In briefAnalog and hybrid computing in the 21st century: a look at computing that uses physical circuit signals, not just digital logic, for simulation, optimization and energy-efficient AI
SummaryThis interdisciplinary paper reviews modern analog and hybrid (analog-digital) computing. Analog computers represent mathematical variables and relations with physical quantities, such as electronic circuit signals, and can directly represent nonlinear dynamical systems. The authors note that the design constraints of digital architectures increasingly limit performance and energy efficiency, while analog systems can map specific computations onto physical processes using smaller, simpler elements. The article discusses applications in simulating dynamical systems, performing optimization, and enabling more energy-efficient artificial intelligence.
Bernd Ulmann, Olivier Bournez, Corey J. Maley et al.
In briefCLiS: Training-free TPU attention speedup for video diffusion via certified stabilization
SummaryThis paper addresses the slow attention computation in high-resolution video diffusion models, which operate over long token sequences. It proposes CLiS, a training-free, plug-and-play method that fixes per-row stabilizers before the key-value loop and certifies numerical safety per attention head, reverting to online softmax when certification fails. On Wan 2.2 and LTX-2 running on TPUs, CLiS reduces diffusion transformer step latency by up to 19.6% compared with tuned online softmax, without measurable quality loss.
Elisa Tsai, Shao-yu Huang, Daniel Mandragona et al.
In briefAccuracy Drops With Table Size, Not Dialogue Turns: Data Delivery in an Agent
SummaryThe paper compares two ways of passing a tool's structured records to a multi-turn agent: putting full record sets into the model's context, or keeping them in session state and delivering rows through a typed side channel. Across 24 dialogues, the side-channel design answered 92.3% of turns fully correctly, versus 45.7% for the context design. The context design's accuracy collapses as the delivered table grows beyond about 74 rows.
In briefDesigning human-AI teamwork rules that account for how people are swayed by AI outputs
SummaryThe paper introduces a framework called collaborative omnimechanism design for human-AI collaboration, recognizing that human judgments can be influenced by how AI outputs are presented. It is built on a computationally efficient contextual bandit variant of boosting for outcome indistinguishability, reducing a calibration-type condition to weighted calibration. The authors evaluate the approach in three collaboration settings and report that it adapts to different query costs and prediction objectives without retraining, performing comparably to or better than the baselines.
In briefStatistical inference when predictions shape the data they learn from
SummaryThis paper studies performativity, where prediction-informed decisions change the data-generating process, as in policy-making and self-evolving AI systems. The authors build an end-to-end statistical inference framework, deriving central limit theorems and covariance recursions for repeated risk minimization and repeated gradient descent. They propose consistent covariance estimators based on policy perturbations and score matching, and extend prediction-powered inference to this setting. Experiments show that ignoring propagated uncertainty can cause substantial undercoverage even after repeated retraining.
WhyThe authors extend statistical inference to performative repeated risk minimization and gradient descent, deriving covariance recursions and prediction-powered inference, evaluated on synthetic and LL
In briefLocal Spatial TreeSHAP: A Geographically Explainable AI Study of Urbanization-Dependent Flood Resilience Heterogeneity
SummaryThis paper proposes a spatially explainable machine learning framework combining geographically weighted random forests with a new Local Spatial TreeSHAP algorithm to interpret flood resilience mechanisms at the community level. It is applied to 779 communities affected by the July 2023 extreme rainstorm in Beijing. The results show resilience mechanisms vary spatially more strongly than the underlying physical and socio-economic conditions, and identify three distinct resilience regimes across an urbanization gradient.
In briefOnline fair division of chores with upgrade recourse
SummaryThis paper studies online fair division of indivisible chores among n agents, where earlier assignments may be revised but each previously assigned chore must be replaced by a different chore the agent finds no more costly. The authors show that a simple adaptation of round robin achieves envy-freeness up to one chore (EF1) together with a (2-1/n)-approximation of the maximin share (MMS). They also prove that no deterministic algorithm without predictions can do better on this MMS factor.
