In briefWorldFoundry: A Unified Open-Source Infrastructure for World Intelligence
SummaryThe paper defines a generative world model as a predictive dynamical system that keeps a persistent, history-conditioned state of an environment and updates it in response to actions. Building on this definition, it proposes a framework covering perceptual grounding, state modeling, controllable transformation, dynamics-aware prediction and closed-loop interaction. The authors release WorldFoundry, open-source infrastructure for reproducible inference and evaluation across video, 3D/4D, embodied and interactive systems, with a catalog of 270 models, 74 benchmarks and 62 metrics.
In briefHow reliable is a seabed image embedding? Repeat-look tests of frozen encoders on sonar
SummaryThis paper treats embeddings from pretrained encoders applied to seabed sonar images as repeated measurements of the same patch, taken from different looks on overlapping survey passes. Using synthetic-aperture sonar data, the authors test a frozen DINOv3 encoder and find that standard single-look metrics and a coral linear probe are stable, but repeat-look reliability is poor (dbICC 0.20). A stable probe direction reaches an ICC of 0.60, while acquisition geometry explains under a tenth of the within-cell spread.
Harold Horsley, Oscar Pizarro, Hayat Rajani et al.
In briefLook Ahead: Search-driven reasoning for robots that test possible futures before acting
SummaryThe paper studies how a robot policy can search over candidate action sequences before acting, testing them with rollouts from cloned simulator states on the PushT task with a frozen ACT policy. The authors introduce One Step Ahead (OSA), a gate that triggers full beam search only when predicted coverage drops beyond a threshold. In the reported ablations, OSA keeps or improves success while cutting runtime to about one third of full beam search.
Ahmed Zeer, Osama Shbib, Muzaffer Kaan Yuce et al.
In briefBlenderFORGE: A Framework for Training AI Agents to Edit 3D Scenes Reactively
SummaryThis paper presents BlenderFORGE, a framework for training open-weight multimodal language model agents to edit 3D scenes by generating Blender Python code. It covers four state-verifiable editing tasks: object placement, shape-key editing, lighting adjustment, and material modification. The framework combines perturbation-based data generation, an isolated Blender Sandbox for deterministic execution and reward computation, and a two-stage training pipeline. Experiments show that the resulting Qwen3-VL-8B and Qwen3.5-9B agents substantially outperform their base models and achieve competitive results against some zero-shot proprietary baselines on selected metrics.
In briefCascading Circle Loss: a contrastive learning method for ranking legal texts by graded relevance
SummaryThis paper introduces Cascading Circle Loss (CasCL), a contrastive learning framework for legal retrieval with graded relevance. It combines an InfoNCE objective for separating positives from negatives with a cascade of ranking boundaries that model ordered relevance levels between queries and legal clauses. The framework adds a negative-aware self-paced optimization mechanism and a relevance gain weighting strategy. Experiments on legal passage retrieval benchmarks report that CasCL consistently outperforms InfoNCE, Circle Loss, MNRL and Margin MSE on both binary and graded ranking metrics.
In briefBAPLe++: Transfer and Defense of Prompt-Based Backdoors in Medical AI Models
SummaryThis paper studies backdoor attacks on medical foundation models adapted through prompt learning, where a poisoned prompt can leave clean accuracy intact while steering triggered inputs to an attacker-chosen class. It is the first to examine whether such prompt-based backdoors transfer across different models, measured with Clean Accuracy Difference and Backdoor Accuracy Difference. The authors also propose DePLe, a defense that uses the infected prompt as an anchor and learns a new prompt on a small clean set. Experiments on four models and six chest X-ray and histopathology datasets show that transfer is asymmetric and dataset-dependent, and that DePLe reduces backdoor accuracy more consistently than four training-free input transformations.
In briefConwayNet: Game-of-Life-style lateral inhibition for more interpretable, consistent, and noise-resistant neural networks
SummaryThe paper proposes a differentiable lateral inhibition mechanism based on continuous Game-of-Life rules, which derives inhibitory and excitatory responses from layer inputs and blends them into activation maps in a state-conditioned way. Its rule parameters are fixed rather than learned. When added to fully connected, convolutional, wide-residual and graph convolutional networks, it makes activation maps more coherent across space, depth and seeds, and gives hidden-unit co-activation graphs a spatial layout. The authors report that under label noise, networks from nearly identical initialisations disagree on thirty times fewer test images, saliency maps focus on objects rather than backgrounds, and robustness to unseen pixel noise improves substantially.
In briefA unified framework for assessing clinical movement from skeleton data using convex-concave and sparse graph learning.
SummaryThis paper addresses skeleton-based rehabilitation assessment and proposes CC2-STG, a multi-stream graph convolutional network. It decomposes joint relationships into complementary convex-concave distance functions, uses sparse regularization to select relevant geometric features, and applies adaptive topology to capture non-anatomical joint dependencies in pathological movements. The authors report state-of-the-art results on the KiMoRe and UIPRMD datasets for clinical movement assessment.
Quang-Anh N. D., Duc Pham Minh, Minh-Anh Nguyen et al.
In briefConstrained look-ahead guidance for flow-based image editing that reduces conflicts between edit and fidelity goals
SummaryThe paper proposes an inversion-free framework for editing images with flow models, working from the generative trajectory itself. At each step, a one-step look-ahead evaluates how candidate updates affect estimated edit and fidelity rewards, and corrections are applied under an explicit fidelity constraint. This is meant to reduce gradient interference between the two objectives. Experiments report stronger semantic edits with better preservation of source content than existing strong baselines.
In briefAERO: Adaptive Ensemble-Disagreement Routing for Oracle Feedback in Sample-Efficient Online RLHF
SummaryThis paper proposes AERO, an online reinforcement learning from human feedback (RLHF) algorithm for aligning language models. It uses ensemble vote entropy to decide which contexts need costly oracle preference labels and which can receive synthetic labels from a learned reward model. Experiments across model families and scales report competitive alignment performance with a smaller oracle-query budget than online DPO and active-exploration baselines.
In briefREINS: A Self-Evolving Runtime Harness for Real-Time Trajectory Planning
SummaryThis paper introduces REINS, a runtime harness that wraps an LLM-based trajectory planner so it can operate within real-time control limits. LLM outputs are grounded in a learned dynamics model, verified through calibrated predicates and forward rollout, and consolidated into an indexed rule memory that handles familiar scenes without calling the LLM. The authors report 98%+ compliance, nearly 50% fewer collisions, and 1.9–4.6 ms latency at the tactical decision layer, with LLM invocation declining as memory grows and no changes to LLM weights.
In briefRoadState: A vision-language system that turns road images into lane-level map data for updates
SummaryRoadState is a production system that converts sequential road-scene observations into structured lane-level semantics for Baidu Maps' map-updating pipeline. Its core model, RoadState-VLM, treats road-scene understanding as autoregressive state estimation and uses Geometric Chain-of-Thought, a reasoning method guided by a dependency graph of road-state features. On a real-world dataset it reaches 94.33% mean accuracy across six road-state dimensions, above the strongest fine-tuned VLM baseline. Since March 2025 it has supported updates in more than 360 cities and reduced overall production costs by 30%.
