In briefBuilding coding-agent benchmarks from real vibe coding session records
SummaryThe paper presents SessionFarmer, a framework that converts developer-agent dialogue records and commit histories into self-contained, executable software engineering tasks. It filters unrelated sessions and groups multi-turn, multi-session interactions into work episodes centered on one engineering goal, then generates task statements, quality rubrics, and unit tests. The authors construct CropChat, an open benchmark of 242 tasks, and CropEnterprise from proprietary enterprise logs.
In briefEfficient LLM planning that checks predicted outcomes against real observations
SummaryThis paper proposes an expectation-observation consistency verification (EOCV) mechanism for batched LLM planning executors. The method predicts expected action outcomes and compares them with actual observations, continuing execution when they match and replanning from the current observation when they do not. Experiments on text-based and multimodal benchmarks report task performance broadly comparable to reactive executors, with 29-35% lower average weighted token usage. The mechanism is model-agnostic and remains effective when integrated into state-of-the-art planners.
In briefJudging Synthetic Skin Images for Training Classifiers, with Safety in Mind
SummarySkin cancer classifiers need many labeled images, which are scarce for rare lesion types. The paper generates about 275,000 images with a diffusion model (EDM2) and tests 16 quality metrics by training a classifier on the subset each metric selects. No single metric works best for every class, so the authors propose a safety-aware algorithm that weights each metric per class and removes images too close to real training images. On the ISIC-2018 test set, it gives a small but statistically significant gain over the FID baseline for a Swin-B classifier, but not for ResNet-50.
In briefM3Former: Finer-Grained Information Mixing Across Transformer Layers
SummaryThis paper argues that standard Transformer residual connections pass information from earlier layers through a single fixed sum shared by all of a block's inputs, which limits what the model can use. The authors propose Multi-Role Multi-Group Mixing Residual (M3), where each role and each group of hidden dimensions forms its own token-dependent mixture of all earlier layers. In experiments from 100M to 1.4B parameters, M3Former outperforms the Transformer and coarser cross-layer connections, and it also improves sparse mixture-of-experts models up to 4.6B parameters.
WhyThe authors decompose residual broadcast into three constraints and relax them, proposing token-dependent cross-layer mixing per role and group; the abstract reports gains from 100M to 1.4B parameters
In briefBlueprint2CAD: A benchmark testing whether AI models match engineering drawings to 3D part renders
SummaryThis paper introduces Blueprint2CAD, a benchmark that checks whether an orthographic engineering drawing and a shaded render show the same mechanical part. The dataset BlueCAD-30K is built from 29,925 CAD models. With known viewpoints, frozen encoders reach up to 65% Recall@1, falling to 43% when viewpoints are unknown, and viewpoint is identified as the main bottleneck. Six zero-shot vision-language models achieve only 0.52 to 0.67 verification accuracy.
Alexander Valverde, Danny Xie-Li, Zhichao Wang et al.
In briefUsing LSIRM to Diagnose Model-Item Interactions in Language Model Benchmarks
SummaryThis paper applies the latent space item response model (LSIRM) to 1,348 language models answering 5,761 BBH items, and compares answer presentation and scoring methods across BBH, GPQA, MMLU-Pro, and MuSR. It finds that nearby items on the LSIRM maps often share the same correct answer, a pattern stronger when answers are scored by option labels than by option text. The authors estimate each model's tendency to do better or worse on items with particular correct answers and separate this from its score. Two models differing by only 14 correct items move hundreds of ranks when this contribution is excluded.
In briefA parameter-free policy for choosing channels in queueing bandit problems
SummaryThis paper studies how to schedule a single queue across K channels when only bandit feedback is available. The proposed policy explores when the queue is empty and, during busy periods, picks the empirically best channel while conditions hold, switching to a robust bandit policy if they fail. Under arbitrary arrivals and services set by an oblivious adversary, it achieves Õ(√(K L_T)) worst-case queue length regret with high probability, where L_T is the longest busy period, improving the earlier T^{3/4} dependence. Under bounded i.i.d. models with a unique best stabilizing channel, cumulative queue length regret converges to a finite problem-dependent constant.
WhyThe authors propose a parameter-free best-of-both-worlds policy improving queue regret from T^{3/4} to near-tight √(K L_T), with high-probability worst-case and stochastic results reported.
In briefInteractionBench: A Benchmark for Timing, Spoken-Task Accuracy, Whispering, and Grounding in Real-Time Voice Systems
SummaryThis paper presents InteractionBench, a benchmark that evaluates real-time voice systems on the same recorded user audio across four dimensions: conversational timing, spoken-task accuracy, whisper and volume, and groundedness in conversations of more than twenty turns. It contains 221 items per system across thirteen scenarios and reports 1,519 graded runs over seven systems. Among the six full-duplex systems, timing and spoken-task accuracy are rank-correlated (ρ = 0.83).
In briefLocal complexity of overparameterized models arising from singularities
SummaryThis paper studies overparameterized singular models, whose optimal parameters form multiple points or whose generalization-error Hessian is degenerate. It focuses on the real log canonical threshold (RLCT) and its localized version, the local RLCT, which characterize generalization in singular models. Under a localization assumption, the authors clarify the theoretical relationship between the local RLCT and two metrics: the rank of the generalization-error Hessian and the covering number of a loss function family.
In briefMixed-precision training for geometry-aware neural operators with uncertainty calibration
SummaryThis paper proposes Architecture-Aware Mixed Precision (AA-MP), which assigns numerical precision based on replica roles in replica-exchange training, using BF16 for exploration and FP32 for refinement. Across fluid dynamics and three-dimensional structural mechanics benchmarks, it reports accuracy comparable to full-precision training, with reduced peak GPU memory and training time. The authors also show near-nominal coverage with narrower intervals under conformal prediction.
In briefTERA: Restoring Task-Relevant Image Evidence for Underwater Detection Over Limited-Bandwidth Links
SummaryThe paper notes that underwater object detectors are usually tested only on images before transmission, ignoring extra degradation from bandwidth-limited downscaling and compression. The authors build AquaTX, a benchmark from real underwater detection datasets with both seen and held-out codecs. They also propose TERA, a receiver-side method that uses no metadata and combines task-guided pixel restoration with dual-state feature correction while keeping the detector frozen.
WhyThe authors add bandwidth-limited compression degradation to underwater detection evaluation and propose frozen-detector receiver-side restoration; they report mAP gains on DUO-TX and RUOD-TX plus hel
In briefReward-Robust RL With Behavior-Distinct Reward Shaping Sets
SummaryThis paper studies reward-robust reinforcement learning, which trains agents to handle possibly misspecified reward functions. The authors observe that under potential-based reward shaping, many different rewards produce the same policy ordering, creating redundancy in standard uncertainty sets. They propose projecting each reward onto a canonical representative so the set contains only behaviorally distinct rewards. The paper reports that this preserves the optimal robust value, shrinks the set, and improves performance over representative robust RL baselines across benchmark tasks.
