In briefSafety Reconstructed: Masked-Diffusion Guard Models for Stronger AI Safety Filtering
SummaryThe paper proposes LLaDA-Guard, a safety guard model that asks which safety label better explains a prompt or response, rather than predicting a single verdict token. It fine-tunes the masked diffusion model LLaDA-8B-Instruct with LoRA and a class-conditional reconstruction objective. On seven held-out safety benchmarks it ranks best on average against discriminative baselines on stronger backbones, with better calibration and less over-defense. The generative design also enables token-level risk localization and prompt rewriting to safe equivalents, with a reported 60.7% average conversion-to-safe rate.
In brieflapanda: A matrix-free differentiable solver for nonconvex constrained optimization layers
SummaryThis paper proposes lapanda, a matrix-free differentiable solver for nonconvex optimization with general constraints, so it can be trained end-to-end inside neural networks. It reformulates the problem as a sequence of augmented Lagrangian subproblems solved by a proximal averaged quasi-Newton method with adaptive linesearch. The authors evaluate it on constrained Rosenbrock benchmarks, imitation learning for constrained optimal control, and robotic obstacle-avoidance tasks, reporting lower computation time and memory use than state-of-the-art differentiable solvers while maintaining constraint satisfaction and learning performance.
In briefCounting Wingbeats from Event Data Under Camera Motion: A Controlled MuJoCo Benchmark
SummaryThis paper presents a controlled MuJoCo simulation benchmark for counting completed wingbeat cycles using only event-camera observations while the camera moves. It compares motion-aware training with event-background compensation against ridge, Fourier, autocorrelation, and adapted EEPPR baselines. Under moderate motion, motion-aware training substantially reduces count error, while the tested compensation increased it, and a Fourier baseline sometimes performed best. The authors state that the findings apply only to this small synthetic benchmark and do not establish real-sensor or aerodynamic flight performance.
In briefStoryEngine: A state-tracking AI framework for consistent video storytelling
SummaryThis paper presents StoryEngine, an agentic framework for generating long-form multi-shot videos. It tracks entity placement and story states, propagating event-driven changes so each shot has defined start and end states, and uses canonical references and a bounded repair loop to fix inconsistencies. The authors built a benchmark across diverse scenarios and visual styles to evaluate storytelling quality, narrative coherence, and visual consistency. Reported experiments indicate it outperforms state-of-the-art methods on all evaluated dimensions.
In briefUnified view of continual learning dynamics linking data attribution, forgetting, and plasticity loss
SummaryThe paper studies language models that keep updating over their lifetime and decomposes how learning from one token changes another prediction, token by token and layer by layer. It separates the softmax force, shared readout geometry, and residual connections, yielding two interaction channels and a forward-computable approximation. The authors argue positive interactions identify useful data, while negative ones cause concentrated collisions or accumulated erosion, and over time updates reshape the geometry that carries future learning signals. Data attribution, forgetting, and plasticity loss are framed as regimes of the same evolving dynamics.
In briefParaAgent: Reinforcing Parallel Acting in Open-World Tool Environments
SummaryThis paper introduces ParaAct, a structured parallel-action loop that coordinates phase-level exploration and execution with action-level parallelism for language model agents. ParaAgent learns this loop through multi-agent cold-start demonstrations and reinforcement learning with multi-level advantage decoupling, supervised by step, phase, and trajectory rewards. On two open-world tool benchmarks, ParaAgent-4B reports the best average success among all baselines, including GPT-4.1 systems, with the largest gains on multi-tool tasks.
In briefSpatialSpeak: Question-answering pretraining for multi-view 3D spatial reasoning
SummaryThis paper proposes SpatialSpeak, a two-stage framework that helps vision-language models reason about space across multiple views. Stage I pretrains the model through text-based question answering on local 3D point queries and global object-center queries, while Stage II trains spatial chain-of-thought reasoning with visual compensation. The authors report state-of-the-art results on ReVSI, VSI-Bench, and SPAR-Bench, with a ReVSI score of 62.8.
In briefEAT: An Expert Account Tracker for Faster Mixture-of-Experts Inference
SummaryThis paper introduces EAT (Expert Account Tracker), a method that speeds up Mixture-of-Experts (MoE) models by activating fewer experts during inference. It uses history-aware metrics and adaptive thresholds to dynamically choose the most important experts while aiming to preserve performance. Experiments report an average reduction of over 25% in activated experts compared with the vanilla method, along with better token generation speed than the Top-P baseline across several models and datasets.
In briefQuizzing the Translation: A Prover-Grounded Metric for Natural Language to First-Order Logic
SummaryThe paper addresses evaluating translations of natural-language problems into first-order logic, where small errors can invalidate reasoning. It notes that common metrics like BLEU and BERTScore often rate badly broken translations higher than less broken ones. The authors propose SIV, which uses a theorem prover to check candidate translations against positive and contrastive probes derived from the target formula. SIV's scores track error severity closely and yield labeled error traces.
In briefAnguinus Sculpturae: Synthesizing Peak-Enhancement Breast MRI Without Contrast Agents
SummaryThis paper addresses the MAMA-SYNTH challenge of predicting peak-enhancement breast DCE-MRI from pre-contrast images, aiming to reduce reliance on gadolinium-based contrast agents. The authors show that under a straight-line path the generative process reduces to a single step: the synthetic image equals the pre-contrast image plus a predicted enhancement map. Their pipeline uses nnU-Net segmentations to guide two generators whose outputs are blended region-wise. On the held-out Duke subset of MAMA-MIA, the model achieves the best FRD and Dice among evaluated variants.
Benjamin Hamm, Nico Albert Disch, Maximilian Rokuss et al.
In briefRecovering anomaly supervision for time series detection using context-matched normal-anomaly pairs
SummaryThis paper addresses time series anomaly detection, where labels are scarce and anomalies depend heavily on context. The authors propose Context-Anchored Pair Supervision (CAPS), which learns from simulated normal-anomalous pairs to recover supervision without target-domain anomaly labels. CAPS generates context-anchored counterparts of target trajectories to train a discriminative detector. Experiments on nine datasets report the strongest aggregate performance across four metrics compared with the baselines.
In briefYou Only Edit Once: Local Demonstration Editing to Improve LLM In-Context Learning
SummaryThis paper addresses how to choose demonstration sets for in-context learning in large language models, which is costly to search exhaustively. It proposes local demonstration editing (LDE), which applies a single structured edit (keep, delete, or replace) to an initially retrieved set. A small model, Jev-LDE, is trained with reinforcement learning with verifiable rewards and works as a plug-and-play module. The authors report consistent ICL gains on standard classification benchmarks and transfer to held-out benchmarks and models without retraining.