In briefIdentifying handedness from offline Bangla handwriting under controlled collection conditions
SummaryThis paper examines whether a writer's handedness, left or right, can be identified from offline Bangla handwriting. The authors use their BanglaHand dataset of 42 writers, 21 left-handed and 21 right-handed, collected and photographed under controlled conditions. Test writers' images are held out, and whole-writer permutation tests and photometric controls are applied. An ink-gradient descriptor reaches 77.4% balanced accuracy on unseen writers and 64.5% after ink density and brightness are regressed out.
Tanzim Hasan Prappo, Sayeed Hassan, Sayem Ali et al.
In briefTesting whether vision-language models' step-by-step instructions really work, using crochet as a real-world check
SummaryThis paper proposes crochet as a test of whether vision-language models can produce working procedures. Given an image of a basic crochet object, models must infer how it was made and write a pattern that reproduces it. The authors crocheted 40 patterns generated by four models by hand, and only four were free of errors. The results expose weaknesses in visual, geometric, and procedural reasoning.
In briefWhether pulse wave shape carries blood-flow signs of an impending stroke
SummaryThe study analyses continuous photoplethysmography (PPG) recordings made before physician-recognised stroke in intensive-care (MIMIC-III, 176 patients) and emergency-department (MC-MED, 154 patients) cohorts. Beat-by-beat pulse waveform descriptors separated pre-recognition windows from stable windows at a 360-minute horizon, with models transferring unchanged between cohorts and outperforming heart rate alone. The authors propose pulse morphology as a candidate continuous cerebrovascular biomarker.
In briefMatchability-aware conformal prediction for open-ended LLM answers
SummaryThis paper introduces CAELUM, a framework that gives language model answers calibrated uncertainty by building semantic prediction sets. It samples answers from clarification prompts, clusters them by meaning, and separates aleatoric from epistemic uncertainty using kernel-based von Neumann entropy before applying split conformal prediction. Across 30 model and dataset combinations (4-32B), mean coverage is 0.74 in a stratum identified by a lightweight classifier, compared with 0.65 overall.
In briefThe Lagrangian Brain: A Constraint Mechanics of Mind and Machine
SummaryThe paper proposes that brains follow a constraint mechanics in which the world, experience and goals act as constraints on a single state, each enforced by its own multiplier that pushes the state back. From this it derives several neurobiological features, such as Dale's law and brain rhythms. A gradient-free machine trained on 1,000 solved puzzles is reported to solve all 422,786 Sudoku-Extreme test puzzles and to handle larger mazes and longer strings.
In briefLocating Tiny Evidence Regions with Selective Contrastive Self-Distillation
SummaryThis paper addresses the weakness of multimodal large language models on fine-grained visual tasks that depend on evidence in tiny image regions. It proposes selective positive-negative on-policy self-distillation (SPN-OPSD), which uses supervision from both evidence and background regions to resist irrelevant interference, along with selective token-level distillation to focus on evidence. On Qwen3.5-2B, accuracy rises from 76.44%, 45.92%, and 57.43% to 84.82%, 54.56%, and 59.56% on V* Bench, ZoomBench, and MME-RealWorld-Lite, with similar gains on Qwen3.5-0.8B.
In briefAnomaly detection in machining vibration signals using a spectrogram-based CNN autoencoder
SummaryThis paper presents a one-class anomaly detection method for monitoring machining conditions from vibration signals, using a convolutional neural network autoencoder. Three-axis vibration signals are converted into spectrograms and combined into a three-channel input, and the model is trained only on normal process data to flag anomalies by reconstruction error. A new metric, Segmented Vertical Error (SVE), is introduced to better detect short-lived instabilities such as chatter. Tests on about 10,000 broaching strokes show clear separation between normal and anomalous conditions, with preliminary checks on a public turning dataset.
In briefBreaking precision walls in optimizers with one preset constant
SummaryThe paper studies how optimizers behave when signal strength decays as a power law across directions, so strong directions are fit while weak ones can be missed even though average loss looks fine. It defines a precision wall as the steepest decay at which the weakest direction still reaches within 5% error in a fixed budget. A rule that sets one constant from true curvature or spectrum raised this wall on low-rank matrix sensing and quadratic tasks across shapes and seeds, while results on logistic regression were mixed.
In briefLIFT: A Few-Shot Depth Completion Method Using Local Metric Alignment and Foundation-Guided Propagation
SummaryThis paper presents LIFT, a framework for few-shot depth completion that recovers dense metric depth from sparse measurements using only a small number of target-domain annotations. It leverages depth and feature priors from a frozen pretrained monocular depth estimation foundation model, estimating local scale and shift parameters that are interpolated into dense fields. The initialization is then refined through foundation-guided multi-scale depth propagation and a lightweight residual decoder. The authors report state-of-the-art few-shot performance on NYUv2 and KITTI-DC.
In brief3D Segmentation in the Clinic: A Grand Challenge
SummaryThis paper describes a segmentation competition for automatically or semi-automatically extracting the liver from CT scans and the caudate nucleus from brain MRI data. It was held as a workshop at the 2007 MICCAI conference. The paper discusses the rationale for the competition, describes the training and test datasets for both tasks, and presents the scoring system used to evaluate segmentation results.
In briefHow well large language models resist cognitive biases in industrial safety hazard reviews
SummaryThis paper tests how three frontier large language models behave when used for HAZOP studies, a method for preventing catastrophic industrial accidents. Across 420 trials, the models resisted 96.8% of cognitive manipulations, and overall susceptibility was statistically inconclusive. However, safety-critical risk downgrades appeared under low anchors and gain frames, so the authors argue that automated hazard analysis needs disciplined human-in-the-loop validation.
Vinicius M. M\"uller, Vitor A. O. Silva, S\'avio S. V. Vianna et al.
In briefLocating and repairing domain shift in VLM driving trajectory planning
SummaryThis paper studies how vision-language models used as end-to-end driving planners lose accuracy when moved from one city to another. Activation patching reveals that the shift is structured across image versus text tokens and across network layers. The authors propose MoLaRx, which first locates where the shift occurs and then repairs only those modality-layer zones with matched operators. On three VLMs and two cross-city benchmarks, it achieves the smallest cross-city trajectory-error gap in all six main settings while using 50-60% fewer trainable parameters.
Zhihong Cui, Hengyu Liu, Michael Alexander Riegler et al.
In briefCan self-improving AI loops work on frontier pre-training runs?