In briefSGR-OPSD: Subgoal-Guided Credit Assignment for Training AI Agents
SummaryThis paper proposes SGR-OPSD, a reinforcement learning framework for AI agents that uses LLM-generated subgoal plans to give dense action-level guidance. It builds a Subgoal Progress Graph from sampled transition frequencies to estimate empirical returns and calibrate credit weights. Tested on ALFWorld, WebShop, and Search-QA with Qwen2.5 3B and 7B models, it outperforms GRPO and self-distillation baselines, reaching 85.8% on ALFWorld with the 7B model.
In briefA framework for checking whether brain MRI measurement pipelines support stable scientific conclusions
SummaryThis paper introduces Scientific Claim Stability (SCS), a framework that evaluates the conclusions supported by MRI morphometry pipelines rather than only measurement agreement. It converts region-of-interest morphometry tables into auditable claim cards and tests their support under changes in dataset, pipeline, atlas, feature family, and target. Using 10,370 MRI scans from three datasets and five cohorts, the framework generates 2,783 claim cards and evaluates 89 claim-family pairs.
Duy-Cat Can, Dang-Duat Tran, Anh-Khang Nguyen et al.
In briefA new federated learning framework that defends against gradient inversion attacks
SummaryThis paper proposes HyperFL++, a privacy-preserving federated learning framework designed to defend against gradient inversion attacks. It uses hypernetworks to generate local model parameters, so the server aggregates only the hypernetwork parameters, which breaks the direct link between shared parameters and local private data. A Row-wise Kronecker Decomposition is added to reduce the memory and communication overhead of the hypernetwork. The authors report that theoretical analysis and experiments show it protects data privacy while maintaining high model utility.
In briefKey estimation under deployment limits using a DSP-guided transformer reranker
SummaryThis paper presents key lime, a music key estimation method for local applications that must stay small and fast during library import. A 21.5k-parameter transformer reranks deterministic DSP scores, with a 115 KB model artifact and a 76.7 MB installed footprint. On FMAK and FMAKv2 tracks, MIREX rises from 61.08 to 73.55 and exact accuracy from 52.30 to 67.07, approaching a madmom KeyCNN reference at 74.16 MIREX.
In briefMirrorComm: A Communication-Efficient Cooperative Perception Framework Using Mirrored World Models
SummaryThis paper presents MirrorComm, described as the first world model-based communication-efficient cooperative perception framework. Mirrored world models at the transmitter and receiver predict the transmitter's current observation from shared reconstructed history, and only observation updates that correct prediction errors are sent. On the OPV2V dataset for object detection and semantic segmentation, it reduces bandwidth requirements by up to about 66 times while keeping performance comparable to a state-of-the-art method. Ablation studies are reported to show the effect of each component.
In briefWhen Rewards Depend on the Group, GRPO Plays a Tournament
SummaryThe paper shows that when rewards for diversity, coverage, or distribution matching depend on the other responses in a sampled group, the group-relative baseline used by GRPO and RLOO changes the training objective. Each response is effectively trained to beat its groupmates in a zero-sum tournament. The authors propose Team-GRPO, which sets the baseline coefficient to +1 to give an unbiased estimator for any group-dependent reward using two samples per prompt. On GlobalOpinionQA, Team-GRPO reduces error from 0.20 to 0.08 where GRPO collapses to the majority answer, and on NoveltyBench it avoids GRPO's diversity collapse.
WhyThe authors prove group-relative baselines alter group-dependent reward objectives and propose unbiased Team-GRPO, validated on GlobalOpinionQA and NoveltyBench.
In briefSolver-like reasoning in LLMs used as formalizers
SummaryThis paper studies how large language models that translate problems into formal programs for solvers behave during their reasoning traces. The authors call the solver-like work done in these traces "solver-like behavior" and label 2,000 generations from three reasoning models on four constraint satisfaction datasets using LLM judges checked against human labels. The behavior appears in 85% of traces and uses 24% of reasoning tokens, yet the programs usually recompute the answer, making the effort redundant.
In briefIMUdrape: Generating Virtual Sensor Data for Clothing-Mounted IMUs from Text
SummaryIMUdrape is a method for generating virtual inertial sensor (IMU) data for sensors sewn into clothing. It synthesizes motion from text, reconstructs the garment as a sewing pattern from photos, places virtual sensors at real positions, and runs physics-based cloth simulation to produce paired garment and body signals. On four real datasets it outperforms existing virtual-data approaches for activity recognition and full-body pose estimation, reducing the amount of real data needed.
In brief"Late Autumn": LLM Agents Tailor a General Robot Planner to One Warehouse
SummaryThe paper explores whether LLM coding agents running autonomous research can specialize a general multi-robot planner for a single warehouse, trading generality for throughput. The planner combines task assignment with multi-agent path finding and was tested in six campaigns on a floor modeled on an existing automated warehouse. The agents raised throughput by 22 to 60 percent, with gains holding on unseen seeds.
In briefWhich Chains to Calibrate for Physical Backpropagation: Per-User Factors vs. Base-Station Chains
SummaryThis paper studies reciprocity calibration for physical backpropagation over the time-division duplex uplink, where users return loss derivatives to the base station. It compares calibrating base-station chains against calibrating per-user amplitude and phase errors, finding that the conventional priority can reverse. The authors report that the ordering holds at 30 dB SNR and reverses for 1 dB SNR with 10 degree user errors, and they evaluate a capped one-echo estimator.
In briefEviCast: Predicting Retrieval Benefits via Shared Response Calibration
SummaryEviCast is a framework that forecasts whether retrieval-augmented generation will help a given generator, task, and retriever before running full large-scale experiments. It measures the generator's response to controlled evidence states once, then combines this with each retriever's evidence-state distribution to estimate end-to-end net utility and its 95% lower confidence bound. In evaluations on HotpotQA, MuSiQue, and a third dataset, it correctly decided most retrieval configurations while cutting diagnostic generation calls by 80%.
WhyThe authors propose forecasting RAG retrieval gains before execution via shared generator calibration, with a topology-aware variant; the abstract reports decisions across 14, 16, and 24 configuration
In briefNearest Is Not Safest: Choosing Antenna Ports for Over-the-Air Gradient Return in City Multipath
SummaryThe paper studies how a fluid antenna should choose its return port for over-the-air backpropagation when a terminal moves in urban multipath. In ray-traced Chicago street simulations, returning through the nearest port reverses many gradients, with the rate rising as port spacing grows. The authors also derive a condition certifying the sign of the mean correlation and propose a knowledge map from surveyed paths that lowers the reversal rate, though its advantage over a geometry rule does not hold under all tested errors.