In briefLearning reaction mechanisms that transfer to new chemistry from visual chemical knowledge
SummaryThe paper introduces MechaVLM, a visual framework that combines transferable chemical representations with external mechanistic knowledge to predict step-by-step reaction mechanisms. It retrieves precedents from 70,384 literature mechanism figures and generates executable electron edits with an atom-indexed language decoder. The authors also release MechBench, a benchmark of 2,184 mechanisms and 9,146 elementary steps, and report strong zero-shot prediction performance.
In briefHierarchical Response Preservation for Continual Adaptation of Zero-Shot Graph-Text Models
SummaryThis paper studies how pretrained graph-text models can learn new classes over time without losing their zero-shot transfer ability or memory of earlier tasks. The authors propose Hierarchical Response Preservation (HiRP), which keeps each historical class probability, sums the new-class probabilities, and uses the resulting geometry to constrain updates. Across three class-incremental settings, HiRP improves average accuracy by 1.84 to 7.95 percentage points over the strongest baseline in each setting, while reducing zero-shot transfer degradation.
In briefDeep-learning waveform design for short-blocklength integrated sensing and communication
SummaryThis paper proposes an autoencoder-based framework for designing ISAC waveforms, which share one signal for communication and sensing, in noncoherent settings. The authors derive a modified Cramér-Rao bound for multi-target delay estimation and analyze maximum-likelihood decoding for noncoherent communication under correlated fading. The autoencoder learns waveforms that jointly optimize both functions, with a tunable parameter controlling their trade-off. Simulations indicate the design outperforms conventional schemes in communication reliability and sensing accuracy, particularly under short blocklength and fading conditions.
Muah Kim, Shuangyang Li, Tayyebeh Jahani-Nezhad et al.
In briefViCoR: Automatically verifying and fixing molecular structures recognized from chemistry images
SummaryThis paper studies Selective Structure Recognition, which automatically produces reliable molecular structures from chemistry literature while rejecting unresolved cases. It proposes ViCoR, a framework that iteratively verifies and revises predicted structures by aligning them spatially with source images and mapping discrepancies to executable graph edits. On two real-world benchmarks, overall accuracy improves substantially over the baseline recognizer.
In briefJustQuant: 4-Bit Activation Quantization Without Smoothing, SVD, or Rotation
SummaryThis paper proposes JustQuant, a framework that achieves 4-bit activation quantization using only plain low-bit operators by shifting complexity into training. It introduces Theseus QAD, a quantization-aware distillation method that progressively applies multi-level supervision. Experiments on DiT and diffusion large language models show that the method improves low-bit quantization quality, with its effect differing between smaller and larger models, while avoiding the complex operators required by many existing PTQ methods.
In briefPoE-Fuse: a precision-weighted expert fusion method for understanding changes between two satellite image dates
SummaryPoE-Fuse compares two satellite images of the same place taken at different times to locate and describe what changed, with uses in disaster response and environmental monitoring. It keeps geometry, grounding and language foundation models frozen and trains only a lightweight fusion trunk that weights their features by learned per-cell precision. A single shared trunk handles all three tasks, reaching a mean F1 of 59.2%, versus 40.7% for an instruction-tuned temporal vision-language assistant and above dedicated change-detection models retrained under the same protocol.
In briefBeyond One-Step Accuracy: A State-Affine Latent Model (SALT) for More Reliable Visual Planning
SummaryThe paper notes that visual planning world models are usually trained for one-step prediction, while planning applies them recursively, so one-step accuracy does not reflect how errors accumulate. The authors show that state-affine dynamics have state-independent Jacobians, making error propagation depend only on the action sequence, and use this to propose SALT, trained with recursive multi-step rollout supervision. Across four visual planning environments, SALT has higher one-step prediction error than the matched LeWM baseline but improves closed-loop success by 10.0 percentage points on average.
In briefLoopLUT: Real-time 4K image enhancement using region-by-region refined 3D lookup tables
SummaryThe paper proposes LoopLUT, a cascaded 3D lookup table method for real-time color enhancement of 4K images. A global LUT first applies an overall correction, then loop iterations predict at low resolution which regions still need correction and build residual LUTs from those regions' color statistics. Cascaded gates form a partition of unity, fusing the lookup results per pixel without a separate fusion module. Experiments on four benchmarks report PSNR gains of up to 2.81 dB over the strongest prior method while maintaining real-time 4K throughput.
In briefPretraining Transformers with quantized softmax in attention, a study of how approximating the softmax step affects model training
SummaryThis paper studies how approximating the softmax in attention with low-precision quantization affects Transformer pretraining. It introduces K-interval attention, which approximates the exponential using K+1 grid values, and compares choices such as per-row calibration, interpolation versus hard rounding, and the placement of a straight-through surrogate. Experiments matched on model, data, and optimizer report that these choices change the validation loss gap relative to standard softmax, with the best reported setting at 124M parameters and 2.5B tokens showing gaps of +0.004 to +0.019 nats.
In briefTGRL: Temperature-grouped reinforcement learning for more efficient exploration in LLMs
SummaryThis paper proposes Temperature-Grouped Reinforcement Learning (TGRL), which uses the diversity created by sampling temperature as an explicit training signal for large language models. For each prompt, it splits rollouts into low- and high-temperature groups, estimates the exploration gain from their reward contrast, and assigns this signal to tokens using Jensen–Shannon divergence between temperature-scaled next-token distributions. The authors report that TGRL reaches equivalent accuracy up to 36% faster than strong RLVR baselines without a larger rollout budget, and improves results across 11 benchmarks.
In briefApproximating Softmax in Pretrained LLMs for Faster Attention Kernels
SummaryThe paper studies how accurately softmax must be computed at inference in ten frozen decoder-only Transformer models (0.5B to 72B parameters). It finds the number of positions receiving probability and within-row resolution can be reduced substantially, but uniform weighting of those positions is harmful, with resolution near the row maximum mattering most. Based on this, it proposes Rowmax-PoT, a coarse logarithmic weight representation anchored at each row maximum, and Rowmax-H15, a hardware-specialized version in FlashAttention-4. On NVIDIA B200, the patched FP8 attention forward runs 12.4% faster at causal 8K and 25.8% faster at non-causal 8K in host-side latency measurements.
In briefIVT-Guard: A Unified Reasoning Model for Detecting AI-Generated Images, Videos, and Text
SummaryThe paper notes that existing AI-generated content detectors mostly handle only one modality and output binary verdicts without reasoning. The authors introduce IVT-Set, a dataset of over 152K image, video, and text samples with multi-granularity chain-of-thought reasoning, and propose IVT-Guard, a framework for interpretable detection across all three modalities. A three-stage training paradigm is also presented, and the authors report state-of-the-art detection results in in-domain, out-of-domain, and cross-dataset settings.
In briefFill2SR: Adapting Inpainting Diffusion Transformers for Real-World Super-Resolution
SummaryThis paper presents Fill2SR, which repurposes a masked-inpainting Diffusion Transformer for image super-resolution without extra spatial branches. Its IIEA module writes the low-quality image into the native masked-image slot and is trained with LoRA-only tuning. The paper also introduces RCDT, an offline pipeline that transfers degradation descriptors from unpaired real images onto clean targets.