SummaryThis paper asks whether recursive self-improvement (RSI), previously studied only in post-training, can be applied to foundation model pre-training. The authors apply RSI to an entire frontier pre-training run from scratch by splitting it into smaller loops that can each be closed. The result is 801-30B-A3B-Base, a 30B-A3B MoE model trained on 11.8T tokens over three months. The model and lightweight SFT data are open-sourced.
WhyThe authors apply recursive self-improvement to a full frontier pre-training run, decomposing it into closable loops, and report comparisons on several benchmarks with open-released model and data.
In briefFedSGN: Federated Sparse Grassmann Newton Learning for IoT Anomaly Detection
SummaryThis paper proposes FedSGN, a federated learning method for detecting anomalies in Internet of Things devices. It applies an ℓ2,0-norm constraint to the global projection matrix to remove redundant features and adds an orthogonality penalty to preserve structure. An augmented Lagrangian solver with a semismooth Newton scheme enables parallel client updates. Experiments show good overall detection performance with reduced communication overhead, and the source code is publicly available.
In briefGenerating stylized images with style and content separated by rearranging attention
SummaryThis paper proposes a method for disentangled stylized image generation using IP-cross-attention rearrangement in diffusion models. It notes that style and content are coupled in diffusion latent spaces, causing content leakage, and that style is hard to describe with text. The authors also introduce negative content guidance, inspired by classifier-free guidance, to reduce content leakage.
In briefPermute-then-Adapt: a weak-to-strong method for contrastive image-text adaptation
SummaryThe paper proposes Permute-then-Adapt (PTA), a framework that improves contrastive training of image-text models such as CLIP. A efficient weak model identifies semantically related pairs that would otherwise be treated as negatives, and a random-pairing reference distribution with a significance level α selects auxiliary positives. A single multi-positive contrastive objective then trains the model on this calibrated positive set. The authors report gains over heuristic soft-label approaches in object recognition, cross-modal retrieval, and data efficiency.
In briefFaster painting rendering via structured stroke initialization and audits
SummaryThis paper presents a framework for stroke-based painting rendering that models continuous curved strokes analytically with Quadratic Bezier Splatting and differentiable alpha compositing. A feedforward multi-scale stroke decoder provides structural initialization, and test-time stroke optimization refines reconstruction quality. A single-stroke audit mechanism evaluates local stroke quality so defective primitives cannot survive optimization by being hidden under overlapping layers.
In briefAbductionBench: A Benchmark for Abductive Reasoning in Large Language Models
SummaryThis paper introduces AbductionBench, a unified framework for evaluating abductive reasoning, which is inferring plausible explanations from evidence. It applies the framework to 27 datasets across ten domains and tests eight LLMs on single-choice tasks, analyzing both final answers and observable reasoning traces. Results show models match or beat small LLMs on compact evidence but fall behind on longer cases, and rarely revise their initial hypotheses in passive tasks.
Moein Salimi, Amir Qeysarbeigi, Babak Hosseini Mohtasham et al.
In briefR-BED: A Framework Linking Consequences to AI System Evaluation and Continuous Assurance
SummaryThis paper presents R-BED, a specification-level methodology that connects evidence on technical failures, deployment-control effectiveness and real-world consequences to authorization decisions for deployed AI systems. It defines thirteen diagnostic dimensions, a failure-mediated expected-loss identity, bounded evidence statuses, a control assurance ladder, five authorization dispositions, and a Continuous System Evaluation layer in which evidence expires when the system configuration changes. A synthetic refund-agent walkthrough shows that a task-success view can miss a consequential control weakness. The methodology has not been empirically validated.
In briefReducing spectral bias in gated local ReLU networks
SummaryNeural networks trained by gradient methods learn low-frequency content faster than high-frequency content, a phenomenon called spectral bias. The paper studies shallow ReLU models under a general bilinear loss, covering least-squares regression and H^1 variational losses, and proposes local ReLU dictionaries gated by hat partition-of-unity weights. For L^2 regression, the spread of frequency-dependent learning rates drops from 1.9×10^5 to 4.5, while for the H^1 stiffness matrix gating inverts the frequency ordering instead of removing the bias. The authors conclude that a preconditioner built from the same stencil is needed to fix spectral bias for that case.
In briefPLS-GS: Compact 3D Gaussian Splatting Aligned to LiDAR Surfaces
SummaryPLS-GS is a method that builds a compact 3D Gaussian Splatting scene representation by combining camera images with LiDAR data. It fits a hierarchical Gaussian mixture model over position and color to form a continuous scaffold, and constrains rendering Gaussians to the LiDAR-derived surface. Across five large outdoor scenes, it places primitives closer to LiDAR measurements than image- and LiDAR-initialized baselines, using substantially fewer primitives for a more compact model and faster rendering.
Krishnan Harikumar, Alejandro Fontan, Thierry Peynot et al.
In briefHallucination-guided unlearning: using hallucinations to find overfitted memories
SummaryThis paper proposes Hallucination-Guided Unlearning (HGU), which treats a language model's hallucinations as diagnostic signals to locate unsupported text spans. HGU traces a sparse set of high-responsibility MLP neurons and edits them with compact gates under a KL constraint that preserves utility. The authors report that HGU gives the best forgetting-utility trade-off on TOFU and the strongest forgetting on the WMDP auxiliary task while improving retain and MMLU scores.
In briefPerception, not reasoning, limits visual theory of mind in AI models
SummaryThis paper examines why vision-language models perform far worse on false-belief tasks presented as images than as text. The authors hold the reasoning task constant while varying perceptual fidelity across five conditions, and measure how much of the gap each fidelity closes. Testing seven models on abstract 2D panels and 3D game scenes, they conclude that structured visual scene-state extraction, rather than belief inference, is the dominant bottleneck.
Mohamed Rayan Barhdadi, Syed Talal Wasim, Juergen Gall et al.
In briefMachine Hearing: Can audio LLMs go beyond human hearing?
SummaryThe paper introduces the Machine Hearing Benchmark (MHB), with 6,128 instances across 13 tasks and seven domains, using inputs with sampling rates from about 10^-5 to 10^6 Hz. Evaluations show that current open-source and proprietary audio LLMs transfer poorly to these tasks zero-shot, and supervised adaptation on machine-hearing data alone does not close the gap. The authors propose MachEar, a lightweight encoder extension that improves machine-hearing performance of a pretrained audio LLM while largely preserving its existing audio understanding.
In briefA geometric theory of how Transformer attention and representations evolve together
SummaryThis paper builds a statistical-physics-based variational and geometric theory for how Transformer attention and token representations change jointly while model parameters stay fixed. It defines a free energy for attention and models attention weights with a Fisher-Rao gradient flow, showing that the softmax matrix is the unique minimiser and that positive attention trajectories converge to it. The authors also couple Hamiltonian dynamics for representations with attention relaxation, yielding an energy-dissipation law, and prove global well-posedness of the resulting systems.