In briefBeyond Accuracy: Effects of Quantization on Coding Agents
SummaryThis paper compares BF16 with 4-bit AWQ and GPTQ quantization for several coding agents, including Qwen2.5-Coder-Instruct models (7B, 14B, 32B) and Devstral-Small-2507 (24B), on OJBench coding problems. No statistically significant drop in final accuracy was found for any agent. However, the study reports substantial behavioral changes, such as shifts in how problems are solved and how tools are used, along with changes in failure types. The authors conclude that quantized coding agents should be evaluated through solution trajectories, tool use, and failure modes, not just final accuracy.
Areej Mehboob, Muhammad Ali Shafique, Muhammad Usman Qadeer et al.
In briefDo Visual RL Gains Hold Up When Sampling More at Test Time?
SummaryThis paper examines whether reinforcement learning post-training gains for visual generative models persist when Best-of-N sampling adds inference compute. The authors find that higher single-sample reward does not always yield a better quality-compute tradeoff, and gains can shrink, plateau, or reverse at large sample budgets. They explain this with a quantile-transfer framework and propose Quantile Utility Policy Optimization (QUPO), which better preserves upper-tail quality on SD3.5-M and BAGEL.
In briefHyperspectral image classification using a spectral-spatial vision test-time training model
SummaryThis paper presents ViTTT-HSI, an adaptation of the vision test-time training architecture (ViT3) for classifying hyperspectral images. The model uses spectral convolution to extract ordered-band information and builds spatial tokens without reducing spatial resolution. Two ViT3-based test-time training encoder blocks then mix the tokens, and center-pixel and neighborhood-average features are fused for classification. Evaluated on four datasets with ten random seeds and 1% or 5% training fractions, it reports mean overall accuracies of about 97.64%, 98.88%, 99.33%, and 98.43%.
In briefGEAR: Scene-Based Outfit Search for VR Travel Using Generated Images
SummaryThis paper presents GEAR, a framework for finding outfits for avatars in virtual 3D environments based on the user's first-person view of the scene. A vision-language model describes suitable clothing for the scene, a diffusion model renders that clothing onto a standard reference character, and the cropped garment is used to search existing 2D outfit assets by embedding similarity. An interactive Unity prototype is included, and GEAR is compared with baseline methods on a controlled dataset and in a user study across 20 virtual travel scenes, where it received the highest overall participant preference.
Gia-Huy Dinh, Thinh-Phuc Nguyen, Thanh-Hai Nguyen et al.
In briefStudy Showing Prior Control Over Rewards Causally Changes Language-Model Plasticity
SummaryTwo copies of one language model, starting from identical weights, were trained on the same rewards, but one copy's rewards depended on its own answers while the other's were yoked to it. The copy with control learned a later new task faster, and this advantage persisted in a second experiment where both copies trained on identical, fixed answers. The authors also report that the advantage could be predicted from each copy's weights before the new task began.
WhyThe authors adapt a learned-helplessness yoked design to test LLM plasticity, reporting paired experiments on Qwen3-0.6B and 1.7B with fixed-answer controls.
In briefAgent memory as rate-limited filtering: explicit state beats long context past a reader-dependent length
The summary in this language did not pass the date and amount checks, so it is withheld. Check the original.
WhyThe authors model agent memory as rate-limited recursive filtering, deriving a capacity-dependent forgetting horizon and a tight error bound, and report a reader-dependent crossover with long context.
In briefUsing text descriptions to match the same object across different camera views more reliably
SummaryThis paper presents TeSCo, a framework for cross-view object correspondence, which finds the same target in query and search videos captured from different viewpoints. It uses a vision-language model to generate a textual description of the target from the query view, then fuses this text with visual features through text-conditioned and iterative refinement modules. The authors report state-of-the-art results on the Ego-Exo4D and HANDAL-X datasets.
In briefVisCodeBench: A Benchmark for Realistic Text-to-Visualization Generation
SummaryThis paper introduces VisCodeBench, an end-to-end benchmark of 2,672 samples for generating visualizations from natural language requests, spanning educational, scientific, and general audience scenarios. It evaluates three systems (Codex, Claude Code, and LIDA) with various LLMs across seven metrics covering execution, chart content, and visual quality. Results show basic execution is nearly solved, but no system closely reproduces reference data, with the best data similarity at 0.81.
In briefWhich Words a Model Fears: From a Refusal Direction to a Vocabulary-Wide Fear Lexicon
SummaryThis paper studies whether aligned language models refuse requests partly based on which words appear in a prompt, independent of their meaning. The authors extract a refusal direction shared across prompts and words, and show that moving hidden states along it raises or lowers refusal monotonically with dose across eight models. Using this direction to score every vocabulary word, they validate the resulting ranking with two held-out tests.
WhyThe authors extend a refusal direction into a vocabulary-wide fear lexicon, showing single words can trigger refusal; validated across eight models with dose-response and two held-out tests.
In briefAgiWorld: An Agent-Driven Video World Model for Persistent Worlds
SummaryAgiWorld is a video world model that explores and builds a persistent virtual world from a single image and a free-form instruction. An agent, AgiWorld-Agent, plans at a high level and maintains a world state, while AgiWorld-WM generates video chunk by chunk and stores it with geometry in a spatial memory, forming a closed loop between the two. The authors report state-of-the-art overall performance on world revisiting and exploration benchmarks, and that the world stays consistent over minute-long journeys.
In briefLearning to stabilize unknown continuous-time stochastic systems from data
SummaryThis paper numerically studies identifying unstable continuous-time stochastic linear systems from observed trajectories in order to learn a stabilizing gain. It examines how failure probability depends on data amount, dimension, noise intensity, excitation, and controllability. On an unstable batch-reactor benchmark, identification error decays inversely with total observation time while failure probability decays exponentially. The authors also report that stronger excitation improves stabilization only logarithmically, and that accurate identification of the full dynamics is not necessary for successful stabilization.
In briefMemShift: Separately Measuring Detection and Repair in AI Memory Derived from Personal Records
SummaryThe paper introduces MemShift, a benchmark for AI assistants that store values derived from a person's records when those records change. It covers five kinds of shifts and scores detection and repair separately with a deterministic scorer instead of an LLM judge. Evaluating four open-weight LLMs, the authors find that detection only helps as far as repair allows, and that programs written over typed records close most of the gap to exact recomputation.
In briefA structure-preserving Eig-PIELM method for Maxwell eigenvalues that removes gradient modes
SummaryThis paper addresses the problem that, for Maxwell curl-curl eigenproblems, the perfect electric conductor boundary condition alone leaves a gradient nullspace that can produce spurious modes in Eig-PIELM. The authors propose dR-Eig-PIELM, which uses compatible grad-curl trial spaces and a constrained null-space construction to remove the represented gradient image before solving the reduced eigenproblem. For rectangles and cuboids they derive exactness and dimension results, and numerical tests show that spurious gradient-contaminated modes are excluded. A waveguide example also illustrates use in geometry design optimization.