In briefForeFly: A dual-horizon world action model that helps drones follow language instructions while navigating in the air
SummaryThis paper presents ForeFly, a model for Aerial Vision-Language Navigation that lets drones follow instructions over long flights in complex 3D spaces. It predicts two kinds of future: a nearby future for local continuity and an adaptive route-critical future for long-range guidance. A new method, Foresight-Guided Action Refinement (FGAR), uses each prediction differently to improve actions. Experiments on the TravelUAV and UAV-ON benchmarks show ForeFly consistently outperforms strong baselines in seen and unseen settings.
In briefCompressing predicted values for learning-augmented online task scheduling
SummaryThis paper studies whether predicted dual values for metrical task systems can be compressed to a small set of landmark states. Predictions are given only at landmarks, and other values are reconstructed via Lipschitz extension. The authors prove an excess cost bound depending on the landmarks' covering radius, a matching lower bound for sparse landmarks on finite lines, and PAC guarantees for learning compressed prediction tables.
In briefOpenFC: Learning Verification Policies for Open-Search Fact Checking
SummaryOpenFC is a training framework that turns Qwen3-8B into a compact controller deciding each step of fact verification: reasoning, evidence gathering, and when to stop. It uses two stages, Stepwise-Calibrated Cold Start to build reliable trajectories for supervised fine-tuning, then Verification-Aware Reinforcement Learning to refine the policy on unresolved claims. The authors report 70.39% average accuracy and 63.30% macro-F1 across six fact-checking benchmarks, and say both stages contribute complementary gains.
In briefHiLoRe: Deciding What to Store, Compress, or Recompute in Efficient GRPO Training
SummaryThe paper notes that fixed gradient checkpointing schedules in GRPO training can leave substantial GPU memory unused. It proposes HiLoRe, which uses the current GRPO loss coefficients to estimate each activation state's approximation sensitivity and allocates states among high-precision storage, low-precision compression, and deterministic recomputation. Across five model-task settings, it reports actor-update throughput gains up to 13.5% over gradient checkpointing, with paired mean downstream-score differences below 0.6 percentage points.
In briefCorrect-Then-Forecast: Observer-Based State-Space Models for Time Series Prediction
SummaryThe paper proposes Observer State-Space Models (OSSMs), which treat an observed input time series as measurements of an underlying autonomous dynamical system. A single transition governs state propagation across both context and forecast intervals, while available observations correct the estimated state through an observer. The authors note that classical control-theoretic properties such as observability and convergence of estimation error follow naturally, and that conventional and recent state-space models can be recovered as special cases. Experiments on several benchmarks, with the same parameter count and training setup as the corresponding baseline, report substantial improvements.
Alexis-Raja Brachet, Guillaume Clavier--Fr\'emond, Abdelhakim Ziani et al.
In briefDr. Free: Self-Evolving Search Agents Without Difficulty Rewards
SummaryThis paper presents Dr. Free, a self-evolving framework for training search agents without data, which removes difficulty-based rewards for question proposers. It samples relational chains from a knowledge graph, pairs them with aligned passages, and rewards questions only when the target answer is more likely under full evidence than under any shortcut context. Across seven open-domain QA benchmarks, it outperforms prior data-free search agents and a supervised baseline, with large gains on multi-hop QA, while reducing proposer training time by over 7 times.
In briefMA-JEPA: Joint-Embedding World Models for Multi-Agent Reinforcement Learning
SummaryThis paper introduces MA-JEPA, a stochastic world model for multi-agent reinforcement learning that predicts target representations instead of reconstructing observations. It uses centralized training with decentralized execution, including a training-only joint predictor and a centralized critic for value learning. On the SMAC benchmark, it matches or exceeds the strongest reported comparator mean win rate on four of eight evaluated maps.
Brandon Gary Kaplowitz, Osaze James Obahor, and Christian Schroeder de Witt
In briefTreats out-of-distribution generalization as a bifurcation problem and guides sampling only in a critical time window
SummaryThis paper studies why continuous-time generative models often fail to synthesize unseen combinations of concepts. The authors find that compositional binding happens within a localized temporal window of the diffusion trajectory. They propose surgical guidance, which applies exact multi-pass decomposed scoring only within that window to reduce computational cost. On an OOD bi-digit MNIST testbed, intervening in the first 15% of the trajectory yields a 5.3% absolute pairwise accuracy gain over the joint baseline.
Nguyen-Thanh-Luong Doan, Quang-Vu Nguyen, Tang-Phu-Quy Le et al.
In briefGraphSelect: choosing fewer text and image inputs for multimodal graph prediction
SummaryThis paper studies how to pick a limited subset of text and image representations, with a separate capacity for each modality, while preserving the predictions of multimodal graph models. The authors propose GraphSelect, which starts from individual candidate gains and refines the subset through jointly evaluated exchanges that reduce prediction loss. On six graphs it recovers more of the objective than six adapted attribution and explanation methods. Across nine trained architectures on two graphs, keeping 20% of candidate representations per modality yields a mean accuracy drop of 0.10 percentage points.
In brief"Fine Until Fine-Tuned": A paper showing that repeatedly training on the same solutions makes later-trained models' reasoning fragile
SummaryThe paper examines recipes like s1 and LIMO that teach reasoning from small datasets by repeating the same worked solutions. Repetition looks harmless at first, but later training, even unrelated to reasoning, makes the model's reasoning fragile, dropping accuracy from about 95% to 86.0% and to 59.3% or below under harsher stages. Fresh solutions or replaying a small share of originals prevented the damage, and most of it could be restored with brief reasoning training.
In briefPGL-3D: Progressive Geometric Learning for 3D Visual Query Localization
SummaryThis paper addresses 3D Visual Query Localization, which finds the latest contiguous occurrence of a queried object in an RGB point-cloud sequence and predicts a 9-DoF cuboid for each response frame. The proposed PGL-3D framework predicts, selects, refines, and re-predicts, using intermediate cuboids to guide search-evidence aggregation and update query and proposal features. It reports a mean stAP of 0.270, compared with 0.044 for LaF, and includes ablations and a reproduction result on GSOT3D.
In briefSparsity by Default: Theory and Practice of ARD for Variable Selection in Gaussian Process Regression
SummaryThis paper reviews automatic relevance determination (ARD), a common way to select input variables in Gaussian process regression. Each input gets its own lengthscale, learned by maximizing the marginal likelihood, so irrelevant inputs receive very large lengthscales and are effectively switched off. The authors analyze this pruning mechanism, survey practical algorithms and asymptotic theory, compare ARD with spike-and-slab and shrinkage priors, and discuss its limitations and remedies.