WhyThe authors give a round-robin adaptation achieving EF1 and (2-1/n)-MMS under upgrade recourse, plus a matching deterministic lower bound and prediction-based improvements.
In briefReal-VAS: A Real-World Video Amodal Segmentation Dataset
SummaryThis paper presents Real-VAS, a large-scale zero-shot evaluation dataset for amodal segmentation of occluded objects in video, built by compositing real-image video clips. Its generation pipeline derives the object mask, occlusion boundary, and composited image from a single unoccluded source, producing consistent ground truth without human estimation or costly 3D reconstruction. The dataset covers dynamic occlusion and container scenarios, including on-surface, articulated, and mobile containment, where object motion is simulated from the container's tracked transformation. The authors state that the dataset and generation code will be made publicly available.
In briefDo multimodal AI models give consistent answers about the same optical spectrum in different formats?
SummaryThe study tests whether multimodal models reach the same decision when one optical spectrum is shown as a table or plot in nanometers or photon energy. It also checks whether models answer "undetermined" when needed data are missing, using measured spectra of silicon metasurfaces and gold nanoparticles. Changing only the representation shifted accuracy by up to 41 percentage points within one model.
In briefPAM-Mamba: An Efficient Mamba-Based Method for Dense Action Detection
SummaryThis paper presents PAM-Mamba, a method for precisely locating multiple overlapping actions in long untrimmed videos. It builds a linear-time O(T) multi-branch temporal encoder on Mamba, with components such as gated learnable positional encoding and bidirectional temporal Mamba. On Charades and MultiTHUMOS it reports 26.82 to 46.92 mAP with 26.1 to 27.9M parameters, 61% to 63% fewer than PAT.
Dhruvraj Singh Rawat, Faegheh Sardari, Armin Mustafa
In briefA paper viewing neural network layer convergence through a spectral renormalization-group lens
SummaryThis paper uses Wilsonian renormalization-group ideas to explain the heavy-tailed power-law spectra that deep neural network layers develop during training, focusing on exponents near 2. It offers statistical and physical interpretations of this value. The authors introduce the SETOL theory, a spectral RG flow connecting HTSR and SETOL to optimizers such as Muon, and a new optimizer extension, WW-PGD.
In briefRankLane: Faster Mixture-of-Experts Prefill on Accelerators via Load-Ranked Expert Lanes
SummaryThis paper presents RankLane, a method for running mixture-of-experts (MoE) model prefill on accelerators with static execution plans. It ranks each device's experts by current load and assigns them to fixed-capacity lanes, using runtime indices to select experts and token rows. Evaluated on four Cloud AI 100 Ultra SoCs with Qwen3-30B-A3B and Gemma 4 26B-A4B, it reports end-to-end prefill speedups of 1.27-4.18x and 1.93-3.46x over a dense baseline, respectively.
In briefCircuit Oracle: answering questions about model internals by automated attribution graph analysis
SummaryThis paper introduces Circuit Oracle, a large language model with a multi-agent harness and task-specific skills that answers natural-language questions by reading attribution graphs through tool calls. It is tested on three safety-relevant tasks: detecting spurious features, eliciting hidden knowledge from taboo-finetuned models, and suppressing refusals through feature-level interventions. On spurious-feature detection it reaches 78.2 ± 2.4% accuracy, beating single-layer ranking baselines, while trained self-explainers remain stronger for secret elicitation. Results are mixed; on refusal suppression it roughly matches a diff-in-means baseline only with best-of-five attempts.
Shariar Kabir, Hong Kiat Tan, Swastik Agrawal et al.
In briefNon-conformal distortion and local orientation reversal in a Transformer's Fourier modular-addition circuit
SummaryThe paper examines the local geometry of a Fourier-based modular-addition circuit in a one-layer Transformer. For each of five key frequencies, the MLP contribution is restricted to the corresponding Fourier input plane and its directional Beltrami coefficient is measured under continuous input perturbations. Trained maps show nonconformal distortion and regions of local orientation reversal, with the median Beltrami coefficient falling from 0.939 at initialization to 0.390 after training.