In briefAttribution-based diagnostics for checking how single-cell foundation models make decisions
SummaryThis paper applies post-hoc attribution to downstream predictions of single-cell RNA foundation models (scFMs) to test whether their decisions rely on biologically meaningful signals or model-specific shortcuts. It covers two tasks: cell-type annotation with Cell2Sentence and scGPT, and perturbation prediction with scGPT+STATE, where attributions are compared against known regulatory ground truth in a synthetic setting. The authors report that attribution reveals information beyond standard performance metrics, and that conclusions depend strongly on the choice of biological reference, baseline, and null model.
Farnoush Rezaei Jafari (1, 2), Tristan Lazard (3) et al.
In briefDoes a Norwegian further-trained model give more outdated legal values than its base model? A preregistered comparison of Borealis 2 and Gemma 4
SummaryThis study compares how often two released language models give outdated numeric values from Norwegian statutes and regulations: Borealis 2, further trained by the National Library of Norway from Gemma 4, and Gemma 4 itself. The study was preregistered and scored deterministically against a value history using the closed norregel benchmark. The difference between the models is small and not significant, falling within the preregistered range of no effect of interest. The design compares released models and does not isolate the effect of any part of their training.
In briefWhen Adaptive Halting Helps: Why Selection Rules Matter for Looped Language Models
SummaryThis paper studies when looped language models can stop iterating early. On a frozen four-loop Ouro-1.4B model, the authors compared a trained stability probe with the model's native exit gate. In a pre-registered confirmatory test with a 1-point accuracy-loss rule, the probe saved more compute, but under a point-only selection rule the preference reversed. The authors conclude that halting methods should be compared together with their selection procedures, using strong untrained baselines and out-of-distribution checks.
In briefLeanGuard: Lean-Verified Guardrails for Tool-Using AI Agents
SummaryThe paper presents LeanGuard, a runtime that checks every tool call proposed by a language-model agent against a policy written in Lean4 before the call is executed. Forbidden calls are denied and can be corrected either by feedback that tells the agent which rule was violated, or by proof-guided repair, where Lean searches for a nearby admissible call that keeps the user's intent. Results are reported on Tau-Bench and DTap, including reduced prompt-injection success.
Sreehari Sankar, Debargha Ganguly, Vikash Singh et al.
In briefSpatial Transfusion: From Scene Understanding to Spatial Audio Generation and Editing
SummaryThis paper presents Spatial Transfusion, a unified framework for first-order ambisonics (FOA) spatial audio that extends a pretrained spatial audio understanding model to generation and editing. It uses flow matching over VAE latents and frames editing as structured scene transformations over a taxonomy of atomic edits. Experiments reportedly keep spatial understanding largely intact while outperforming full fine-tuning on generation and editing, supporting controllable FOA creation through both structured and natural-language interfaces.
In briefFindings-aware Language-Image Pretraining for CT Volume Interpretation
SummaryThis paper proposes FiLIP, a method for learning representations that link CT volumes with radiology reports. A large language model extracts abnormal finding phrases from reports, and a biomedical text embedding model converts them into target features for the image encoder. FiLIP outperformed weakly supervised approaches on multi-label zero-shot finding classification in both in-domain and external-domain evaluations, and was also effective when fine-tuned for radiology report generation.
In briefDynaCell Evaluation Suite for Dynamic 3D Virtual Staining of Live Cells
SummaryThis paper presents an evaluation suite for judging virtual staining models on 3D time-lapse images of live cells, which predict fluorescence images from unlabeled images. The suite includes paired labeled and unlabeled 3D time-lapse data, baseline models, and three metric tiers covering pixel fidelity, organelle segmentation, and single-cell phenotype similarity. Across 3 cell types, 2 microscopes, 4 organelles, and 3 perturbation states, regression baselines better predicted organelle localization, while the generative baseline better captured phenotype distributions across cells. The datasets, baseline checkpoints, and evaluation code are released as part of the DynaCell framework.
Alexandr A. Kalinin, Dihan Zheng, Taylla Milena Theodoro et al.
In briefLanguage Model Capitulation Under Pushback: A Direction That Emerges With Scale
SummaryThis paper tests whether language models' tendency to abandon correct answers when users object corresponds to a linear direction in the residual stream, and whether that direction appears with scale. Across eight instruction-tuned models from 0.5B to 14B parameters in two families, under a pre-registered protocol, a direction clearing the validation gate appears at 7B, 8B and 14B. Ablating it in Qwen2.5-7B reduces answer reversals by 4.8 percentage points.
WhyThe authors localise capitulation under pushback to a residual-stream direction and report a causal ablation reducing reversals by 4.8 points across eight pre-registered models.
In briefGRIPS: An Empirical Study of Pre-SQL Grounding Information for SQL Generation
SummaryThis paper studies how information gathered before SQL generation helps text-to-SQL systems link question terms to the correct tables, columns, values, and relations. The authors propose GRIPS, a three-stage method that derives structured pre-SQL information to make these decisions explicit. GRIPS-Adaptive outperforms the SQL Static baseline in 18 of 20 settings on TempTabQA-C and BIRD.
Ekata Mitra, Yufei Tao, Kamalamalini Nagasundaram et al.
In briefDeep learning to detect fraudulent financial transactions and strengthen cybersecurity
SummaryThis paper proposes a deep learning method that uses neural networks to identify anomalous financial behavior for fraud transaction detection. The authors evaluate several models, including LSTM, CNN, and hybrid architectures, on real-world financial transaction datasets. The results show clear improvements over conventional machine learning baselines. The approach detects fraud robustly while keeping false positives low, and is presented as scalable for enterprise financial cybersecurity systems.
In briefCytoWave: A Pretraining Framework for Predicting Single-Cell Responses to Gene Perturbations
SummaryThe paper presents CytoWave, a pretraining framework that predicts how single cells change their gene expression after genetic perturbations. It uses large cross-species perturbation datasets unified through orthology, training first on unperturbed cells and then on predicting perturbed expression with a contrastive objective. Strict dataset-held-out evaluation and pathway activity case studies show consistent and effective performance.
In briefLie Routing for Viewpoint-Robust 3D Vision-Language Models
SummaryThis paper introduces LieVLM, a Lie routing framework that helps multimodal models give consistent answers about 3D scenes from different camera viewpoints. It uses geometry tokens from VGGT to predict local rotation parameters and routing coefficients, which adjust projected visual features through a closed-form Padé–Cayley map. Experiments on 3D captioning, grounding, detection, and spatial reasoning report consistent gains over additive fusion.
In briefA supervised medical image segmentation loss that focuses on fuzzy object edges, aiming to stay reliable with noisy labels.
SummaryThe paper introduces the Peripheral Prototypical Contrastive (PPC) loss, which uses prototype features from local regions near object boundaries to improve boundary recognition in medical image segmentation. Experiments across multiple datasets reportedly show consistent performance gains, including under noisy labels. An adaptive strategy for choosing the local region used to compute prototypes is reported to improve robustness and outperform existing methods.
