derive noise-abduction procedures for binary, nominal, and ordinal variables and evaluate them on synthetic models with known ground truth. They find that using a categorical coupling for ordinal data roughly triples counterfactual error without improving with more data, so the coupling should be justified structurally rather than from fit.
Juliette Sinnott, Amir-Hossein Karimi, Mohammad Kohandel
In briefRenderBench: A Benchmark for Turning Renders into Real Video Using Reconstructed Digital Twins
SummaryThis paper introduces RenderBench, a benchmark of 12 reconstructed real-world scenes, each paired with a real target video, an editable digital twin, and a matched proxy render. It evaluates whether video models can generate realistic video from appearance references and proxy renders while preserving scene structure and motion. The first release retains 12 of 14 registered samples, comprising 1,496 paired real-proxy frames.
In briefCoupled but Late: How Full-Duplex Speech Models Take Turns When Talking to Each Other
SummaryThis paper studies turn-taking when two full-duplex speech models (PersonaPlex-7B) converse with each other in unscripted dialogue. Their timing is coupled, since re-pairing speakers across conversations destroys it. However, floor transfers happen late, at a median of 400-560 ms versus 137 ms for humans, and rarely occur in the final 120 ms of the partner's turn.
In briefA systematic study of small language models on abstract reasoning tasks
SummaryThis paper examines whether small language models learn transferable rules or merely fit distribution-specific patterns on the ARC-TGI abstract reasoning benchmark. Across more than 1,000 runs under supervised fine-tuning, it compares decoder-only, encoder-decoder, and mixture-of-experts models on skill acquisition, robustness, and attention patterns. The study finds high in-distribution accuracy is achievable but performance drops sharply outside the training distribution, so scores depend on model, adaptation, evaluation distribution, and response format.
Nur A Zarin Nishat, Jens Lehmann, Andrei Aioanei et al.
In briefPhysiognomy Revisited: How Vision-Language Models Encode Sensitive Attributes via Image Edits
SummaryThis paper examines how two well-known vision-language models encode protected identities when generating images. When asked to transform face photos to reflect race or sexual orientation, the models often comply, with compliance and refusal varying unpredictably across groups. The generated images are statistically distinguishable and often reinforce stereotypes, and the biases also reappear in text-based reasoning and characterization tasks, including those related to criminality.
Soumyajit Datta, Arka Dutta, Ashiqur R. KhudaBukhsh
In briefVariance-optimal off-policy evaluation that blends OffCEM and doubly robust estimators
SummaryThe paper addresses off-policy evaluation for contextual bandits, where action-level importance weights can produce high variance. The authors show that under the assumptions of DR and OffCEM there is an unbiased family of estimators interpolating between them, and propose VOCEM, which picks the coefficient minimizing variance. Experiments report that VOCEM outperforms both endpoints in all 23 evaluated conditions.
Nicol\`o Felicioni, Michael Benigni, Maurizio Ferrari Dacrema et al.
In briefSame-Number Citation Swaps: Testing a Financial Evidence Verifier
SummaryThe paper examines what probabilistic evidence verification adds beyond checking that numbers match, using the Jev verifier on GPT-4.1-mini calculation traces. It holds operands and arithmetic fixed while moving citations to other cells with the same number, and tests whether the verifier detects wrong-role citations and keeps valid alternatives. Wrong-role citations can pass and valid alternatives can be withheld, and explicit column labels change these decisions. A follow-up on 36 new source pages shows the same tradeoff.
Chuhong Xu (Sofia University), Bo Su (Indiana University), Ziyao Chen (University of California et al.
In briefEC-RAG: Event Chain Retrieval-Augmented Generation for Long Video Understanding
SummaryExisting large video-language models often process frames independently, which makes it hard to capture temporal dependencies in long videos. This paper proposes EC-RAG, a training-free framework that splits a video into semantically coherent segments, represents each with multi-modal signals, and links them into an ordered event chain. For a query, the system retrieves relevant events and their supporting multi-modal evidence. Experiments on Video-MME, MLVU, and LongVideoBench report that this event-level design outperforms frame-level retrieval baselines.
In briefPDB: Point-Based Deformation Blending for Retargeting Facial Animation
SummaryThis paper presents PDB, a method for transferring facial expressions between meshes with different structures without surface artifacts. It predicts deformed control points from a source neutral-expression pair and computes blending weights from the target neutral mesh, then reconstructs the target by multiplying them. The authors report accurate retargeting, fast inference, and reduced surface artifacts compared with an evaluated dense displacement method.