In briefCalibrated derivative sensitivity for selecting inputs in Gaussian-process regression
SummaryAutomatic relevance determination, the standard variable-selection tool for Gaussian-process regression, ranks inputs by inverse lengthscales, which reflect how fast a function varies rather than how much an input affects predictions, and it has no calibrated rule for choosing inputs. The paper proposes selecting inputs by derivative sensitivity and calibrates the GP derivative process with a studentized multiplier bootstrap. The authors prove asymptotic family-wise and false-discovery-rate control, and in 100 replications the rule controls FDR for truly null inputs, whereas uncalibrated derivative rankings and a Bernstein-von Mises cutoff exceed the target.
In briefTerMeZO: A sparse zeroth-order method for fine-tuning BitNet models on edge devices
SummaryThis paper addresses the high memory cost of fine-tuning BitNet language models, whose ternary weights still require updating full-precision latent weights. The proposed TerMeZO method uses the geometry of the ternary quantizer to select the latent weights most likely to change during fine-tuning, without extra data or memory. The authors provide a convergence analysis and experiments on BitNet models from 1B to 3B parameters across classification, instruction-following, and math reasoning. TerMeZO matches or exceeds full-parameter MeZO while substantially reducing fine-tuning memory use.
Houssem Sifaou, Prabodh Katti, Bipin Rajendran et al.
In briefOSCC: Certified safe coupling to reduce gradient noise in imperfect-information learning
SummaryThis paper proposes observation-safe counterfactual coupling (OSCC), a framework that defines which coupled rollouts are admissible in imperfect-information environments, based on marginal preservation and information-state safety. It introduces OSCC-Select, a calibration-based selector that chooses among independent, root-only, continuation-only, and fully coupled sampling. On Leduc comparisons, the fully coupled CP-GRPO reduces return-contrast variance from 41.1158 to 18.1441, a 55.87% reduction, and in the three-action setting OSCC-Select attains a lower gradient-noise trace than return-variance selection.
In briefReasoning on the Simplex: Geometric Fixed-Point Models
SummaryThe paper proposes Geometric Fixed-Point Reasoning (GFPR), where the iterated state is a field of categorical beliefs whose argmax is the answer at each step. At about 7M parameters it reports 95.1% on Sudoku-Extreme, 92.0% on Maze-Hard, and 100% on S_5 length 128, above published FPRM results at the same scale. The same update is applied to a 201M language model trained on FineWeb-Edu, which the authors report outperforms GPT-2 small on four zero-shot multiple-choice tasks.
Talgat Daulbaev, Ilya Glazkov, Maxim Rakhuba et al.
In briefJev Matches 7B Models for Speech-Neuroprosthesis Rescoring
SummaryThis paper addresses the final step of a speech neuroprosthesis, where candidate sentences decoded from brain activity are rescored by a large GPU-based language model. The authors frame rescoring as a single typed call that returns a probability for every candidate, served by the hosted model Jev, and combine it with the decoder's score. On 978 held-out sentences from a participant with ALS, Jev matches or slightly beats OPT-6.7b and Qwen2.5-7B on word error rate without requiring a GPU.
In briefDoes execution need full-precision KV cache? A mixed-fidelity runtime for LLM serving
SummaryLLM serving is limited by the GPU memory used by key-value caches, and existing runtimes only execute with the configured target KV representation. This paper presents ElasticKV, which adds a compact intermediate KV state so that fidelity becomes a runtime-managed property rather than a gate on execution. It combines a pair-structured paged layout, a dual-mode attention backend that consumes the compact state, and pressure-aware fidelity management. Under high concurrency, the authors report 3.8-4.0x lower time-to-first-token and 9.1x lower P90 TTFT than vLLM while preserving generation quality.
In briefNearly Group-Separable Elections: Measuring How Close an Election Is to Group-Separable
SummaryThis paper studies how close a given election is to being group-separable, measuring distance by swaps of adjacent candidates in the votes. It also covers related domains such as caterpillar group-separable, balanced group-separable, single-peaked, and single-crossing elections. The problem is generally intractable, but the authors give practical FPT algorithms parameterized by the number of candidates or swaps. The swap-parameterized algorithm applies to all domains defined by finite forbidden subelections, and experiments support the findings.
Piotr Faliszewski, Jan Jabrocki, Stanis{\l}aw Ka\'zmierowski et al.
In briefManiEdit: Sequentially Editing Unstructured Knowledge in Language Models via a Manifold View
SummaryLarge language models sometimes produce incorrect or outdated content, so they need efficient ways to update knowledge continually. This paper frames knowledge editing as moving a local edit sub-manifold within the global knowledge manifold, and proposes ManiEdit, which combines pivot localization with manifold-aware preservation. The authors report state-of-the-art results on two base LLMs and four unstructured editing benchmarks, while keeping general capabilities near their original level across six downstream tasks.
In briefScaling laws for transformer-based operator learning in physics
SummaryThis paper develops a theoretical framework for the approximation and generalization errors of transformer-based operator learning, which is used to learn solution operators of physical systems. It derives a universal approximation error for Hölder-regular operators and establishes a power-law relation between prediction error and training data size. The scaling exponent depends on input and output dimensions, function regularity, and the intrinsic dimension of the input function class. Numerical experiments confirm the predicted power-law scaling and show the convergence rate varies with intrinsic dimension.
In briefE-CONAN: An Arabic NLU Diagnostic Dataset for Linguistic Phenomena
SummaryThis paper addresses the lack of standard naming and linguistic phenomena categories for error analysis in Natural Language Understanding (NLU), proposing an initial hierarchy for cross-lingual NLU error analysis and a method to build an NLI hierarchical framework, applied as a case study to Arabic. The authors manually annotate and release the freely available E-CONAN diagnostic dataset, which they use to evaluate 9 pretrained language models and 5 LLMs. Results show LLMs outperform pretrained models on world knowledge and commonsense reasoning but underperform on syntax. Reasoning is the hardest phenomenon for all models, and syntactic phenomena are easiest for pretrained models.
In briefAutomatically finding hidden symmetries in neural network parameters
SummaryThis paper presents a framework for systematically discovering symmetries in neural network parameter spaces, which matter for understanding loss landscapes, training dynamics, and generalization. The authors define data-dependent parameter symmetries through infinitesimal conditions and learn group generators and nonlinear action maps jointly. The framework uncovers symmetries across various architectures, including pretrained transformer models, among them previously unknown ones.
In briefIn-Token Learning: Diffusion Transformer Framework for High-Fidelity Image Restoration
SummaryThe paper presents In-Token Learning, a framework that adapts a pretrained diffusion transformer to restore degraded images using conditional rectified flow matching. Degraded-image tokens are fused with latent tokens along the channel dimension, and the model is conditioned through a frozen degraded-image embedding and a fixed task prompt, without a ControlNet-style branch. The authors evaluate super-resolution, denoising, and colorization on several datasets, reporting competitive fidelity and perceptual quality but weaker generalization on RealLQ250. They also demonstrate QHD and tiled 12K restoration, while noting that attention cost still grows with resolution.