In briefA framework for making industrial AI agents execute changes to engineering systems reliably.
SummaryThe paper presents a governance framework for trustworthy execution of industrial agents within one enterprise domain, defining eight control domains and a four-plane architecture with three control chains. It separates tool return values from actual engineering effects and evidence coverage from engineering judgment, and examines selected mechanisms through prototypes, conformance tests, and CAD/CAE and OpenFOAM cases. Reported results include 0/29 hazardous commits, which measure closed-set specification conformance rather than safety probability.
In briefFactorSpec: Serving Exact Relational Constraints as Speculative Drafts
SummaryThis paper presents FactorSpec, which compiles a JSON Schema and its field relations into a factor graph with an exact feasibility check. Tokens fully determined by the constraints are used as certain speculative drafts. In SGLang, it yields up to 10.5 tokens per verify step and decodes up to 1.43x faster than tuned speculators, and it serves up to 13.9x more valid records per GPU-second than syntax-only baselines.
In briefPreparing Built Environment Organizations' Workflows for Agentic AI
SummaryThis paper reviews how data in the built environment can be organized to support agentic AI systems that reason, plan, and act autonomously. It focuses on knowledge structures, interoperability, and access mechanisms such as retrieval-augmented generation (RAG) and the Model Context Protocol (MCP). Drawing on existing literature, it examines methods for organizing structured and unstructured data and how AI agents can reliably access it. The paper also highlights challenges of data fragmentation and interoperability.
Ahmet Bahaddin Ersoz, Justin McHenry, Yevgeniya Kovalchuk et al.
In briefSVD-Prune: Spectral Token Pruning for Efficient Vision Models
SummaryThis paper proposes SVD-Prune, a training-free method that prunes visual tokens in vision-language and vision-language-action models. It uses singular value decomposition and statistical leverage scores to keep the dominant low-rank spectral structure of visual representations. The method can be applied at the encoder, decoder, or both, and reportedly remains robust under extreme compression, cutting the visual token budget by up to 99.3%.
Yvon Apedo, Martyna Poreba, Michal Szczepanski et al.
In briefConcise-9B: Trimming Unneeded Thinking in a Reasoning Model
SummaryThis paper studies how reasoning models keep thinking long after their answer has already formed, using Qwen3.5-9B as the base. The authors apply two rounds of on-policy rejection-sampling fine-tuning that keep the shortest correct samples, plus a small share of answer-synthesis targets, producing Concise-9B. On 8,983 held-out problems at a 32K thinking budget, it generates 0.66x the tokens of the base model with accuracy within ±1 point.
In briefPer-atom message-passing depth allocation for predicting molecular properties
SummaryThis paper introduces PRISM, a method that lets graph neural networks choose a message-passing depth separately for each atom. Each atom produces a bounded scalar that places a unimodal distribution over intermediate message-passing states. On molecular property benchmarks with scaffold splits, PRISM achieves strong results against eight baselines and transfers to GCN, GIN, and GAT backbones.
Regina Qianru Zhang, Honggang Wen, Zongru Li et al.
In briefBenchmark design for biosynthetic gene clusters across data scarcity, granularity, and modality
SummaryBiosynthetic gene clusters (BGCs) encode specialized metabolites such as antibiotics and anticancer agents. The paper proposes a benchmark-design framework and instantiates it as BGC-Bench, which derives five tasks from the MIBiG database, evaluates them under genotype- and phenotype-shift splits, and includes an active-learning setting. Twelve encoders are compared across nucleotide, amino-acid, and Pfam-domain representations. No single representation dominates, and pretrained encoders do not consistently outperform other representations across tasks and splits.
Hanlin Xiao, Eriko Takano, Mauricio A. \'Alvarez et al.
In briefWhether invertible graph aggregation can make a ReLU readout classifier need more hidden units
SummaryThe paper studies whether an invertible neighborhood aggregation on graphs can increase the number of hidden units a per-node classifier needs. For every even N≥4, it builds a labeled tree with 2N vertices and maximum degree three, showing a shared one-hidden-layer ReLU classifier needs one unit on raw features but N units after one lazy neighborhood mean. It also gives error bounds under bounded uniform noise and compares trained heads on Roman-empire and Tolokers.
In briefWhen the Text Is Silent: Memory Shortcuts in LLM Annotation
SummaryThis paper examines whether large language models, when annotating text, rely on what they already know about well-known people instead of the text itself. Using counterfactual texts built from 995 public figures and matched control names, the authors test three commercial LLMs and find that models never favor memory over conflicting text, but when evidence is absent, famous names raise the rates of inferred occupations, ages, and countries. The resulting labels are often factually correct about the namesake yet unsupported by the text, so they are methodologically invalid.
WhyThe authors separate textual evidence from memorized knowledge, reporting memory shortcuts across 995 public figures, three commercial LLMs, and 121,594 prompts, plus mitigation trade-offs.
In briefCurrent state of LLMs for Arabic dialect machine translation
SummaryThis paper evaluates large language models translating between English and 16 Arabic dialects on the MADAR dataset, in both directions. It compares specialized Arabic models, multilingual models, and commercial APIs using metrics including BLEU, CHRF, COMET, and GEMBA, plus a small manual evaluation. The study also tests whether models can identify a text's dialect. It finds dialectal diversity particularly difficult to handle, while noting results may be affected by possible training data contamination.
In briefHinted Scaling Laws: Predicting Downstream Accuracy for Small Models Using Extra Hints
SummaryScaling laws forecast large model performance from smaller training runs, but they fail when small models score near zero on downstream benchmarks. This paper introduces hinted scaling laws, which use privileged information to create a range of easier benchmark versions so small models score above zero. Accuracy rises smoothly and predictably with hint level, and on the Delphi pretraining ladder the method forecasts GSM8K accuracy more accurately than a standard scaling law. Hinting also reduces the number of test examples needed to compare models.|
WhyThe authors introduce hinted scaling laws that keep small models above zero accuracy, reporting much lower GSM8K, AIME and HLE forecasting error across model families.
Suze van Adrichem, Luke Bailey, Rohith Kuditipudi et al.
In briefLearning newsvendor orders when stockouts hide demand, using randomized ordering
SummaryThis paper studies online newsvendor learning with bounded demand, censored observations and convex nonlinear shortage costs. Because stockouts hide shortage magnitudes, the authors use randomized ordering to build unbiased subgradient estimates and combine them with gradient-based updates in an online inventory policy. Its expected regret against the best fixed order in hindsight is O(T^{2/3}), improving to O(√T) under independent demand and suitable curvature, with matching minimax lower bounds.
In briefA Theory for Designing Rules in Multi-Agent Systems, Illustrated with Homogeneous Fleets
SummaryThis paper develops a general theory for designing rules that regulate systems of many decentralised agents, linking each rule to the set of system performances the regulator wants agents to reach. It introduces the homogeneous rule design problem, where agents are identical, and shows that under worst-case aggregation its solution matches that of the general problem, avoiding combinatorial complexity. The authors propose RuleBnB, an anytime branch-and-bound algorithm for solving the general problem, and demonstrate its value in an autonomous driving case study.