In briefJoVox: Joint Hierarchical Flow Matching for Lightweight Voxel Generation
SummaryJoVox is a discrete flow matching framework that learns the joint distribution of all levels of a voxel hierarchy, so coarse cells can be predicted using finer cells already generated nearby. It adds a two-level prediction that outputs cells two levels finer at once, and an unmasking strategy that revisits early decisions to correct inconsistencies. The paper reports that JoVox outperforms baselines on unconditional and image-conditioned voxel generation, and compares favorably with TRELLIS.2 in training throughput, forward-pass FLOPs, GPU count, and Chamfer distance.
In briefRetinaGuard: risk-controlled AI triage for eye screening across several diseases
SummaryThis paper presents RetinaGuard, an agentic triage framework for retinal fundus photographs in which several specialised agents gather evidence and a learned predictor estimates how likely each automated decision is to be unsafe. Acceptance thresholds are set with conformal risk control. On the RFMiD test split, at matched referral budgets and equal certified safety, it captured more unsafe decisions than confidence-based referral. The authors report that the benefit of agentic evidence gathering depends on how much the primary model's own uncertainty already reveals about its errors.
Sohaib Sulman, Muhammad Atif Saeed, Akhtar Jamil et al.
In briefTransMem: Transition-Aware Retrieval for Changing Personal Memory
SummaryThis paper proposes TransMem, a memory retrieval framework for long-term personalized agents. It stores user preference updates as explicit retrieval units alongside factual memories, first retrieving factual anchors and then adaptively expanding over transition records. Experiments on PersonaMem and PersonaBench with two Qwen backbones report that TransMem outperforms representative retrieval-based and structured-memory baselines. The authors conclude that preference-update records are useful retrieval units for long-term personalization.
In briefRankShift: How Benchmark Design Choices Affect Document Parsing Model Rankings
SummaryThis paper examines whether model rankings on document parsing benchmarks are stable across benchmark design choices. Using OmniDocBench as a case study, the authors evaluate 16 OCR and vision-language models under variations in dataset composition, matching strategy, and scoring metric. They find that the interaction between matching strategy and ground truth format causes substantial ranking instability, while metric choice is the most stable factor. The authors argue that leaderboards should report ranking robustness alongside headline scores.
In briefMGE: A membership-guided attack that extracts training data from diffusion models
SummaryThis paper examines how the memory properties of diffusion models can leak training data, a risk known as data extraction attack. The authors propose MGE, a membership-guided extraction framework that treats extraction as conditional generation, conditioned on a membership inference classifier. They add adversarial gradient and parallel branch decoding to improve success rate and efficiency. Experiments on image generation benchmarks show MGE can extract hundreds of images, highlighting the need for greater attention to privacy protection in diffusion models.
In briefFOCAL: Safe offline reinforcement learning with behavior regularization and feasibility guidance
SummaryThis paper studies safe offline reinforcement learning, which learns policies from a static dataset that maximize return while satisfying safety constraints. The authors propose FOCAL, which combines behavior-regularized policy gradients with a conditional variational autoencoder actor, optimizing separately for transitions estimated to be feasible and infeasible. At inference, the scale of the latent prior is adjusted based on the estimated feasibility of the state. On the DSRL benchmark, FOCAL satisfies constraints on all evaluated tasks while achieving competitive or higher returns than baselines.
Mahesh Keswani, Parth Bhardwaj, Ankur Kumar et al.
In briefHeterogeneous Asymmetric Contrastive Distillation (HACD): A Method for Knowledge Transfer Across Different Network Architectures
SummaryThe paper addresses knowledge distillation between different architecture families, such as from a Transformer teacher to a CNN student. It proposes HACD, which maps teacher and student logits onto a shared unit hypersphere and transfers the relational structure of predictions via contrastive matching instead of per-sample logit imitation. HACD adds no learnable parameters, and the authors report the best or second-best accuracy on most heterogeneous model pairs on CIFAR-100 and ImageNet-1K.
In briefLearning causal structure from incomplete time series with changing noise levels
SummaryThis paper studies how to discover causal structure in multivariate time series when values are missing, instantaneous effects exist, and noise variance varies. The authors propose Cheesefill, an optimal transport framework that models causal mechanisms with conditional normalizing flows and jointly learns a stochastic correction map to refine imputed values. Experiments on synthetic and real-world benchmarks report that Cheesefill consistently outperforms existing baselines across different missingness regimes.
In briefPhysics-guided TeX map generation for low-light vision tasks
SummaryThis paper addresses fusing visible and infrared images for vision tasks in low-light conditions. It proposes TeX-Gen, which generates Temperature, Emissivity, and Texture maps from paired visible and infrared inputs using branch-specific gates. The authors also introduce TeXMetric to evaluate these maps without pixel-level physical labels, and TeX Tune to jointly optimize task-guided adjustments while keeping the generator frozen. Experiments on FLIR, LLVIP, M3FD, and MSRS report competitive task performance.
In briefAligning Task-Critical Parts Improves Imitation of Object Interactions
SummaryThis paper improves the DINOBot framework for imitation learning of object-object interactions. It adds RANSAC to reject outliers during alignment, reducing the effect of noisy depth data and spurious feature matches, and focuses semantic correspondence on task-relevant interaction regions. In real-world robot tasks including picking, insertion, shaking, and pouring, the approach raised overall task success rates by 13.8% over the baseline and converged faster during alignment.
Denis Hadjivelichkov, Sebastian {\AE}gidius, Valerio Modugno et al.
In briefDynamics-Informed Action Learning for Goal-Conditioned Latent Planning: A Method That Improves Visual Control in Latent World Models
SummaryThis paper addresses visual control with latent world models, identifying a gap between teacher-forced training and recursive prediction during planning, and the failure of planners to use action-conditioned dynamics when initializing action search. The authors propose Dynamics-Informed Action Learning (DIAL), which adds rollout alignment and multi-step inverse dynamics supervision during training and autoregressively couples inverse and forward dynamics at test time to initialize the Cross-Entropy Method (CEM). The paper reports that DIAL outperforms existing latent world models and planning baselines on four visual control benchmarks, in both standard and long-horizon scenarios.
In briefPandemonium: Long-form speech recognition that tracks multiple speakers together
SummaryThe paper presents Pandemonium, a target-speaker speech recognition system for conversations with several speakers. It extends the Diarization-conditioned Whisper approach to a longform Parakeet backbone (DiCoP), adding cross-speaker attention so speakers are not decoded in isolation, plus an STNO mask refiner. The authors use a pseudo-labeling pipeline to mine 5.4k hours of multi-speaker audio from YODAS2 for training. The combined approach achieves state-of-the-art results on several meeting benchmarks, and the models and pseudo-labeled data are open-sourced.
Alexander Polok, Dominik Klement, Ivan Medennikov et al.