In briefLLM4Trust: Exploring Large Language Models for Trust Evaluation
SummaryThis paper introduces LLM4Trust, the first benchmark framework that systematically tests how well large language models can evaluate trust. The authors build trust graphs modeling five basic trust properties and assess eight LLMs under nine prompt methods, then validate the best combinations on five real-world datasets. The results show LLMs understand basic trust properties well, but they are vulnerable to attacks and costly to run, so the authors propose a defense mechanism and batch inference.
In briefAI-assisted collaborative assembly line inspection: integration and deployment challenges
SummaryThis paper describes an AI-assisted collaborative inspection cell designed and deployed at the Silverline kitchen-appliance factory. The cell combines two cobots, machine-vision defect detection, and multi-modal data logging with real-time visualization. The authors report deployment issues such as close-proximity safety, AI robustness under glare, and ROS 2 namespace collisions, along with the solutions adopted. After deployment, per-unit quality-check time fell from 82 s to 61 s, and operator mental demand was significantly lower (p = 0.005, NASA-TLX).
Asya \"Unal, Amr Okasha, Ege \c{C}{\i}rakman et al.
In briefAnatomy-structured hierarchical multiple instance learning for weakly supervised chest X-ray disease detection
SummaryThis paper addresses detecting thoracic diseases in chest X-rays using only weak, image-level labels. It proposes ASH-MIL, which uses parallel cardiac, pulmonary, and agnostic branches with hierarchical multiple instance learning, injecting anatomical priors as soft spatial biases into decoder cross-attention. Localization is derived from MIL-weighted attention maps without bounding-box supervision. The authors report consistent gains over prior weakly supervised and anatomy-aware methods on CXR8 and cross-domain MIMIC-CXR, especially under stricter localization criteria.
In briefSceneScaffold: Actively Building Structured Scene Representations for 3D Understanding
SummaryExisting 3D multimodal models compress scenes into flat, object-centric tokens, which weakens spatial relations and makes reasoning unstable in cluttered or ambiguous scenes. This paper proposes SceneScaffold, which builds a role-aware spatial scaffold before language reasoning, with entity states, scene-frame states, relation states, and a global summary. The method is evaluated on 3D visual grounding, question answering, and dense captioning, with diagnostic results on relation-intensive and spatially ambiguous cases. Code is publicly available.
In briefTRACE: Managing Memory Validity When Agents Rejoin Changing Teams
SummaryThis paper presents TRACE, a training-free layer that decides which of a returning agent's saved memories are still valid to act on after the shared state has changed. It treats re-entry as an eligibility decision, reconciling the departure checkpoint with absence-period updates and releasing a bounded Return View only when it covers the role's open obligations. Evaluated on Memora, STALE Type II, and a derived ManBench-Return setting, TRACE is reported to keep valid memory while rejecting stale memory, with 92.6–98.3% valid-information availability and 98.4–99.5% invalid-information rejection on ManBench-Return.
In briefPPG-LM: A Photoplethysmography-Language Model with Multi-Level Clinical Alignment
SummaryThis paper introduces PPG-LM, a family of models that align photoplethysmography (PPG) waveforms with language, learning from both signal-derived physiology and clinical context in electronic health records. The authors generate segment-, event-, and visit-level captions automatically and train in two stages using contrastive learning, waveform-conditioned captioning, and time-aware alignment. Pretrained on about 73k hours of PPG, the model supports recognition, cross-modal retrieval, and captioning, and shows gains on MC-MED, MIMIC-III, and VitalDB.
In briefSLP-ProbHard: Probabilistic Learning with Hard Constraints via Structural Latent Parameterization
SummaryThis paper introduces SLP-ProbHard, a framework that generates probabilistic predictions satisfying exact constraints by using a structural feasible latent parameterization, rather than correcting predictions after the fact. The method is tested on affine constraints, ordering and simplex constraints, nonlinear manifolds, and hydrological flow-duration-curve data. The authors find that exact feasibility alone does not determine the predictive distribution, so direct structural generation is preferred when meaningful feasible coordinates are available.
In briefPrintable adversarial decals for aerial vision, tested with measured ink colors and printed patches
SummaryThis paper optimizes adversarial patches under three physical constraints: printer color gamut, the flat panel size of a vehicle, and detail loss at altitude. Using a gamut measured from a printed chart improves attack success on some closed-set segmentation models but not others, and palette chromaticity alone does not explain the difference. A physical test with printed decals detected no transfer and bounded the transferred rate at 0.133.
Sandesh Shrestha, K. T. Yasas Mahima, Asanka G. Perera
In briefSource Anchoring: Keeping Generated Physics Samples Consistent and On-Distribution
SummaryThis paper proposes SAPC, a flow matching method that encodes physical constraints into the source noise before generation rather than correcting samples afterward. It is evaluated on five partial differential equation systems covering six tasks, against five baselines and an unconstrained backbone. SAPC reproduces the target distributions most accurately on every task while matching the constraint precision of the best projection-based methods.
In briefWhat happens when an AI research agent works alone after the user leaves? This paper studies how initial user details shape its search process and final recommendations.
SummaryThe paper introduces DRaligned, a counterfactual evaluation framework built on PDR-Bench, which changes one task-relevant user factor at a time and compares acquisition requests, drafts, and final reports. Experiments find that agents follow a largely shared research process but allocate requests differently, and their final recommendations distinguish user conditions more clearly than explicit requests do. These directional differences recur across tested agent models, execution harnesses, and evaluator models.
In briefWhen Evidence Changes the Subject: Subject-Typed Claim Licensing for Learned Routing
SummaryThe paper argues that benchmarks of composite learned systems measure the whole system, while scientific claims may concern only one component, creating an attribution problem. It proposes subject-typed claim licensing, which separates weaker conclusions about the requested subject from positive but non-substitutive credit about another subject. The approach is instantiated as SCOPE-Routing for preference-conditioned multigraph routing, with reported results from non-author reproduction, held-out review, and a controlled routing study.
In briefRelaxKV: Query-guided recomputation for efficient reuse of KV cache in RAG
SummaryThis paper addresses KV cache reuse in Retrieval-Augmented Generation, where cached states from independent chunks miss cross-chunk interactions. It proposes RelaxKV, which jointly allocates repair targets and recomputation context, using the user query to pick layer-specific targets and restrict recomputation to query-relevant context to reduce attention cost. Across four decoder models at a 15% anchor ratio, RelaxKV improves aggregate LongBench performance over ProphetKV, and on Qwen3-14B it gives a better quality-TTFT trade-off.