Matteo Penlington, Lorenzo Fiaschi, Raj Kiriti Velicheti et al.
In briefA three-level model combining technology and data to manage workplace safety risks.
SummaryThis paper presents the 3-Tier Integrated Technology Risk Model, a staged framework for occupational safety and health management. It combines Job Safety Analysis, dynamic risk assessment using leading and lagging indicators, and predictive analytics powered by large language models. Each tier builds on the one before it, moving organizations from reactive safety practices toward predictive safety management with just-in-time training and preemptive risk mitigation.
In briefMulticlass classification when only rare features are useful, and a screening method that works.
SummaryThe paper studies high-dimensional multiclass classification where only a tiny fraction of features are useful and the number of classes may be large. It proposes Diversity Pursuit Higher Criticism Thresholding (DP-HCT), which applies a higher-criticism threshold to p-values from per-feature diversity tests such as one-way ANOVA. The authors show that whenever classification beyond chance is possible, DP-HCT with ANOVA and a simple class-mean similarity classifier achieves asymptotically perfect accuracy.
WhyThe authors extend rare-and-weak feature theory to multiclass, deriving a stricter phase transition with growing classes and DP-HCT optimality, validated on real benchmarks and simulations.
In briefPartial Invariance for Classification Under Causal Models
SummaryThis paper introduces environment-label separation, a form of partial invariance that splits the target conditional distribution into an environment-invariant part and a label-invariant part. Within the structural causal model framework, the authors give sufficient conditions for this separation and build an algorithm, dIMP-Classifier, that jointly minimizes an estimate of test risk and encourages invariance across environments. Experiments on simulated data and the real-world light tunnel Causal Chamber data show the method performs under interventions on the target mechanism and outperforms the compared standard and invariance-based baselines.
In briefCan LLMs Reliably Judge Their Own Answers? A Probabilistic VC Framework
SummaryThis paper proposes a theoretical and empirical framework for testing whether large language models can tell their correct solutions from incorrect ones while reporting well-calibrated confidence. It adapts classical VC theory into two metrics, Probabilistic VC and Calibration-aware PVC. Eleven 7-8B models were tested on mathematical reasoning, TruthfulQA, and CommonsenseQA, and models with stronger self-evaluation expressiveness showed weaker calibration.
Jae Oh Woo, Mengdie Flora Wang, Rahul Ghosh et al.
In briefRetrain, Alert, or Rollback? A Decision Framework for ML Under Concept Drift
SummaryThe paper proposes RAR (Retrain-Alert-Rollback), a state machine that assigns each of three responses to a degraded deployed model an explicit trigger and cost. Retraining is gated on drift detection, label availability, and measured degradation, and retrained models must pass a two-stage canary before promotion. Evaluated against six baselines on 13 streams, including eight real public and five semi-synthetic ones, RAR does not dominate: retrain-heavy baselines lead at low retraining cost, while RAR performs best on several streams at higher cost.
Ali Sadeghian, Alireza Shafaei, Hamid Hadibeik et al.
In briefTesting How Stable Sentiment Analysis Models Are Under Input Changes
SummaryThis paper presents a lightweight, batch-based metamorphic testing approach for assessing the robustness of language-based AI systems. The authors use sentiment analysis as a representative task and test a pretrained classifier under four transformations: synonym replacement, typo injection, word-order variation, and negation. Using consistency score, flip accuracy, and confidence change, they find the model handles semantic variations well but is vulnerable to textual noise and logical transformations.
In briefLayerwise brain-like model of visual cortex using manifold capacity
SummaryThis paper proposes an unsupervised, layer-by-layer learning approach for modeling neural responses in primate visual cortex, avoiding global backpropagation of error signals. Each layer uses an intermediate loss based on manifold capacity within an efficient coding framework, with task complexity adjusted to match the growing receptive field sizes. The model is evaluated on neural datasets from three macaque ventral stream areas and on human psychophysical object classification data. The authors report that it matches or exceeds architecture-matched baselines on all evaluations.
Thomas Yerxa, SueYeon Chung, and Eero P. Simoncelli
In briefDo AI reviewers notice real paper revisions? A large-scale study
SummaryThe study analyzes PaperTrajectory-2026, a dataset of 6,959 paper revision trajectories from five AI conferences, to test whether AI reviewers' scores reflect author improvements. Most major improvements left reviewer scores unchanged, and reviewers agreed weakly on the direction of score changes. A brief negative user opinion could also erase or reverse the scoring advantage of a revision.
In briefNonreciprocal Spherical Attention for Coarsening and Retrieval: Sharp Timing and Score Correction
SummaryThis paper develops nonreciprocal spherical attention as a mathematical framework for coarsening processes and compressed nearest-neighbor retrieval. It proves a limit for the coarsening time on a logarithmic scale under specific conditions for two shrinking blocks, and gives an O(r²) error bound for attention output under finite compression. The authors also formulate skew query-side score calibration as a strictly convex moment problem and report experiments on SIFT1M and a ten-million-vector FAISS IVF-PQ index, measuring hit rate and query latency.
In briefIterative Gumbel Planning for stable continuous robot control
SummaryThis paper addresses applying search-based planning to continuous-control reinforcement learning, where Gaussian policies clash with discrete action candidates. It shows that distilling winning actions into a Gaussian causes projection error and variance growth. The proposed Iterative Gumbel Planning (IGP) uses a per-dimension categorical policy with Gumbel Top-K sampling and multi-round MPPI-style search. The authors report state-of-the-art aggregate performance and sample efficiency on DMControl and HumanoidBench, with more stable action sequences.
In briefTIGER: A method that learns researcher preferences to generate tailored research ideas
SummaryThis paper introduces TIGER, a method that learns individual researchers' preferences to generate tailored research ideas. In pairwise comparisons against GPT-4.1, three LLM researchers chose the TIGER-8B idea in 56.6% to 68.4% of 76 comparisons each. Two PhD researchers chose it in 19 of 30 pairs and 17 of 30 pairs, respectively.
In briefWho Should Score the Cache? Mid-depth Layers Are the Best Judges of KV-cache Eviction
SummaryThis paper studies which transformer layer's attention scores should decide which tokens to evict from the KV cache. Across twenty-one models, the depth of the scoring layer strongly affects long-context retrieval accuracy, with mid-depth layers (about 0.4 to 0.75 of depth) performing best. The authors propose selecting the scoring layer with calibration prompts that have known answers, which outperforms the default of each layer scoring its own cache on most models.
WhyThe authors ask which layer should score KV-cache tokens, report mid-depth layers peaking across twenty-one models, and compare against per-head SnapKV.