In briefThe Delta Gap: Why Cell Response Model Scores Are Overstated
SummaryThis paper examines how cellular world models predict cell responses to genetic perturbations. It shows that the common global Pearson correlation metric is dominated by thousands of non-responsive genes, which inflates apparent performance. On the Replogle K562 screen, GEARS scored 0.977 globally but only 0.076 on delta Pearson, below a trivial baseline of 0.267. The authors propose delta-space evaluation as the primary benchmark and report partial gap reduction through delta-aligned training and zero-shot transfer to another cell line.
In briefMMGraph-Agent: Agentic Multimodal RAG Using Cache-Inspired Multimodal Knowledge Hypergraphs
SummaryThis paper proposes MMGraph-Agent, a unified multimodal retrieval-augmented generation framework built on cache-inspired multimodal knowledge hypergraphs. It enables extremely lightweight hypergraph construction with zero API cost and uses dynamic memory-based retrieval to alleviate long-context issues. Experiments in offline, online, and joint retrieval settings report consistent gains in performance, efficiency, and architectural unification.
In briefSTABLE: an optimizer that controls parameter drift to reduce forgetting in continual learning
SummaryThis paper tackles catastrophic forgetting in continual learning, where learning new tasks erodes earlier knowledge, by working at the optimizer level. For distance-based classifiers, the authors prove that bounded parameter drift guarantees bounded margin erosion under a local score-stability condition. They propose STABLE, a primal-dual optimizer that raises a dual variable when the drift budget is exceeded and lowers it otherwise. Experiments show vanilla PEFT baselines trained with STABLE match or outperform specialized continual learning methods, and in multimodal streams it also preserves zero-shot transfer.
In briefFG-DSA: Frequency-Guided Attention for Efficient Dense Prediction
SummaryThis paper proposes FG-DSA, a framework that adds a 2D FFT high-pass filtering module to spatial attention, injecting frequency-domain priors. It lets lightweight networks recover sharp object boundaries with linear complexity. Experiments report consistent gains in semantic segmentation, and the module can be used in diffusion-based U-Nets to improve generated image detail with very little extra parameter and FLOP cost.
In briefSHIN: A scalable layered 3D Gaussian Splatting method for incremental rendering quality
Summary3D Gaussian Splatting (3DGS) renders high-quality novel views in real time, but device and network budgets differ. SHIN builds a scalable hierarchy with a reusable base and additive enhancement layers, so higher-quality versions extend the decoded representation. It orders Gaussians using multi-scale position-scale sensitivity and splits each stage's budget between new Gaussians and spherical-harmonic bands using calibrated gradient-based scores per byte. Across three benchmarks, SHIN progressively improves one cumulative representation and achieves the best PSNR-size efficiency in all overlapping comparisons with multi-rate and progressive baselines.
In briefMACaQ: A Fast, Model-Agnostic Calibrated Quantization Method for Low-Precision Weights and Activations
SummaryThe paper introduces MACaQ, a post-training quantization method that compresses both weights and activations to MXFP4 (W4A4) precision. It treats each linear layer independently, applying per-channel scaling and a block-Hadamard transform to activations at inference time while absorbing the inverse into the weights. A Hessian-aware coordinate-descent step, Flip-Refinement, improves on round-to-nearest quantization. The authors report only 2.3% end-to-end overhead over pure MXFP4 on Qwen3-32B, lower perplexity than DuQuant++ and MR-GPTQ, and quantization of a 122B-parameter model in 64 minutes.
Lorenz K. Muller, Philippe Bich, Chiara Boretti et al.
In briefWhy edits to language models often fail to carry over between languages
SummaryThis paper studies why factual knowledge edits made to a multilingual language model in one language often fail to transfer to other languages. The authors use locate-then-edit methods (ROME, MEMIT, AlphaEdit) as an analytical framework and split the cross-lingual propagation gap into key-space propagation, parallel residual mismatch, and orthogonal residual mismatch. They find a large orthogonal residual component is the main bottleneck, and a contrastive alignment objective improves key-space propagation and partly improves transfer, while value-space mismatch remains.
Aastha A K Verma, Anwoy Chatterjee, Sparsh Jain et al.
In briefAgent2BIM: A Systematic Evaluation of Capture Setups and Prior Models for As-Built BIM
SummaryThis paper presents Agent2BIM, an agent-based Scan-to-BIM framework that generates as-built Building Information Models from reality-capture data. The authors evaluate seven capture configurations, including 360-degree panoramas, smartphone video and LiDAR, and terrestrial laser scanning, in a laboratory and an MEP scene. Results are compared against TLS-verified reference models for component detection, geometric accuracy, and MEP connectivity. The findings highlight the complementary roles of visual and metric evidence in agent-based Scan-to-BIM.
In briefVisProve: Learning Typed Visual Provenance for Scientific Document QA
SummaryThe paper introduces ReViE, a dataset of 3,941 questions over complete scientific documents, annotated with page, section, and region labels and typed as Quotation, Compression, or Inference. It proposes VisProve, a method that jointly generates answers and provenance using differentiable spatial evidence reading and evidence-answer relation feedback. On a document-disjoint test set, VisProve improves section-level and typed F1 over joint supervised fine-tuning, though absolute typed F1 remains low.
In briefSame contacts, different meeting setups: how format changes LLM group opinion
SummaryThis paper studies how the format of meetings among large language model agents affects group support, holding members, initial stances, and pairwise contact counts fixed. It compares merging triads into five-person meetings with only reshuffling who meets in triads, finding that merging shifts group support more and in a consistent direction. The authors argue that meeting format is an important variable in LLM social simulations, beyond what contact networks capture.
WhyThe authors test pairwise contact networks in LLM group simulations via a contact-matched intervention, explaining meeting-format effects through a Jensen gap across six open-weight LLMs and 14 topics
In briefSLA-aware IoT task offloading: class-weighted reinforcement learning compared with an exact optimum
SummaryThis study examines how to decide whether Internet of Things tasks with different latency limits run at the edge or in the cloud. The authors build four small systems that dynamic programming can solve exactly, then measure how far reinforcement learning agents fall from the optimal policy. A class-blind penalty stays 65% to 118% above the optimal cost even with a perfect learner, while a corrected per-class penalty comes within 0.13% of the optimum. In larger systems the agents remain 41% to 76% above the optimum, and training longer does not close this gap.
In briefVAComp: View-Aware Full Projector Compensation for Moving Viewers
SummaryThis paper presents VAComp, a view-aware framework that adjusts a projector's input to correct geometric and photometric distortions for a moving viewer in a naked-eye setting. It resolves SfM scale ambiguity with metric calibration and calibrates an oppositely facing camera using two checkerboards. Two non-iterative hybrid methods, VAComp-SD and VAComp-SI, predict compensation images in a single forward pass, and real-world experiments report competitive quality with real-time and interactive rates respectively.