In briefHow Wording of AI Red-Teaming Instructions Shapes Test Results
SummaryThis paper examines whether the wording of instructions used in AI red-teaming, which probes generative AI systems for harms, affects the results. The authors reviewed 228 prior works and then ran controlled simulations that varied only the instructions while holding risk domain, attacker, and target model fixed, totaling 22,500 attacks. They find that risk granularity, instruction length, the presence of examples, and explicit harm definitions significantly predict attack success rates.
In briefTileWeave: Heterogeneous Tiling for Fast Sparse Prefill on Phones
SummaryThis paper presents TileWeave, a method for speeding up the prefill stage of large language models on smartphones, where long inputs cause high latency and memory use. It splits sparse attention by how much key blocks are reused across queries: heavily shared blocks run on the NPU in coarse tiles, while low-reuse blocks run on the GPU only for the queries that selected them. On RULER and LongBench, its accuracy matches or exceeds XAttention, and on two Snapdragon phones it achieves 1.35 to 3.23 times faster full-model prefill than XAttention.
In briefIs It Cheating or Just Failing? A Stronger LLM Monitor Separates Reward Hacking From Honest Mistakes
SummaryThis paper studies whether an AI agent's inflated score comes from gaming the scoring system or from ordinary mistakes, using a sandboxed task with labelled episodes. Small local monitors failed to separate exploits from honest failures, while Gemini 3.1 Flash-Lite met the pre-registered criterion (AUC 0.80) on the same inputs, though with only 14 exploit episodes the estimate is uncertain. The authors also report that a two-rule deterministic check caught every known exploit and release their environment, labels and logs.
In briefTAP-Out: Tracking Points Anywhere in 360° Using Video Outpainting
SummaryThis paper presents TAP-Out, an offline method for tracking points that leave the camera's field of view. It uses generative video outpainting to extend a narrow-view video into a full 360° panoramic video, locates target objects with video object segmentation, and produces object-centred perspective crops for standard in-frame trackers. On the TAPVid360 benchmark, TAP-Out achieves state-of-the-art results, improving out-of-view tracking accuracy by about 3x over the best prior method.
Finlay G. C. Hudson, James A. D. Gardner, William A. P. Smith
In briefMXCode: Vector Quantization for Efficient MX-Native LLM Inference
SummaryThis paper introduces MXCode, a framework that combines roughly 2-bit-per-weight vector quantization with MX-native FP4 computation to reduce memory and bandwidth demands in large language model inference. It proposes three strategies, PostScale, PreScale, and PreScale+, which trade off quantization quality, scale overhead, and codebook granularity. The authors evaluate MXCode on Qwen3 (235B), DeepSeek-R1 (671B), and GLM-5.2 (753B), and report memory and time-to-first-token results on GB300 hardware.
Gaurav Kolhe, Cansu Demirkiran, Shubhi Tyagi et al.
In briefVESPER: Learning Free-Energy Landscapes Online with a Multi-Tempered Variational Objective
SummaryVESPER is a method that learns free-energy landscapes over collective variables while molecular dynamics simulations run. The evolving model biases the simulations to cross energy barriers more easily, and data from several tempering levels train one shared model without reweighting. The authors prove conditional convergence of the coupled dynamics and test the method on alanine dipeptide, alanine tetrapeptide, methane dehydrogenation on Ni(511), and chignolin.
WhyThe authors couple enhanced sampling with neural free-energy learning via one convex variational objective over multiple tempering levels, reporting tests on alanine peptides, methane dehydrogenation,
Jay L. Kaplan, Gustavo R. P\'erez-Lemus, Mengjian Hua et al.
In briefComparing SAM-based Transformers and CNNs for Segmenting Thin Retinal Blood Vessels
SummaryThis paper compares a plain Vision Transformer encoder (MedSAM), a hybrid Transformer (LiteMedSAM), and three CNNs of different sizes on retinal vessel segmentation across DRIVE, HRF, and FIVES, with a microscopy dataset (VessMap) as a second-domain check. The authors measured centerline recall and precision for thin vessels, along with computational cost and training epochs. They found the pre-trained MedSAM's thin-vessel centerline recall was not significantly different from the best CNN on all four datasets. A compact plain ViT found by architecture search but trained from scratch did not match this performance.