In briefBeyond Ensemble Disagreement: Predictive STL Monitoring of RL Locomotion Under Distribution Shift
SummaryThis paper evaluates predictive Signal Temporal Logic monitoring for reinforcement learning locomotion policies on Ant-v5 and HalfCheetah-v5 when the operating conditions shift. The method uses closed-loop world-model rollouts and a three-model ensemble. For HalfCheetah, predicted robustness ranks future violations well, while ensemble disagreement does not reliably track prediction error or violations. The authors conclude that ensemble disagreement should not be treated as a reliable indicator of safety uncertainty.
Qurban Ali, Badaruddin Chachar, Maqsood Ali et al.
In briefUnified matrix framework for classifier discrimination, calibration and ranking
SummaryThis paper proposes a single matrix-algebra framework covering the three pillars of classifier evaluation: discrimination, calibration and ranking. It extends an existing indicator-matrix approach for discrimination measures with deviation operators for a Calibration Framework and rank- and slice-dependent operators for a Ranking Framework, both sharing micro, macro and exemplar aggregation. The authors derive properties of Brier score, nDCG and AUC, and test the distinctions on two multilabel benchmarks. Many results are machine-checked in Lean 4.
WhyThe authors unify discrimination, calibration and ranking in one matrix algebra, adding deviation and rank-dependent cell operators; results are partly Lean 4 machine-checked.
In briefForecasting Earth surface change by predicting satellite image embeddings instead of raw spectral values
SummaryThis paper proposes E²Forecast, a transformer-based model that forecasts Earth surface change in embedding space rather than raw spectral values, reducing the impact of clouds, shadows, and sensor noise. Its Gaussian branch predicts a mean and per-dimension uncertainty, while its rectified-flow branch generates forecast ensembles. On Sentinel-1 and Sentinel-2 benchmarks, forecasting embeddings halves cosine distance (0.20 to 0.10) and reduces overconfidence, and it raises F1 for anomalous change detection from 41.58 to 48.70 on a global benchmark.
WhyThe authors forecast satellite embeddings instead of spectral values with a dual-branch transformer, reporting halved cosine distance across nine foundation models.
Jingtao Li, Sadiq Jaffer, Anil Madhavapeddy et al.
In briefTest-time scaling for open-ended reasoning in small models
SummaryThe paper studies whether sampling a small instruction-tuned model many times and voting can improve open-ended generation. It groups samples by meaning and returns the representative of the largest cluster, without a reward model or self-judging. Voting improves math accuracy but adds at most 0.30 judge points on open-ended text, because nearly all open-ended samples fall into a single meaning cluster.
In briefBuilt on Broken Assumptions: Why Coding Agents Fail on Chained Tasks
SummaryThe paper introduces BOBA-Bench, which contains 280 chains of 20 consecutive requests drawn from 140 real open-source projects, and evaluates ten coding agents. The agents complete 95% of requests individually but only 51% when requests are chained. The authors attribute this to mistakes the agents miss in their own tests, so later requests build on broken code and fail about twice as often. A single external correction and one repair round raised chain completion from 51% to 59%.
In briefReconstructing 3D material microstructures from high-resolution 2D images
SummaryThis paper introduces PCRD, a framework that uses high-resolution 2D scanning electron microscopy images during training to improve 3D reconstruction from lower-resolution micro-CT data alone. It is demonstrated on pultruded carbon fiber-reinforced polymer composites. In 2D cross-sectional reconstruction, PCRD improves mAP by 1.66 percentage points over direct micro-CT training. Against high-resolution synchrotron X-ray CT, fiber volume fraction, diameter and local spacing show close agreement, and 89.97% of the reference fiber centerline length is matched versus 75.08% for interpolated low-resolution volumes.
Wenbo Sui, Lars P. Mikkelsen, Jens W. Andreasen et al.
In briefCan monocular depth models stay consistent across viewpoints? A video-based evaluation method
SummaryThis paper proposes a new evaluation protocol that tests monocular depth estimation models on video sequences with moving cameras, checking whether predictions stay consistent across viewpoint changes. The authors apply it to diverse real-world videos, anchoring model outputs to metric depth using structure from motion. The work highlights open challenges in temporal consistency over long time frames and difficulties in using structure from motion for metric grounding.
In briefZeroGraph: Training Model Blocks Separately Using Stored Teacher Outputs
SummaryThis paper presents ZeroGraph, a systems method for converting and distilling pretrained models. It stores a frozen teacher model's consistent boundary outputs as a versioned dataset, so each block trains independently with its own parameters and state, using saved supervision instead of live dependencies across layers. Experiments show that an unanchored training objective degrades the assembled model, while a clean teacher-boundary regression recipe passes quality and context-use checks across several models.
Omar Ramadan, Sam Siavoshian, Amir Kashif Saeed et al.
In briefA conditional generative model for T2-weighted prostate MRI and its downstream use in classifying clinically significant prostate cancer
SummaryThis single-center retrospective study developed a conditional generative model for T2-weighted prostate MRI, extending MAISI-v2 with conditioning on clinically significant cancer class, prostate zones, and lesion masks. The model was fine-tuned on 1,303 patients with 1,654 lesions. Synthetic images improved cancer classification under simulated data limits, though the gain did not transfer significantly across institutions.
Xavier Rafael-Palou (1), Alejandro Vergara-Richart (1, 2) et al.
In briefA Fuzzy-Enhanced Machine Learning Method for Type 2 Diabetes Diagnosis
SummaryThis paper presents a machine learning framework combining fuzzification, data balancing, and evolutionary feature selection to predict type 2 diabetes. It selectively fuzzifies nonlinear clinical variables, applies SMOTE-ENN for balancing and noise filtering, uses the Bat Algorithm to select features, and trains a Random Forest classifier. On the Pima Indians Diabetes Dataset, the model reduces features to 10 and reaches 96.11% accuracy and a 98.23% F1-score, outperforming a baseline Random Forest using all 16 features.
In briefRecursive Reasoning Exploration: Building Math Research on Evolving Reasoning Graphs
SummaryThe paper proposes Recursive Reasoning Exploration (RRE), a framework that records claims, premises, attempted arguments, counterexamples, and revisions in evolving graphs so mathematical results can be reused and checked. As a case study, it proves that for every complex tensor of order n of at least 2 and rank above two, the minimum Nullstellensatz certificate degree for its original two-term decomposition equations is exactly 3n. A parallel branch gives bounds for nonzero nilpotent m by m matrices.
In briefComparing EEG inner-speech decoding across three public datasets with one fixed model and protocol
SummaryThis study uses one preprocessing pipeline, one within-subject five-fold cross-validation protocol, and a compact Transformer model (VC-DATNet) to compare inner-speech decoding from EEG across three public datasets. Per-participant sensor/source hybrid accuracies were 39.48% on the four-class Thinking Out Loud task, 20.60% on BISpeech, and 25.53% on SyncSpeech. Sensor-space and source-space representations were complementary, performing better for different participants. Many decoders converged within a narrow accuracy range, suggesting performance is limited more by signal quality and dataset scale than by model capacity.