In briefCoverage-verified sparse attention: closed-loop quality control for long-context LLM decoding
SummaryThe paper proposes Coverage-Verified Sparse Attention (CVSA), a training-free decoding loop that uses the fraction of softmax attention mass on selected tokens, called coverage, to decide whether to accept a sparse attention result. Heads with insufficient coverage are recovered rather than accepted. On LongBench, CVSA is within 1.13 points of dense attention and statistically matches it at 70B, while losing 0.8 points on RULER. Measured median decode-step speedup reaches 6.21x at 128K context.
In briefAn improved algorithm for solving quantified Boolean formulas
SummaryQuantified Boolean formulas (QBF) extend Boolean satisfiability (SAT) by adding quantifiers and are a fundamental problem in constraint solving. This paper presents a new memoized-search algorithm that solves a QBF with n variables and m clauses in 1.685^m·poly(n) time and space. It is the first improvement in over 20 years on Williams' 1.709^m·poly(n) bound. The method uses the minimum vertex cover of conflict graphs, and the authors show QBF is slicewise polynomial-time solvable under this parameter.
WhyThe authors report improving Williams' 20-year-old 1.709^m bound to 1.685^m via memoized search and extend conflict-graph vertex cover ideas to QBF.
In briefLatent-WAM++: Shaping World and Action Representations for Driving
SummaryThis paper presents Latent-WAM++, a representation-centered world action model for end-to-end autonomous driving. It assigns each intermediate world and action representation a supervision suited to its role in planning. The method reports 90.1 EPDMS on NAVSIM v2 and, under zero-shot evaluation on HUGSIM, 47.0 average RC and 30.1 average HD-Score. Training cost falls from 1,600 to 190 GPU-hours, an 88.1% reduction.
In briefDFT-Agent Bench: Scaffold Effects and Execution Provenance in Autonomous DFT
SummaryThe paper presents DFT-Agent Bench, over 200 Quantum ESPRESSO materials simulation tasks covering relaxation, electronic structure, equation-of-state, and elastic-tensor pipelines. It compares a near-bare ReAct baseline with a bundled v5.1 scaffold on two language models, finding gains for the full scaffold on bands+DOS and d-electron vc-relax tasks, while elastic results vary by model. A provenance audit checks whether the required execution contract was met, catching incomplete or unmarked runs that may still score highly.
In briefWhether LLM agents keep their identity after a model swap
SummaryWhen the underlying model of an LLM agent is upgraded or replaced, operators usually load the same persona, memory, and rules onto the new model and assume behavior is unchanged. The authors ran 5,760 compliance tests across 12 regulated domains with three open-weight models and found that some model transitions cause significant behavioral drift while others do not, even with identical external state. This drift is hidden from verbal checks, since every model claims full policy compliance while its actual decisions under pressure differ. They conclude that model transitions should be treated as behavioral changes requiring action-level empirical testing.
In briefCONSTRAINER: A method for prompting graph optimization solvers with constraint conditioning
SummaryThe paper presents CONSTRAINER, a framework that treats constraint specifications as structured inputs. Constraints are parsed into operator graphs, encoded by a graph neural network, and used to condition a shared backbone model that predicts solutions. The approach is evaluated on power grid optimization and mixed-integer linear programming problems, where it reports improved solution optimality and more efficient transfer learning and generalization than baseline methods.
In briefEndoSCOP-V: A Video Benchmark for Multimodal Models in Endoscopy Reporting and Clinical Reasoning
SummaryEndoSCOP-V is a multi-turn evaluation framework built on 1,307 real endoscopy and colonoscopy videos totaling 63.2 hours. Its questions are generated deterministically from clinician-filled structured endoscopy reports rather than written by hand or by an LLM. The best open-weight 27B-class model reaches 59.8% strict accuracy, compared with 78% for a panel of 15 board-certified endoscopists, with the largest gaps in quantitative measurement and multi-variable grading.
In briefWhat the collusion probes actually detect: three evaluation pitfalls for multi-agent monitors
SummaryThe paper examines activation probes meant to detect secret collusion among LLM agents. Using a benchmark of 2,500 committee and 1,476 code-review episodes that includes agents who were instructed to collude but refused, the authors find that probes mostly detect the instruction rather than the agent's decision. They identify and document three evaluation pitfalls that can inflate reported detection results.
WhyThe authors separate collusion instructions from decisions, showing probes mainly read instructions, validated on 2,500 committee and 1,476 code-review episodes, plus a testing protocol.
In briefNeuroHorizon: Hierarchical-Memory Long-Horizon Forecasting of Neural Population Activity
SummaryThis paper examines how well models can predict neural population activity far beyond observed data, comparing direct multi-step prediction with autoregressive rollout. It introduces NeuroHorizon, which combines event-based spike encoding with a causal decoder and a hierarchical prediction memory that keeps recent predictions in detail and compresses earlier ones. On macaque motor-cortex recordings it achieves the highest forward-prediction score among evaluated autonomous models at horizons from 100 to 1000 ms, and it uses scheduled sampling to reduce the gap between training and rollout.
In briefDetent: a method for batch-invariant LLM serving and RL rollouts at near-stock speed
SummaryThis paper presents Detent, a method that makes an LLM engine return the same output for a request regardless of which other requests share its batch. Detent expresses kernel selection as a program, fixes only the few batch-dependent values it needs, and tunes them for the workload before deployment. On multi-GPU vLLM and SGLang deployments, it runs at 1.04 to 1.17 times Stock's latency per output token, compared with 1.67 to 10.3 times for the deterministic modes.
WhyThe authors express kernel selection as an executable program fixing only few batch-dependent values, reporting latency versus deterministic modes in vLLM and SGLang and bitwise-reproducible RL rollou
Peixu Xin (Independent Researcher, Boston University), Runchuan Feng (Independent Researcher et al.
In briefAdapting Large Language Models Using Only Forward Passes
SummaryThis paper proposes a method for adapting large language models to specific tasks using only forward passes, the same computation used during inference, without gradient-based training. It measures how each singular value decomposition (SVD) component affects the angle concentration of hidden states to decide which weight matrices to reduce in rank, removing higher-order components and keeping lower-order ones. According to the paper, the approach achieves competitive accuracy across datasets, with theoretical analysis indicating lower peak memory and greater speedup than gradient-based and exhaustive search methods, and it shows robustness across LLMs up to 57B parameters.
In briefReason first, then retrieve key frames: hypothesis-guided frame selection for long videos
SummaryThis paper proposes HypoKFS, a training-free, plug-and-play framework for long video understanding. It first reasons over sparsely sampled frames to form an initial set of answer hypotheses, and selects additional discriminative keyframes only when the answer distribution is uncertain. The selected frames guide renewed inference and hypothesis refinement in an iterative loop. Experiments on Video-MME, LongVideoBench, and MLVU show consistent gains across multiple MLLM backbones, with accuracy improved by up to 6.7 percentage points.