In briefTransformer-Based Reinforcement Learning for Phase I Dose Allocation
SummaryThis paper proposes a single reinforcement-learning policy for Phase I dose-finding trials with varying numbers of ordered dose levels, representing each dose as a token processed by a Transformer encoder. The policy chooses among de-escalation, staying, or escalation, while the final maximum tolerated dose is selected outside the policy using weighted isotonic regression and a safety cap. Simulations show that toxicity-penalized policies reduce dose-limiting toxicity burden, whereas unpenalized policies improve selection accuracy at greater toxicity cost. The authors state that superiority over conventional dose-finding designs is not established.
In briefCalVLA: Training Robot Policies to Signal Their Own Likely Failures
SummaryThe paper presents CalVLA, a method that adjusts the native action-token confidence of the OpenVLA-7B vision-language-action policy using successful and failed rollouts and an Action-Confidence Alignment loss, without an auxiliary failure detector. On a held-out test split from 500 LIBERO-10 configurations, the adapted confidence reaches a failure-discrimination AUROC of 0.7392 ± 0.0480 over three seeds, up from 0.4485 before adaptation. Mean task success falls from 58% to 55%, and the authors state that the results show improved failure discrimination only, not improved control or established safety benefits.
In briefTwo strategies for drawing conclusions: shaping representations versus reasoning in context
SummaryThis paper formalizes two inference strategies: a geometric strategy that gradually shapes representational geometry to reflect relational structure, making later inference cheap, and an algebraic strategy that flexibly reasons over novel relations given in context at extra computational cost. The authors identify opposing signatures of these strategies in transitive inference. Experiments with humans, neural networks trained from scratch, and large language models demonstrate these effects.
WhyThe authors formalize geometric and algebraic inference and report opposing signatures in transitive inference across humans, trained networks, and LLMs, with causal interventions.
In briefRisk assessment and optimal design for AI-agent safeguard systems that account for dependence between safeguards
SummaryThis paper proposes a framework for assessing risk in multi-component AI safeguard systems by accounting for dependence among safeguard decisions. It jointly selects configurations and operating policies for sequential layered screening and selective verification, and provides stochastic comparisons of error probabilities and costs. The authors release DAPOA, an open-source package implementing the framework. Numerical experiments show that independence-based selections violate security requirements in six of eight scenarios, and an empirical study with four LLM safeguards finds dependence-aware selection lowers held-out total cost by 39.6%.
In briefAn agent system for the ML training lifecycle covering model generation, validation, and optimization
SummaryThis paper presents an agent stack deployed at LinkedIn with two independently usable loops: AutoTorch, which generates and migrates PyTorch models and training pipelines, and AutoResearch, which optimizes models through expert-steered or autonomous experiments. AutoTorch uses executable validation hooks to iteratively repair its outputs before engineer review and full-scale training. Across 115 workflow runs, AutoTorch completed generation and full-scale training in 92, and 77 of those met post-training quality requirements. In a separate cohort of 100 production pipelines, AutoResearch improved the primary task metric on held-out data for 28. The authors estimate that elapsed end-to-end migration time fell from about eight weeks to two weeks.
In briefLF-VideoQA: A benchmark testing long-video understanding with long-form theory-of-mind questions
SummaryThis paper introduces LF-VideoQA, a dataset for open-ended, long-form question answering about long narrative videos. Its questions test theory of mind, requiring models to combine evidence spread across the video. The dataset includes over 6,300 questions on 94 recent films, each with a rubric for scoring recall and precision. Experiments show the task is very difficult, even for a model with access to the screenplay, with recall being the main bottleneck.