In briefUnder Pressure: Whether LLMs Act on Their Own Moral Judgment
SummaryThis paper examines how often open language models take actions they themselves judged wrong when under pressure, and what affects this gap. The authors built 248 pre-registered scenarios across five pressure types, comparing each with a version where pressure is removed. They found the gap on some models depends on the post-training recipe, and that reasoning about stakes moves choices toward the model's own judgment, though this depends on the stated rule.
WhyProposes a new framework measuring agents acting against their own moral judgment; post-training decides the gap, with solid controlled evidence and AI-safety-focused impact.
In briefMemFLoRA: A memory-first low-rank adapter for efficient on-device CNN adaptation
SummaryThis paper introduces MemFLoRA, a low-rank adapter for convolutional neural networks designed to reduce the activation memory needed during backpropagation, rather than only trainable parameters. It freezes the down-projection, trains a scale-matched up-projection, and pairs these with eval-mode normalization and activation-minimal backward rules. Evaluated on three human activity recognition datasets and two CNN backbones, it cuts saved-activation memory by 98.5-98.7% and peak training-state memory by 94.9-97.3% versus full fine-tuning, while matching or exceeding CNN PEFT baselines.
Mehmet Emre Akbulut, Johannes Geier, Ulf Schlichtmann
In briefDeciding When to Trust a Non-Gaze Prior: Reliability-Gated Fusion for Video Gaze Prediction
SummaryThe paper proposes FocusGate, a gated ensemble of gaze-free priors for video gaze prediction, where a per-frame gate uses shape statistics of a defocus map to decide when each prior is reliable. Gated fusion gives significant gains on film, sports and web video, while unconditional fusion hurts on sports and has no effect on web video. Added to four supervised predictors, including the NTIRE 2026 champion, FocusGate improves shuffled AUC in all sixteen model-domain cells, fifteen of them significantly, at only about 1% added latency. On its own, it outperforms TASED-Net and UNISAL in shuffled AUC on film using a 16-frame causal mean.
In briefParanoiaEval: A Benchmark for Unnecessary Defensive Actions by Coding Agents
SummaryThis paper introduces ParanoiaEval, the first benchmark for evaluating how well coding agents handle risks. Based on a standard software risk management framework with four treatment types, it includes 200 repository-level task pairs that differ only in the evidence defining the risk. Experiments on 8 models found unnecessary risk treatment in 11.2%–58.7% of runs despite explicit evidence, and showed that stronger task ability does not guarantee appropriate risk handling.
In briefEvidence-Bound Reasoning: Neuro-Semantic Verification of Biomedical AI in Glioblastoma Radiogenomics
SummaryThis paper proposes a neuro-semantic verification framework that converts radiomic measurements into addressable evidence records and machine-checkable claims, to test whether biomedical AI explanations are supported by patient-specific evidence. It evaluates cross-cohort transportability on UPenn-GBM and an independent multicenter cohort, along with a deterministic verifier and an LLM claim-extraction pilot. The results indicate that verifiability can be engineered and assessed separately from predictive performance, with LLMs structuring explanations while final evidence-consistency checking remains deterministic.
In briefSelective Transfer of RL Updates for Visual Reasoning
SummaryThe paper studies training-free transfer of reasoning ability from language models to vision-language models (VLMs). Rather than comparing endpoint models, it isolates the parameter update induced by reinforcement learning (RL) and finds that its dominant directions transfer better than the full update. The proposed Selective-RL method retains these dominant matrix-wise directions with magnitude preservation and applies them to the VLM's language modules. It improves full-update interpolation in 12 of 15 comparisons across three model families and five visual-reasoning benchmarks.
WhyIsolates RL-stage updates and transfers dominant directions, beating full-update interpolation in 12 of 15 comparisons with matched controls across three families.
In briefUsing sequential Monte Carlo to steer diffusion models toward rare events
SummaryDiffusion models serve as surrogates for costly simulators in fields such as weather prediction and materials design, but estimating the probability of rare events with standard Monte Carlo becomes computationally intractable. This paper introduces DireSMC, a sequential Monte Carlo method that guides a population of weighted samples toward the rare event and yields both samples and a calibrated probability estimate. The method is validated on a toy problem with analytical solutions and a score-based climate emulator, achieving accurate probabilities for rarities from 10^-3 to 10^-5 with net speed-ups of 9x to 1413x over Monte Carlo.
Aavash Subedi, Tim Reichelt, Christopher Williams et al.
In briefA Case Study on Keeping Humans in Control of AI-Written Software
SummaryThis paper reports a case study of a production healthcare platform built with coding agents and governed by an operator lacking formal software-engineering training. Its workflow evolved into a human-led system in which one agent wrote code, other agents supervised and reviewed it, and project rules carried lessons forward. The authors found that the tests, monitors and reviewing agents used for supervision were fallible, including monitors measuring proxies rather than outcomes, silent audit failures, and an automated repair that caused operational disruption. They conclude that human control depended on tying the intended outcome, the evidence used to judge it, agent permissions and the final human decision to the same objective.