In briefAn LLM-orchestrated multimodal chatbot that helps diagnose tomato plant diseases
SummaryThis paper presents a chatbot that helps diagnose common tomato plant diseases through natural-language conversation. Users describe symptoms in text, and the system returns the most likely disease, basic treatment suggestions, and image-based reasoning. It is built on pre-trained Mistral and Llama 3.2 language models, with LangChain and Streamlit providing a lightweight, offline-friendly interface. The authors report promising experimental results and plan further improvements for real-world deployment.
In briefAgent-built 3D simulation models for manufacturing system design, with an open simulator, benchmark and industrial case
SummaryThe paper presents SimuLean, an open-source 3D discrete-event simulator built on Unity that AI agents can program through domain-level commands. On 18 manufacturing modelling tasks, a frontier agent produced valid models in 93 of 108 attempts, and on the benchmark's interface version two agents built 105 of 108 models using domain commands alone. In an industrial case, one agent rebuilt a published robotised welding workshop study, exposing an error in the published model, and its corrected-model results agree within 1%.
WhyThe authors report an open-source 3D discrete-event simulator built for AI agents, evaluated on 18 manufacturing tasks and an industrial welding workshop case.
In briefAI in the Clinical Documentation Loop: How AI-Written Notes Affect Later AI
SummaryThis paper studies how medical records containing content written by large language models (LLMs) affect other LLMs that later read the same record. It examines two artifacts: note bloat from accumulated LLM summaries, and diagnostic overcall, where an unconfirmed differential diagnosis is written into the record. Across seven reader models from four families, LLM-generated summaries modestly but consistently degraded hospital-course generation, and injected overcalls reappeared in outputs at rates varying by model.
In briefInfoOrgBench: Can Agents Turn Scattered Information into Coherent Hierarchies?
SummaryThis paper introduces InfoOrgBench, a benchmark of 150 tasks in which agents organize scattered meeting recaps, issue threads and design proposals into hierarchies. It checks that every record is preserved and uses a Structure Score to measure agreement with the source hierarchy at every level. The evaluation finds that parent-level relations have higher error rates than leaf-level relations across all three domains.
In briefRenAIssance: Caption-Guided Diffusion Models for Generating Historical-Art Visual Media
SummaryThis paper presents RenAIssance, a pipeline that generates scene captions for historical European artworks using BLIP and uses about 15,000 resulting image-text pairs from DEArt to train a compact LoRA adapter for Stable Diffusion v1.5. The adapter shifts generated images toward the historical-art distribution while keeping the base model's ability to render new concepts. Across historical and contemporary prompts of varying lengths, it improves FID to the DEArt reference in every setting, with only a modest drop in CLIP text-image similarity.
In briefMotionGraphicsBench: Testing AI Motion Graphics with Rule-Based Checks
SummaryThis paper introduces MotionGraphicsBench, a benchmark of 100 expert-curated tasks in which models animate a static SVG according to a motion edit prompt. Each task includes a deterministic verification program that reports which requirements pass or fail, instead of comparing against a reference animation or using a vision-language model judge. On a 180-animation test set, the programs agree with human judgments 93.3% of the time. The paper also reports pass@1 and pass@32 results across many large language models, with scores ranging from 6.0% to 95.0%.
WhyThe authors replace reference animations and VLM judges with deterministic verification programs, evaluating 47 configurations on 100 instances against human judgments.
Jiaju Ma, Ashvini Kumar Jindal, Chuhua Wang et al.
In briefReceiver-resolution APS-PDP maps for wireless channels: a benchmark and coarse-to-fine generator
SummaryThe paper introduces UrbanADPSMap, a benchmark that converts ray-traced multipath data into fixed-grid angle power spectrum (APS) and power delay profile (PDP) targets, accounting for bandwidth and aperture limits. It also proposes SpecFlowAR, a coarse-to-fine generator based on multi-scale autoregressive flow matching that first captures coarse dispersion structure and then refines beam-relevant peaks and low-power tails. According to the paper, under a unified NFE=80 budget, SpecFlowAR reduces BPL@5 to 2.03 dB and lowers errors in 3GPP dispersion statistics.
In briefForecasting Solid Waste Generation with Time Series Models
SummaryThis paper compares four time series forecasting models (SARIMAX, Prophet, LSTM and GRU) using real municipal solid waste data from Sri Lanka covering 2012 to 2018. It examines how holidays, weekends, seasons, temperature and humidity relate to waste generation. The results show lower generation on holidays and weekends and higher generation in summer and winter, while temperature and humidity lack a direct linear relationship. LSTM and SARIMAX achieved the best overall forecasting performance.
In briefEgoVeil: Personalized Privacy Reasoning for First-Person Video
SummaryThis paper presents EgoVeil, a system that decides which parts of first-person video, such as footage from smart glasses, should be obscured based on each user's privacy preferences. It turns visual regions into structured semantic observations and combines a privacy reasoner with a user-specific memory that adapts as preference evidence accumulates. Evaluated on in-the-wild data from 10 smart-glasses users, it achieves 91.93% balanced accuracy and 84.42% mask recall, outperforming direct vision-language model baselines.
In briefGlobal Context Guidance for Diffusion Large Language Models: a feature-space method that helps diffusion LLMs use the whole context while generating text.
SummaryThis paper analyzes how diffusion large language models use contextual information during denoising and finds they rely more on local context than global context. Based on this, the authors propose Global Context Guidance (GCG), a feature-space method that improves global-context use without building an explicit weak model. Experiments on LLaDA-8B, Dream-7B, and LLaDA-1.5 across several benchmarks show GCG outperforms existing guidance methods with lower inference overhead.
In briefROLECUE: A Test for Whether Vision-Language Models Match Who-Did-What Across Languages
SummaryThis paper introduces ROLECUE, a diagnostic benchmark pairing 192 two-person action images with sentences in six languages, in canonical and patient-first word orders. It separately tests image-only role recognition, text-only role parsing, and direct image-sentence matching across eight open-weight vision-language models. Several models succeed on individual components but fail when those predictions are combined, with the largest gap reaching 44.9 percentage points.
In briefHow visual attention in CNNs affects decisions, via decision-space alignment
SummaryThis paper presents a decision-space framework that treats multiplicative visual attention as a perturbation of intermediate CNN features and predicts its first-order effect on the target-versus-competitor classification margin. It introduces the Decision Alignment Score (DAS), which measures the directional alignment between that perturbation and the local margin gradient. Experiments on CIFAR-100 with SimAM and CBAM across five CNN backbones report good agreement between predicted and actual margin changes in most configurations, and positive correlations between DAS and actual margin changes. The associations are stronger when attention is placed at later stages.
In briefDeep Backward Multistep High-Order Regression for High-Dimensional Nonlinear Parabolic PDEs
SummaryThis paper proposes a deep backward multistep high-order regression (MHDBR) method for solving semilinear parabolic partial differential equations via their backward stochastic differential equation representations. The method combines multistep time discretization with regression targets built from conditional expectations. Under stated regularity and network assumptions, the authors derive an error bound that separates time discretization, conditional Monte Carlo sampling, and neural-network approximation. The analysis assumes exact population minimization and does not include finite-sample generalization or optimization errors.