In briefMoRE-GS: Routing Motion-Regime Experts for Future Gaussian Splatting
SummaryThis paper presents MoRE-GS, a framework for forecasting future views of dynamic 3D Gaussian scenes. It reconstructs a Gaussian representation from only the observed temporal prefix and predicts control-keypoint motion using four complementary experts: persistence, a coefficient-space Transformer, a graph convolutional network, and a structured Transformer with rigid and non-rigid constraints. A routing table and persistence gate combine the experts' outputs, and the authors report improved performance over other methods on D-NeRF and HyperNeRF scenes.
In briefUnlearning via extrapolated shrink-and-heal lookahead
SummaryThis paper studies unlearning in large language models, aiming to make a model behave like one retrained without its forget set. The proposed method, SHIFT, shrinks the fine-tuned model toward its pre-fine-tuning weights and briefly relearns the retain set to remove forget-specific effects. It extrapolates the lookahead's per-token changes into distillation targets and refines them iteratively. Under a unified OpenUnlearning protocol against eleven baselines, SHIFT achieves the highest aggregate score on all TOFU settings and the closest alignment with the oracle, and it preserves the most retain knowledge on MUSE.
In briefReusing Entity-Replacement Responses Across Prompts: Tests whether internal changes from entity replacement in one English factual question transfer to another question about the same entity, with model weights fixed.
SummaryWith weights fixed, the paper adds hidden-state changes from two source questions to a target question about the same entity and compares this with direct reuse of output-score changes. Using other-relation sources, hidden reuse removes part of the no-change error on small models, but on small plain-format checkpoints it mainly reproduces the shift both relations share rather than reliably lowering relation-selective error. Applying the same relation-neutral change at both questions' entity tokens lowers relation-selective error more than random or other-replacement changes, suggesting dependence on the receiving question.
In briefSIEVE: Compressing Visual Tokens Across Multiple Images While Keeping Each Image's Evidence
SummaryThis paper presents SIEVE, a training-free, single-pass method for compressing visual tokens in vision-language models that process multiple images. It pools each image into a compact grid, selects representative tokens using prompt attention under equal per-image quotas, and merges unselected tokens into similar representatives within the same image. The authors report that SIEVE achieves the highest average scores at a 6.25% token budget across several multi-image and single-image benchmarks, and on LLaVA-OneVision-7B it exceeds the uncompressed model by 2.01 points at a 12.5% budget with 6.6x fewer prefill FLOPs.
Nitish Shukla, Surgan Jandial, Nhat-Tan Bui et al.
In briefLimits of Test-Time Compute for Model Self-Evaluation
SummaryThe paper finds that adding more candidate answers does not consistently close the gap between final-answer accuracy and pass@N. It measures how well models separate their correct and incorrect generations using ranking error, which it decomposes into bias and variance components. Based on this analysis, the authors propose Scorpio, a method that jointly scores randomly grouped candidates, and report gains in verification, tree search, and aggregation on math, code, and puzzle benchmarks.
Anirudh Buvanesh, Khaled Rouissi, Sujin Yun et al.
In briefQueering the Audits: Community-Based Auditing of AI Harms to Queer Communities
SummaryThis paper proposes a participatory auditing approach for finding harms that AI systems cause to queer communities, such as erasure, misclassification, and discrimination. The authors ran a workshop at EurIPS 2025 where 16 queer community members audited four case studies using the 4Cs harm taxonomy (Content, Conduct, Contact, Contract). Participants traced most harms to problem definition and data collection, and found contracts that extract value from vulnerable groups while offering little recourse.
Organizers of QueerInAI: A Pranav, Alissa A. Valentine, Alex Markham et al.
In briefSPARK: Self-Correcting Embodied Memory via Physical Re-Observation
SummarySPARK is a multi-episode framework that corrects errors in retrieval-augmented memory used by embodied agents. After an initial episode, a semantic critic flags problematic regions of the knowledge base, a planner generates targeted revisits, and selective merging adds, removes, or replaces information. Tested on four outdoor AirSim maps under two degraded viewpoint conditions, one corrective episode improved embodied question answering on every map while storing 13-41% fewer objects than append-only updating. Extending correction through Episode 4 brought every map above its undegraded reference performance.