In briefPulseFlow: counterfactual PPG generation via latent transport for rhythm editing
SummaryThis paper addresses the scarcity of labeled atrial fibrillation recordings in photoplethysmography (PPG) data. It proposes PulseFlow, a framework that conditions on a source recording and combines conditional representation learning with invertible latent transport to edit cardiac rhythm while retaining source information. Experiments on two clinical cohorts show effective rhythm transformation, measurable source correspondence, and improved AF classification under limited labels.
In briefThe cost of useful natural gradient updates: knowing the direction is not enough, step size drives the cost
SummaryThe paper asks what information is needed to turn a natural-gradient direction into a useful finite update under a population KL budget. It constructs a four-state exponential family where laws share the same gradient, Fisher information and natural gradient, yet their useful-step sets are disjoint, so step length carries the cost. It also gives sample complexity bounds, NP-hardness results, and experiments on frozen-feature classifier heads.
Subhransu S. Bhattacharjee, Dylan Campbell, Rahul Shome
In briefEfficient AlphaZero-style chess training on limited compute: search, replay, and throughput
SummaryThis paper asks how strong an AlphaZero-style chess system can become under limited training compute when its learning loop is engineered for efficiency. A 6.32-million-parameter model was trained from random initialization over 2.5 days on a single eight-GPU node, reaching 3,251 benchmark Elo against a fixed-node Stockfish 13 ladder. The authors study search allocation, replay, policy representation, progressive model sizing, quantized inference, and throughput engineering, and attribute the strength to the integrated system rather than any single choice.
In briefHamiltonian JEPA: World Models That Plan Actions Using a Separate Control State
SummaryThe paper proposes H-JEPA, which separates a perceptual code from a control state that is a fixed orthonormal slice of that code and evolves under phase-conditioned dissipative port-Hamiltonian dynamics. It introduces port-inverse consistency (PIC), a parameter-free reweighting of prediction error along the port directions. H-JEPA matches or exceeds reconstruction-free baselines on four pixel-based control benchmarks, with its largest gain on OGB-Cube.
In briefChameleon: A Dynamic Number-Format Adapter for Efficient Diffusion Model Quantization
SummaryThis paper proposes Chameleon, a post-training quantization method for diffusion models that keeps bit-width fixed while choosing the number format per weight channel and per activation tensor. Activation formats are selected from cheap statistics, and weight formats are selected by reconstruction error. On SDXL, SDXL-Turbo, and PixArt-α evaluated on COCO-2014, the authors report the best FID in all six backbone and bit-width settings.
In briefLLMs Trust Their Own: Identity-Dependent Conformity in Multi-Agent Systems
SummaryThis paper studies whether large language models change how much they conform to wrong answers depending on the social identity of other agents in a multi-agent setting. The authors tested 12 open-weights models on nine judgment tasks with a single correct answer. They found that in-group consensus increases conformity while out-group consensus decreases it. The authors conclude that group identity shapes how models aggregate information independently of correctness.
In briefFedMHAR: Federated Multimodal Activity Recognition Using Multi-Agent Reinforcement Learning
SummaryThis paper proposes an adaptive, cost-aware framework for recognizing human activities from multiple wearable sensors. It formulates multimodal fusion as a cooperative multi-agent reinforcement learning problem, with one PPO-based agent per sensor modality learning per-sample fusion weights. The framework is extended to federated learning as FedMHAR using a server-side policy called BiFL-PPO, and reports accuracies of 87.30% and 94.98% centrally and 79.74% and 77.49% federated on the MEx and UTD datasets.
In briefDomain-adapted diffusion models for testing whether two variables are conditionally independent
SummaryThis paper studies conditional independence testing, where existing generative-model-based tests accumulate estimation errors and lack guaranteed asymptotic Type I error control. The authors treat conditional generative modeling as a domain adaptation problem and propose DA-Diff, a multi-source diffusion framework, along with the DA-CIT test built on it. Experiments reportedly show DA-Diff improves conditional generation quality over transfer-learning diffusion baselines, and DA-CIT offers strong Type I error control with competitive power.
In briefPredicting which block-update method trains better using cross-block curvature
SummaryThe paper compares simultaneous (Jacobi) and sequential (Gauss-Seidel) block-gradient updates, which are used in neural network training, federated learning, and LoRA. It derives a one-step loss difference from cross-block curvature and a signed loss comparison after K iterations with an error bound. Tested on observed training trajectories over 500 iterations, the theory identifies the lower-loss method in 98.0%, 83.4%, and 97.6% of iterations, rising to 100.0%, 93.2%, and 99.6% when the loss recursion uses measured parameter differences.
In briefWhich masking design works best for EEG foundation models?
SummaryThis paper systematically compares spatio-temporal masking strategies for pre-training EEG foundation models. The authors trained 58 models under MAE and JEPA self-supervised frameworks and evaluated them with a linear probe on 12 OpenEEGBench datasets. Both frameworks agree on an optimal masking configuration and shared failure modes, and a new JEPA-specific failure mode is identified. The authors report that with a well-chosen mask, their pipeline matches REVE-level downstream performance at a fraction of REVE's pre-training compute.
Pierre Guetschel, Bruno Aristimunha, Yassine El Ouahidi et al.
In briefAudio explainability: how masking fill shifts model outputs, shown via logit trajectory decomposition
SummaryThis paper examines perturbation-based explanation methods for audio classifiers, which usually assume that the mask fill used to occlude spectrogram regions has negligible effect. The authors decompose output changes in logit space into an on-axis component along the line between the occluded and original spectrogram, and an off-axis component perpendicular to it. Across three fills, three classifiers and 1,100 AudioSet clips, they report that 41-77% of displacement under partial masking is off-axis, with a stable, model-and-fill-specific direction.
Nico Garc\'ia-Peguinho (1), David Kelly (2), Fabrizio Smeraldi (1) et al.
In briefDISCO: Splitting Long-Text Search and Reasoning Across Models
SummaryThe paper describes why large language models often reason poorly on very long inputs, a problem it calls context rot. It proposes DISCO, in which many worker LLMs locate relevant information in parallel while a driver LLM, trained with reinforcement learning (GRPO), plans and synthesizes the final answer. The authors report 78.4% accuracy on RULER-QA at 1M tokens, up to 9.8 points above full-context models on LongBench v2, and over 80% lower inference cost.
Guanzheng Chen, Viet Dac Lai, Subhojyoti Mukherjee et al.
In briefIs Word Error Rate Enough? Entity-Aware Metrics for Speech Privacy Evaluation
SummaryAs smart devices capture more speech, sensitive content may leak, so this paper studies how well obfuscation techniques protect speech, especially named entities. It adapts entity-aware privacy metrics from natural language processing to speech privacy and evaluates two obfuscation methods under several attack scenarios. The authors find that fine-tuning on entity-rich data improves attacks for some entity categories but not others, and they offer guidance on metric selection depending on whether the obfuscation preserves temporal alignment.
Anjana Rajasekhar, Jule Pohlhausen, Nayana Jacob Alappattu et al.