In briefPixels over Symbols: Sensory Realism and Behavioral Alignment in Cognitive Models
SummaryThis paper compares neural network models that use abstract symbolic inputs with sensory-realistic models on a set of memory-dependent decision-making tasks. The sensory-realistic models match human behavioral response patterns considerably more closely than the abstract models, though a gap remains even for the best models. The authors find that sensory input, rather than learned recurrent weights, accounts for much of this difference, and that it matters more than the choice of sequence architecture.
In briefHandling Uncertainty in Mechanism-AI Dynamical Systems Using Optimal Transport and Information Geometry (Part I: Distribution Learning)
SummaryThis paper presents a mathematical framework for handling uncertainty in dynamical systems that couple mechanistic models with AI, using optimal transport and information geometry. The uncertainty quantification problem is cast as minimizing an energy functional over a probability density space with a Riemannian metric, and gradient flows under Wasserstein and Fisher-Rao metrics are described in a unified way. The authors also propose numerical methods for these gradient flows based on abstract adjoint sensitivity analysis, and compare posterior approximation and predictive uncertainty under a common computational budget.
In briefZARA: single-image method to build animatable, relightable 3D human avatars without per-person tuning
SummaryThis paper presents ZARA, a zero-shot feed-forward framework that reconstructs an animatable 3D Gaussian avatar and its intrinsic material attributes from a single image. It combines a pretrained large-scale human reconstruction model with a Material Transformer, using pseudo-material labels for training. Experiments report reconstruction within tens of seconds and better relighting quality and material generalization than existing baselines. The method also allows direct editing of the avatar's material appearance.
In briefLanguage models without QK normalization: a projection that keeps magnitude information
SummaryThis paper argues that QK-Norm, a common technique in large language models, removes token-specific magnitude information from query and key vectors. The authors propose QK-MAP, which uses multiple learnable frequencies to encode magnitude changes as bounded phase patterns, controlling attention logit growth while preserving magnitude sensitivity. Experiments on models from 190M to 7B parameters report lower training and validation loss than standard attention and QK-Norm.
In briefonǎ: A Language Model Method Using Twin Embedding Tables with Tied Output Layers
SummaryThis paper introduces onǎ, a two-table parameterization for language models that keeps input and output weights shared while letting the output layer mix the softmax distributions of both tables for greater expressiveness. A gate based only on the current input token sets the mixture weight, so its values can be precomputed at inference. Evaluated on Dense, Gated DeltaNet hybrid, mixture-of-experts, and Attention Residual models at 142M, 330M, and 1B scales, onǎ lowers validation loss and external perplexity in most comparisons and raises average accuracy on eight downstream tasks. The cost is one extra embedding table and one extra vocabulary projection.
In briefYRC-Bench: A Benchmark for Culturally Aware Translation of Historical Heritage Texts
SummaryYRC-Bench is a human-curated multi-parallel benchmark for translating historical heritage texts among Classical Chinese, Modern Chinese, and nine foreign languages. It provides a three-level parallel corpus that includes modern explanatory glosses and pairs conserved and substituted culturally specific items. Evaluations of nine leading LLMs show that modern Chinese glosses, especially human-curated ones, substantially improve lexical fidelity and semantic adequacy.
In briefMeta-reinforcement learning combined with zero-shot reinforcement learning
SummaryThis paper proposes MetaBFM, a hybrid method that combines Behavior Foundation Models with meta-reinforcement learning so agents can adapt to new reward functions at test time. During training it mixes supervised reward-following with unsupervised learning, and at test time it can interpolate between exploration, exploitation, and zero-shot reward inference. The authors report that MetaBFM recovers the in-distribution performance of supervised meta-RL and generalizes to reward functions outside the training set.
In brief"Seeing Isn't Enough": TIGER Closes the Gap Between Text and Image Reasoning in Multimodal Models
SummaryThe paper finds that multimodal large language models can read text in images accurately but still reason poorly about those images compared with equivalent text, a gap the authors call a modality reasoning-access gap. They propose TIGER, which mines synthetic examples where a model solves a text problem but fails on the matching rendered image, then post-trains the model on image inputs using GRPO, with SFT and DPO variants. Trained only on synthetic rendered-text images, TIGER improves performance on diverse real-world visual benchmarks and complements existing RLVR visual reasoning methods.
WhyThe authors identify a modality reasoning-access gap and bridge it with synthetic counterfactual training, reporting gains on several real-world visual benchmarks.
Gregory Kang Ruey Lau (1, 2), Minh Huynh Nguyen (1) et al.
In briefDual-granularity learning for regression with noisy continuous labels
SummaryThis paper studies regression with continuous noisy labels, where noise magnitude varies across samples, requiring both instance-level correction and reliance on ordinal structure. The authors propose GRAIN, which at the fine-grained level derives a residual learning mechanism from a robust loss using instance-specific auxiliary variables. At the coarse-grained level, it applies binning-based contrastive regularization that aligns feature order with label proximity. A histogram-based bridge couples the two levels bidirectionally. Experiments report consistent gains in robustness and generalization under diverse noise conditions.
In briefHSGNav: Hypothesis scene graphs for robot navigation when objects are moved
SummaryThe paper proposes the Hypothesis Scene Graph (HSG), which keeps persistent 3D furniture entities separate from revisable hypotheses about small objects on surfaces, allowing local updates without rebuilding the scene structure. Building on HSG, HSGNav uses retrieval, exploration, and verification to support multi-scale lifelong navigation. The authors also introduce FAROS-Bench, a benchmark for mid-episode object relocation and multi-scale find goals. Experiments reportedly show competitive results on GOAT-Bench and substantially better performance than prior methods on FAROS-Bench.
In briefEvaluating hierarchical policies from non-hierarchical data
SummaryThis paper studies how to estimate the expected return of hierarchical policies from previously collected data, without further environment interaction. The authors propose Intra-Option Fitted Q-Evaluation (IO-FQE), which estimates a value function over augmented state-option space so hierarchical policies can be evaluated even without annotations of which option was run. On long-horizon continuous control tasks, IO-FQE removes the bias of naively applying fitted Q-evaluation and achieves substantially lower mean squared error than option-level importance sampling.
In briefAFVE-Bench: A Benchmark for Measuring Leftover Old-Framing Errors in Multi-Turn LLM Revisions
SummaryThis paper introduces the abandoned-frame verbalization error (AFVE), where a model adopts a new organizing frame in a revised text but still reveals the superseded frame. The authors build AFVE-Bench, which uses 192 matched four-turn dialogues in English and Chinese across eight writing domains, and evaluate eleven model configurations with two annotators. AFVE occurs in 25.54% of history-conditioned responses, and 19.77% of responses that pass the target still expose the abandoned frame, most often through contrastive framing.