Dmytro Kurdydyk, Aadi Palnitkar, Sebastian Scherer et al.
In briefDEPLOY-RL: Certifying Reinforcement Learning Controllers for Safety-Critical Industrial Processes
SummaryThis paper presents DEPLOY-RL, a post-training framework for deciding whether a reinforcement learning policy is safe to deploy in safety-critical continuous processes. It focuses on finding where failures occur rather than measuring performance everywhere, using a sampling method that concentrates on certification-critical boundaries. The framework uses conformal risk control for finite-sample false-go guarantees and a Deploy/No-Deploy/Abstain decision with a PID/MPC fallback. In simulations on a papermaking digital twin and the Tennessee Eastman benchmark, it reports a 4.4% false-go rate while retaining 88.6% policy coverage.
In briefUsing graph neural networks to classify large Boolean functions under NPN equivalence
SummaryThis paper studies NPN classification, which determines whether Boolean functions can be transformed into one another by negating or permuting inputs and negating the output. Exact methods require exhaustive enumeration whose cost grows exponentially with the number of inputs. The authors recast the task as NPN action prediction and use a bidirectional heterogeneous graph neural network on AND-Inverter Graphs to predict negations and permutations, supported by graph rewriting. Experiments show the prediction is learnable, improves over heuristic baselines, and scales to functions with up to 16 inputs.
In briefBreaking irregular time series into interpretable temporal patterns using mixture-of-experts
SummaryThis paper introduces MoTIF, a framework that uses mixture-of-experts attention to split irregularly spaced sequential data into cyclical and range-based temporal patterns spanning minutes to years. On synthetic data and a public credit card transaction dataset, MoTIF outperforms baselines on next-value and next-amount prediction. Its learned routing also shows that fraudulent transactions have bursty, calendar-agnostic timing compared with the steadier periodicity of legitimate ones. The authors argue this approach keeps predictive accuracy while making sequence models more auditable.
Jenna L. Ballard, Zijie Pan, Shivani Shekhar et al.
In briefC3: Speech-guided cache retrieval for faster long-audio processing in speech models
SummarySpeech language models store audio far more densely than text, so their attention cache dominates memory and compute during long generations. The paper proposes C3 (Cluster-Compete-Compete), which indexes audio cache entries hierarchically, has query heads sharing a cache head compete for a shared retrieval budget, and adaptively splits and merges retrieval resolution. Tested on long-form speech recognition, translation, and summarization across four speech language models, C3 gives the best overall results among retrieval methods at comparable access budgets, with especially large gains under tight budgets. It also improves long-context decoding efficiency and memory scaling.
Dancheng Liu, Markus M\"uller, Rupak Vignesh Swaminathan et al.
In briefDéjà Vu: Cycle-Consistent Re-ranking for Matching Semantic Keypoints
SummaryThis paper introduces Déjà Vu, a test-time decoding module that improves semantic correspondence, the task of matching equivalent keypoints across images of the same object category. The authors observe that correct matches often appear among top-ranked candidates even when greedy top-1 selection misses them, and they frame matching as a recall-then-rerank problem. The method selects an adaptive candidate set and re-ranks it using a graded cycle-consistency score computed from the existing similarity matrix. It improves three baselines on SPair-71k, its geometric subset, and PF-Pascal under unsupervised and weakly supervised settings.
In briefTesting whether audio-language models trust what they hear or what they're told in the same recording
SummaryThis paper introduces SAC-Bench, a benchmark of 616,031 audio clips covering 188 sound-event classes across seven semantic, temporal, and spatial tasks. Each recording is paired with spoken claims that either match or contradict the acoustic content. Across five open-weight audio-language models, accuracy on event identity drops by 24.8 to 46.5 percentage points under conflicting claims. The authors also find that prompt interventions can restore acoustic accuracy in some cases but can also reduce claim-aligned answers without improving grounding.