In briefHow Synthetic Labels Help Conformal Prediction Cover Uncertainty Across Inputs
SummaryConformal prediction gives reliable overall coverage, but with little calibration data it struggles to capture how uncertainty varies across inputs. This paper studies whether cheap synthetic labels from domain models or language models can help when only a small trusted sample exists. The authors propose prediction-powered quantile learning, which uses synthetic labels with bias correction from trusted data. Across eight regression benchmarks, it improves conditional coverage while keeping marginal validity and yielding tighter prediction sets, and a human-rating study shows similar gains from external LLM labels.
In briefPrefix caching for hybrid LLMs via suffix replay, letting linear-attention states forget distant inputs
SummaryHybrid LLMs mix full-attention and linear-attention layers, but the recurrent states of linear layers cannot be rolled back to arbitrary prefix positions, which complicates prefix caching. The paper presents SuffixReplay, which keeps only a recent suffix of each linear layer's input hidden states as anchors and replays them to approximate the state at any cache-supported page boundary, avoiding stored recurrent-state checkpoints. It is evaluated on three hybrid models and integrated into SGLang.
In briefGeometric identifiability in predict-then-optimize learning
SummaryThe paper studies the decision-focused convex surrogate SPO+ used in predict-then-optimize learning. It characterizes the equality set of the population risk and shows that, under central symmetry, the centered mean class is the unique Bayes minimizer exactly under a face-crossing condition. The results are population-level and provide no finite-sample or generic transfer-regret guarantees.
In briefFORUM: Combining Frozen Vision-Language Models by Agreement for Visual Grounding
SummaryThis paper presents FORUM, a training-free method that fuses several frozen multimodal large language models at test time to locate objects described in text. It selects the region supported by the most distinct models and returns an actual member box via medoid localization, so one inaccurate prediction cannot shift the answer. Fusing three open models, FORUM outperforms a 397B-parameter published reference by a relative 5% on the adversarial Ref-Adv-s benchmark and a plain averaging ensemble by 15%, with gains also on standard RefCOCO+.
In briefPhysics-guided flow-map matching for short-term rain forecasting from radar
SummaryThis paper proposes PG-FMM, a conditional flow-map model for precipitation nowcasting that forecasts future radar fields from past observations. It separates predictable motion from uncertain small-scale detail, using a frozen Lagrangian advection prior for motion and a flow-map generative head that produces sharp detail in four sampling steps. Tested on four radar benchmarks, it outperforms state-of-the-art methods on 18 of 24 metrics, with the largest gains at heavy-rain thresholds, where the critical success index improves by up to 58.9%.
Shunya Nagashima, Takumi Bannai, Makoto Misaizu et al.
In briefA Cheap Verifier Is Enough: LLM Post-Training Tolerates Imperfect Rewards
SummaryThis paper examines whether agreement between verifiers predicts how well large language models perform after post-training on tasks with semi-verifiable rewards in medical, legal, and finance domains. Using over 11k H100 GPU-hours on HealthBench and PRBench, the authors find that higher verifier agreement does not reliably identify the best training verifier. Expensive verifiers do not consistently beat cheaper ones, and open-weight Gemma verifiers yield strong results.
Andreas Plesner, Curtis Northcutt, Francisco Guzm\'an et al.
In briefGSM: Faster Language Modeling via a Shared Global State
SummaryThis paper introduces the Global State Model (GSM), a causal encoder-decoder architecture that separates gathering historical information from deep computation. The encoder aggregates long-range information into a fixed-size shared state through multi-stage sparse retrieval, and the decoder transforms this state without repeatedly accessing history. Experiments across several model scales report downstream performance comparable to matched DeepSeek-V4.1-Flash baselines, while preserving long-context retrieval and improving training and inference throughput and reducing cache overhead.
In briefRethinking token reweighting in SFT: suppressing, reversing, and extrapolating learned features
SummaryThis paper analyzes token-reweighting methods used in supervised fine-tuning (SFT) and argues they can only suppress or amplify supervised updates, not reverse features already learned. The authors propose SCALE, which freezes the pretrained model and the SFT delta and learns bounded token- and module-specific gates by minimizing predictive entropy alone. These gates suppress, reverse, or extrapolate the frozen SFT features. Experiments on Qwen2.5-Math and Qwen3-4B-Base show higher mathematical-reasoning averages than the strongest baselines and the best average code-generation results.
In briefSphMind: A training-free framework helping vision-language models reason spatially with 360° cameras
SummaryThis paper presents SphMind, a training-free, plug-and-play framework for spatial reasoning on 360-degree panoramic images. It separates semantic perception from geometric reasoning, handling spherical geometry outside the multimodal model using a Spherical Harmonics-based Spatial Graph and inference-time geometric grounding. The authors report average directional reasoning gains above 21.4% on MP3D and Stanford2D-3D, an 8.7% improvement over prompt-engineering baselines on ODI-Bench, and a 5.9% gain in rotational invariance.
Shriram Damodaran, Soumyaratna Debnath, Cheston Tan et al.
In briefProtected cores alone are insufficient: certifying AI-proposed revisions to temporal specifications
SummaryThe paper proposes an architecture in which an untrusted AI suggests revisions to temporal specifications and a symbolic governor decides whether to activate them. It argues that aggregate certification can hide systematic failures on protected triggers, so aggregate and core-conditional post-selection certification must be controlled together. Experiments are controlled synthetic studies using a frozen supervised AI proposer to illustrate this masked-core failure.
In briefWhen does a audio LLM form the concept of 'dog'? A study of where dog-related sound information appears inside the model.
SummaryThe paper studies how the multimodal model Qwen2.5-Omni-7B represents dog-barking information internally. The authors use a Jacobian lens readout to derive a 'dog direction' in hidden states and add or remove it to test effects on responses. Dog-related information is decodable without explicit dog or bark cues, and interventions change some answers, mainly in late layers L22 and L24 just before generation. The dog direction shows no advantage over controls on animal/other classification.
In briefA Free Knob: Separating Calibration From Predictive Skill in Threshold Evaluation
SummaryMany dense-prediction benchmarks score rare events by pooling spatial blocks, thresholding them, and computing the Critical Success Index (CSI), which mixes spatial discrimination with amplitude calibration. The paper proposes a symmetric post-hoc monotone calibration audit (FreeKnob Audit), fitted on held-out data and applied separately to each system. It finds that calibration can substantially change CSI comparisons, and recommends reporting pooled frequency bias alongside this audit.
Md Tanveer Hossain Munim, Bijoy Ahmed Saiem, Al-Amin Sany et al.
In briefSharp training-conditional coverage of conformal prediction under covariate shift
SummaryThe paper studies the coverage of weighted split conformal prediction under covariate shift, conditional on the calibration data. It derives explicit training-conditional bounds without unspecified constants, showing the relevant scale is a variance proxy from the chi-squared divergence and an average likelihood ratio rather than its supremum. The authors also compare the method with randomized rejection sampling and importance-weighted learn-then-test, identifying regimes where each yields the narrower valid set.