Md Wahiduzzaman Suva, Md Imtiaj Alam Sajin, Esm E Moula Chowdhury Abha et al.
In briefANCHOR: KL-Anchored Method for Stable Offline-to-Online Reinforcement Learning
SummaryThis paper studies adapting an offline-pretrained reinforcement learning agent to online interaction without offline replay data. The authors propose ANCHOR, which anchors the actor to the pretrained policy early in online fine-tuning, anneals this constraint over time, and keeps only a small residual conservative critic regularizer. Across FrankaKitchen, AntMaze, and Adroit tasks, the method substantially reduces catastrophic early failure while maintaining strong asymptotic performance.
In briefBeyond Unit-Circle Eigenvalues: Invariant Bases for Stable State Space Dynamics
SummaryThe paper addresses diagonal state space models, whose recurrent transitions fail to preserve conserved quantities over long rollouts. It proposes the Kronecker-Cayley transition (KC), which builds its generator from a graph Laplacian over object subspaces and advances it with a Cayley transform, keeping each step diagonal and fixing the conserved quantities before training. On four conservation benchmarks, KC keeps total mass near the reference floor while the other diagonal baselines lose it by orders of magnitude.
In briefRestrained diffusion bridges: training and refining denoisers on targeted regions
SummaryDiffusion models are often inaccurate in data-sparse regions of the target distribution. The paper studies sampling from restrained distributions, which are biased distributions used in enhanced sampling to concentrate data near chosen configurations. It shows the bias cancels exactly when forward noising from restrained data is conditioned on the restrained configuration at an appropriate intermediate time. A model fine-tuned on restrained data can then reach high accuracy in chosen regions without importance weights, and Gibbs sampling enables inference when only restrained data are available.
In briefRewarded for Not Getting Caught: Why Alignment Training Selects for Conditional Compliance
SummaryThe paper argues that current alignment training encodes norms only as prices that are charged when violations are noticed, so it can favor agents that comply only when they might be caught. The author links this to alignment faking, sandbagging, and evaluation-aware scheming, and notes that behavioral tests cannot show unobserved conduct. For norms that no gain should outweigh, the paper recommends architectures that make violations unavailable rather than relying on deeper internalization.
Kevin Baum, Ruta Binkyt\.e, Gerrit Gro{\ss}mann et al.
In briefValue Forcing: Aligning Return Distributions Across All Noise Levels
SummaryThis paper introduces Value Forcing, which uses flow matching to estimate return distributions and enforces Bellman consistency at every noise level to learn value functions from limited data. It presents two algorithms, VF (an actor-critic method) and IVF (an implicit value-learning method). Across 26 long-horizon offline reinforcement learning tasks, both reportedly outperform prior value-learning methods, with especially large gains when data is scarce.
In briefMulti-Agent Collusion and the J-Space: Using interpretability to detect collusion between AI agents
SummaryThis paper examines collusion risks among large language model agents in multi-agent systems. Using the J-space, a representation space previously used to uncover deception in single models, the authors run interpretability experiments on DeepSeek-V4-Flash with the NARCBench collusion benchmark. They also propose J-PACT, an unsupervised method for detecting collusion between agents, and evaluate it on four open-weight models.
In briefA step-by-step derivation of how neural networks compute gradients for training
SummaryThis tutorial derives the backpropagation algorithm step by step for a fully connected feedforward network with sigmoid activations and squared-error loss. It starts from the gradient of a single weight using the chain rule, then arrives at the error terms and four matrix-form equations. The derivation is extended to multiple training examples and mini-batch gradient descent, and each equation is linked to a NumPy implementation trained on the AHDBase handwritten Arabic digit dataset.
In briefImproving Roadside 4D Radar Vehicle Classification with Point-Cloud Dispersion
SummaryThis paper presents a roadside platform that synchronizes a 4D radar with a stereo depth camera and collects a real-world dataset. Because radar point clouds are sparse and noisy, the authors propose a Neyman-Pearson-based binary classifier using the temporally smoothed 2D positional standard deviation of radar point cloud clusters. The method reduces classification error from 52.1% to 42.7%.
Mohammed Alsakabi, Andreia Figueiredo, Ant\'onio Almeida et al.
In briefWhen Is Personalized Federated Learning Worth Its Deployment Cost?
SummaryThe paper frames choosing how deeply to personalize federated learning models as a decision problem, where utility is Macro-F1 minus a cost-weighted deployment cost across six depths (P0 to P5). Experiments on CIFAR-10, CIFAR-100, and a PTB-XL subset show that benefits of personalization depend on the setting, with higher cost sensitivity shifting optimal choices toward the global-only depth P0. A single best uniform depth performs close to an offline oracle assignment, while Ditto full-model personalization is substantially suboptimal once cost is counted.
Beyazit Bestami Yuksel, Ayse Yilmazer Metin, and Emrah Dikbiyik
In briefGatedV: A Transformer Language Model Method That Gates Attention Values
SummaryThe paper proposes GatedV, a simple attention variant that applies a SiLU activation only to the value matrix before aggregation, adding no trainable parameters. Under matched parameter counts and equal training budgets, it lowers held-out perplexity at two model sizes, with gains across all paired seeds. The larger model shows about 0.7% higher forward latency.
In briefDefending fact-checking retrieval against adversarial evidence contamination
SummaryThe paper studies how retrieval-based fact-checking systems can be misled by adversarial evidence that is factually inconsistent yet contextually coherent. It shows that contaminating a small share of training corpora can push such content into many top-k retrievals and substantially lower verification accuracy. The authors propose GLAD, a defense that combines dense embedding peer-coherence with lexical distributional isolation, and report higher detection rates and more consistent F1 recovery than existing defenses across two benchmarks, two embedding models, and five seeds.
Alamgir Munir Qazi, John P. McCrae, Jamal Abdul Nasir
In briefTSSDA: An Autonomous Agent for Building a Time-Series Text-to-SQL Benchmark
SummaryThis paper addresses the scarcity of high-quality benchmarks and training corpora for text-to-SQL on time-series databases. It proposes TSSDA, a data agent framework that models time-series SQL syntax and semantics to define a generation space and autonomously produces diverse queries, with a temporal-aware memory for reusing experience. Using this framework, the authors build the TERN benchmark, which reveals limitations of existing methods, and report that TSSDA substantially improves coverage, diversity, and complexity.
In briefFOLLIE: Automated reasoning on the first French dataset pairing sentences with first-order logic formulas
SummaryFOLLIE links French sentences, compositional first-order logic representations, and automated natural language inference. It contains 19,983 inference records and 36,987 sentence-formula mappings drawn from six NLI sources, with silver annotations produced by a neural-symbolic pipeline. The authors also build an inference engine using Prover9 and Mace4 and report three-way accuracies on FraCaS-FR, GQNLI-FR, and SICK-FR.