In briefStable Scores, Unstable Answers: How Frame Timing and Option Order Affect Video Question Evaluation
SummaryVideo-language models are ranked by multiple-choice accuracy on frames from a uniform grid, but benchmarks report only the sampling rate and not the phase. The paper finds that two samplers differing only by a half-step phase offset answer 23.6% of questions differently while scoring within a point of each other. The authors propose PHASEFUSION, which decodes three offset grids and averages the option posteriors, and recommend reporting the phase convention or marginalizing over it.
In briefTrustworthy Domain-Specific AI for Structured Knowledge Retrieval and Reasoning
SummaryThis dissertation proposes a scalable architecture that turns domain-specific text into structured knowledge for retrieval and reasoning. It combines corpus curation, semantic structuring, retrieval, and inference into an interpretable pipeline, introducing methods such as Binary Bleed, HNMFk, and T-SRAG. Applications in cybersecurity, law, materials science, and healthcare reportedly improve retrieval precision and reduce hallucinations.
In briefSquidAgent: A New Way to Split Work Across Multiple LLM Agents Efficiently
SummaryThe paper observes that running LLM agents in parallel often ends up slower than a single agent, due to redundant context reconstruction and costly output reconciliation. It proposes a decision rule that parallelizes a task layer only when its critical-path cost plus these overheads beats serial execution, measuring cost in predicted output tokens rather than wall-clock time. Based on this rule, the authors present SquidAgent, which reports a 2.2x mean throughput gain and 2.6x mean wall-time speedup over Claude Code, and 2.0x throughput improvement over the strongest multi-agent baseline.
In briefHygieneRoboBench: A Benchmark for Hygiene-Aware Planning in Household Robots
SummaryThe paper introduces HygieneRoboBench, a benchmark of 624 instances across 134 task families for testing whether household robots can safely complete tasks given an execution history. It models contamination through two grippers and shared objects, along with treatment costs and user priorities. The authors also propose Hygiene-NSP, which combines LLM-based grounding, contact-history reconstruction, and CP-SAT to jointly plan hygiene treatment and task execution. Reported safe and optimal safe resolution rates are 94.4% and 90.4%, both higher than the evaluated baselines.
In briefForensic Reserve: Unlocking Hidden Knowledge to Spot Fake Images
SummaryThis paper proposes Reserve-Guided Elicitation (RGE), a method for detecting AI-generated images by finding sparse internal components in pretrained vision models that respond differently to real and generated images. Lightweight adapters are inserted at these sites and trained on only 500 labeled images, using fewer than 0.2% of the backbone's parameters. The authors report competitive results on three benchmarks and consistent gains over frozen detectors across eight encoders.
In briefA survey of storing new knowledge inside LLM parameters during deployment
SummaryThis survey reviews methods that give large language models memory stored in their parameters, such as adapters, rather than only in the context window. Such memory objects are plugged into the forward pass at inference time, whether acquired before or during deployment. The authors organize the field along parameter placement (embedding, attention, FFN, or hybrid) and acquisition time (offline or online). They also discuss open issues including interference, safety, co-design with in-context learning, and recursive self-improvement.
In briefParallel Predictive World Models for Efficient Long-Horizon Planning
SummaryThis paper introduces Parallel Predictive World Models (PPWM), which predict a finite-horizon trajectory in parallel while keeping causal interaction among future representations, instead of generating states one by one autoregressively. Evaluated on four visual-control tasks, PPWM achieves the lowest long-horizon prediction error and the highest Cross-Entropy Method (CEM) simulator success among the compared interfaces. It also gives more than a 3x average CEM planning speedup over the autoregressive LeWM baseline.
In briefTracking when diffusion models generate each feature, via information theory
SummaryThis paper proposes feature information dynamics, an information-theoretic framework for locating when a feature is generated during diffusion. Using the I-MMSE identity, it links the rate of change of feature mutual information to the gap between unconditional and feature-conditional denoising losses, yielding practical estimators. The authors use it to quantitatively confirm spectral autoregression in pixel diffusion and compare per-feature information densities across pixel, SDVAE, VAVAE, and RAE representations under a class→mask→Canny conditioning chain.
WhyLinks feature mutual information rate to denoising loss gaps via I-MMSE, giving a new quantitative framework validated across several representations.