In briefTrajectory Dropout: Strengthening Teacher Feedback in On-Policy Distillation
SummaryThis paper studies how the reasoning prefix a student model generates weakens token-level teacher feedback in on-policy distillation, a effect it calls Prefix-Induced Supervision Attenuation. The authors propose Trajectory Dropout, which randomly drops part of the student's reasoning during training while the teacher still sees the full trajectory. The method reportedly improves average performance across teacher-student pairs and six math reasoning benchmarks, with gains on two out-of-domain benchmarks.
In briefNative Association: Confidence-Aware Person Perception in Sports Broadcasts
SummaryThis paper fine-tunes a 0.77B vision-language model (Florence-2) to output each person's attributes, such as position, team, and jersey number, as one grammar-constrained sequence, avoiding a separate post-hoc binding step. The authors report 0.95 detection F1 on a frozen multi-sport test set, versus 0.65-0.75 for prompted frontier APIs. An extra forward pass provides per-field confidence that enables a reject option.
In briefA dataset and model tests for classifying illustrations in old manuscripts.
SummaryThe paper introduces a manually labeled dataset of 15,000 illustrations from centuries-old manuscripts, sorted into 22 categories. It compares fine-tuned CNN and Transformer classifiers, zero-shot CLIP, embedding-based classifiers, and vision-language models. Fine-tuned image classifiers performed best, with ConvNeXt reaching 88.9% accuracy and 81.3% macro-F1, while zero-shot CLIP and direct vision-language classification performed substantially worse.
In briefFrom PDF to Evidence: Structure-Aware Retrieval for Clinical Guidelines
SummaryClinical practice guidelines in PDF often hold key evidence in tables, flowcharts, and graded recommendations, which standard retrieval flattens into fixed-size text chunks. The authors parse each page image into typed structural elements and retrieve evidence units that follow the document's layout while keeping their structural context. On 26 guidelines with 199 queries, the approach ranks the correct element higher than per-element OCR text, matches page-level recall with less context, and outperforms a ColPali visual-RAG baseline.
In briefConcept Score Relearning: A Unified Attack on Concept Erasure Across Architectures
SummaryThis paper tests whether concepts erased from text-to-image models can be recovered. It proposes Concept Score Relearning (CSR), which optimizes model parameters in each model's native prediction space to reactivate erased concepts without an external target-concept image dataset. CSR works on both U-Net-based Stable Diffusion and Transformer-based FLUX. For strict nudity, it reaches average attack success rates of 50.47% on FLUX and 40.29% on Stable Diffusion.
In briefA unified view of regression-based diffusion reinforcement learning and its design choices
SummaryThis paper examines a family of reinforcement learning methods for diffusion and flow models that replace policy gradients with reweighted supervised regression, including DiffusionNFT, FlowAWR, and RAM. The authors show that each solves a divergence-constrained reward maximization problem, differing only in the convex generator defining the constraint. They analyze the approximations prior methods used for the advantage-embedded regression target and empirically study the design space. The resulting method, DiffusionRFT, is reported to converge faster, train more stably, and achieve top performance.
In briefContext Spanning: Linking Full-Duplex Speech Models to External LLM Backends
SummaryThis paper proposes Context Spanning, a framework that injects information retrieved by an external LLM into a full-duplex speech model via real-time chunked prefill. The retrieved text is fed to the speech model as-is rather than in a compressed latent form, which the authors argue avoids information loss. The authors report high performance on full-duplex benchmarks and strong results on question answering tasks.
In briefMIC: Explaining Mismatches Between Images and Claims in AI-Generated Misinformation
SummaryThis paper presents MIC, a fact-checking framework that detects AI-generated multimodal misinformation and explains where an image conflicts with its accompanying claim using world knowledge. It first applies supervised fine-tuning, then uses Group Relative Policy Optimization with verifiable rewards for verdicts, inconsistency types, visual evidence, and explanations. The authors also introduce MIC-Bench, a benchmark of 8,812 image-claim instances.
Ruihong Zeng, Jonathan Tonglet, Preslav Nakov et al.
In briefMAC-Net: A Deep Learning Model Predicting Thinking Abilities from Brain Scans
SummaryThis paper presents MAC-Net, a deep learning framework that predicts individual cognitive function from task-based fMRI brain scans while accounting for demographic and scanner-related factors. The model processes neural features and participant variables through separate pathways and was tested on baseline data from 6,500 participants in the Adolescent Brain Cognitive Development Study. The N-back plus Monetary Incentive Delay combination performed best, with N-back the most informative task, while adding the Stop Signal Task slightly reduced accuracy.
Md. Tanvir Rahman, Nabil Anan Orka, Asaduzzaman Khan et al.
In briefRaven: a meta-harness that builds and combines specialized AI agent setups
SummaryThis paper introduces Raven, an open-source multi-agent system that automatically builds and evolves modular harnesses for specific models and domains, treating each model-harness pair as a composable unit. A Host Agent breaks down goals, assigns subtasks to specialized agents, and integrates their results, while stored experience and reusable procedures carry knowledge across tasks. The authors report that Raven significantly outperforms state-of-the-art agent systems on complex, long-horizon tasks.
In briefSMAT: A Simple, Efficient Training Method for Better Model Merging
SummaryModel merging combines the abilities of several expert models without retraining them jointly, but standard expert training does not ensure good merged performance. The paper observes that common merging methods can be described by three operations (Scale, Mask, and Perturb) and introduces SMAT, which optimizes expert loss and expected loss on simulated merged parameters. Efficiency is improved through periodic scheduling, kernel fusion, and parameter storage switching. Across four language and vision-language backbones, SMAT raises the mean score over five merging methods by 1.07 to 2.16 points versus the strongest baseline, with under 2% training-time overhead over standard fine-tuning.
In briefSchemaMem: Schema-Indexed Recurrent Memory for Delayed State Retrieval
SummaryThis paper introduces SchemaMem, a recurrent memory architecture that combines chunk-local attention with a persistent, schema-indexed state for retaining and later retrieving information. In a controlled address-value task with roughly matched parameter counts, it shows higher mean written-value retention than two baselines at four times the maximum training delay. It also recovers updated values better in most settings, though Gated DeltaNet performs better at that delay by default and SchemaMem needs substantially more optimization steps.
In briefCORA: A Protocol for Testing Text-to-Audio Search Robustness to Rephrased Queries
SummaryThis paper introduces CORA, a protocol that rewrites each audio caption into five intent-preserving forms (command, question, indirect, key phrase, statement) while keeping the target audio fixed. The authors define RankDrop, a metric that reveals failures hidden by Recall@k, and find it correlates more strongly with target boundary degradation than with text-space query movement. They conclude that robust text-to-audio retrieval requires the target to keep its boundary advantage over competing audio under reformulation.|