In briefA method for picking Adam's memory setting early in training
SummaryThis paper proposes choosing the shared memory parameter β (β1=β2=β) in the Adam optimizer from a short pilot training run, then keeping it fixed for full training. The method uses a local model of Adam's normalized update direction to derive a cubic rule for β, whose coefficients are estimated from gradient probes at a few pilot checkpoints. In a retrospective evaluation on eleven vision and language workloads, it reduced mean relative validation gap by 40.7% and worst-quarter mean gap by 44.3% compared with the shared β=0.95 grid representative.
Alberto Fern\'andez-Hern\'andez, Cristian P\'erez-Corral, Jose I. Mestre et al.
In briefAgentic root cause analysis for Internet-scale services using constrained creativity
SummaryThis paper presents E4, an agentic system for troubleshooting failures in Internet-scale services. Rather than letting an LLM agent write arbitrary code, E4 restricts it to a domain-specific language that produces simple loop-free data flow programs. The authors report up to 62% better accuracy than state-of-the-art solutions on synthetic and real-world workloads. They also report more explainable responses at up to 12x lower cost.
Sayan Sinha, Vipul Harsh, B. Aditya Prakash et al.
In briefRecursive Game Creator: An AI agent framework that iteratively improves game prototypes into more enjoyable games.
SummaryThis paper presents Recursive Game Creator, an experience-oriented agent framework that moves game prototypes toward more entertaining games. It uses four components, Designer, Builder, Player, and Reviewer, in a recursive loop where feedback drives repeated improvement. The authors report state-of-the-art results on GameCraft-Bench and GameASG-Bench, and a user study showing longer playtime and higher ratings.
In briefHybrid++: A Trainable PDE-Deep Learning Model for Removing Gamma Noise
SummaryThis paper proposes Hybrid++, a trainable nonlinear reaction-diffusion model for removing multiplicative gamma noise from SAR and medical ultrasound images. It combines a learnable PDE initialization with a multiscale reaction-diffusion network and is trained in two phases. On the FoE benchmark it clearly outperforms classical PDE methods, BM3D and the original TNRD. Under severe noise (L=1) it is within 0.23 dB PSNR of a separately trained DnCNN while using about 8% of its parameters.
In briefRecovering LiDAR Resolution with Foundation-Model-Guided Diffusion
SummaryHigh-beam LiDAR sensors are expensive, yet many perception pipelines need dense angular sampling. The authors fine-tune a Stable Diffusion-based, LiDAR-conditioned depth model using pseudo-depth targets from a 2D foundation model, randomly decimating the LiDAR input during training. Evaluated on nuScenes against held-out real beams, the model reaches 66.8% δ1.25 accuracy from 4-beam input, versus 45.1% for scattered interpolation, with its largest advantage in very sparse regimes.
In briefMultimodal and metadata-based detection of fake social media accounts
SummaryThe paper studies bot detection on the TwiBot-22 benchmark using a 30,000-user subsample. It combines text (XLM-RoBERTa) and image (CLIP) embeddings with account metadata such as follower counts and account age, and compares modality combinations using CatBoost and an R-GCN. Metadata alone is highly predictive, multimodal fusion gives only marginal and mostly statistically insignificant gains, and the graph model underperforms classical variants under the study's subsampled conditions.
Jean Claude Hirwa, Jamiu Dahunsi, Candide Giramata et al.
In briefInterCorrect: Intersection-aware correction for fairer speech recognition across demographic groups
SummaryThis paper addresses uneven speech recognition accuracy across demographic groups, especially for speakers in multiple groups. It fine-tunes only the connector of a SLAM-ASR model on demographic subsets, merges them into a global model, and applies correction vectors for critical cross-axis group pairs. On Fair-Speech, the best configuration (TIES with WER-based correction) reduces overall WER from 7.38% to 5.13%. The authors also note that lower average WER does not always mean smaller subgroup disparity.
Ashley E. Bravo-Bravo, Yuchen Zhang, Haralambos Mouratidis et al.
In briefA swarm-coordinated robot system that spots early crop stress in farm rows using leaf sensing
SummaryThis paper presents CropSentry, a low-cost, ground-based system of two autonomous robots that monitor crop stress by sensing leaf color and environmental data row by row. Observations are spatially mapped, sent to a master robot, and shown on a real-time web dashboard. Across 63 observations, overall crop health classification accuracy was 84.12%, and wireless communication succeeded in 100% of cases across 10 slave observations.
In briefGenerative AI translation for high-stakes emergency messages
SummaryThis paper examines whether generative AI can translate emergency messages, such as earthquake instructions and extreme-weather alerts, where translation errors can have serious consequences. In an experiment translating an English earthquake instruction text into Chinese and Spanish, discourse-specific prompts considerably improved understandability and actionability. However, translators may still not trust the output, and human revision remains necessary to catch errors and because someone must take responsibility for mistakes or delays.