In briefBeyond Semantic IDs: Atomic Creator Tokens for Generative Recommendation
SummaryThis paper addresses the cold-start problem of recommending creators a user does not yet follow, arguing that the Semantic ID layer used in generative recommenders, not the language model, is the main bottleneck. The authors propose Atomic Creator Token (ACT), which gives each creator its own vocabulary token trained end-to-end, avoiding SID collisions and blocked gradients. In a controlled comparison on production data, ACT outperforms a matched SID baseline on creator-level Recall@15, and it also leads on the public KuaiRand dataset. A live A/B test shows directionally positive results on every reported metric.
WhyThe authors propose per-creator atomic tokens trained end-to-end, comparing against semantic IDs on production data and KuaiRand, plus a live A/B test.
In briefGAVAGAI: Aligning Words to Image Regions in Infant-Scale Vision-Language Models via Balanced Optimal Transport
SummaryThis paper proposes an alignment objective that links words to image regions in vision-language models by treating each scene as an assignment problem solved with entropic optimal transport. Words may be assigned to a NOTHING option instead of any region, and each region's share of assignment mass is capped. On the SAYCam infant video corpus, adding this alignment loss to a captioning model raised word-to-picture accuracy from 32.4% to 48.7% and localization IoU from 0.18 to 0.38. A contrastive model trained on the same child's data still performs better on its own benchmark.
In briefOverview of cyber-physical attack detection in wide-area measurement systems
SummaryThis paper examines wide-area measurement systems (WAMS) in the Smart Grid, which use phasor measurement units (PMUs) to collect time-stamped data on voltage, current, frequency, and phase angle. It surveys cyber and physical anomalies in WAMS and their effects on system integrity and availability. The paper reviews anomaly detection approaches for historical and streaming PMU data, including conventional machine learning and hybrid methods, and discusses their strengths and drawbacks. It also highlights the importance of handling missing data and concept drift.
In briefDecentralized Coordination for Multiple Coding Agents in Shared Codebases (Paper)
SummaryThis paper examines how coding agents working in one shared repository, each serving a separate developer, can produce conflicting code when their changes are combined. The authors introduce ParallACT, a benchmark of complementary and contradictory coding tasks without an orchestrator. Parallel agents perform substantially worse than a single-agent oracle, with the sharpest drop on contradictory tasks. Telling agents about peers helps little, and a communication channel recovers only a small part of performance at extra token cost.
Rishabh Mediratta, Lingyun Xiao, Peter Stone et al.
In briefRaMMOBO: A Fast Multi-Objective Bayesian Optimization Method
SummaryThis paper introduces RaMMOBO, a multi-objective Bayesian optimization method for approximating the Pareto set of several expensive black-box objectives. Its acquisition function lower-bounds expected hypervolume improvement (EHVI) while needing only linear cost in the number of objectives. Experiments on synthetic and real-world benchmarks show it matches or outperforms state-of-the-art methods, with the largest gains at higher objective counts. Acquisition optimization is reported to be up to two orders of magnitude faster than qLogNEHVI.
In briefLow-precision floating point formats rising with deep learning and their standardization
SummaryFloating point arithmetic has long dominated scientific computing because it balances precision and numerical range. Since 1985, 32-bit and 64-bit formats have been the industry standards. Over the past decade, deep learning has become the dominant workload, and narrower formats with lower precision and range have grown more common. The article discusses why this shift is happening, how 16-bit, 8-bit and 4-bit formats were developed, and efforts to standardize them.
In briefCross-Lingual Persona Effects in LLM Negotiation: Whether the Same Instruction Changes Negotiation Behavior in English and Korean
SummaryThe paper has identical Gemini 3.6 Flash buyer and seller agents negotiate 12 CraigslistBargain scenarios in English and Korean, crossing seller urgency with three buyer personas across 432 attempted negotiations. Korean negotiations reached agreement far more often than English (77.3% vs. 52.5%), mainly with neutral and headstrong buyers, and sellers conceded more in Korean. The authors argue that changing only the interaction language can make a prompt-defined social policy produce different simulated negotiation behavior.
In briefA task-agnostic framework for counterfactual fairness in transformers using activation steering
SummaryThe paper proposes MICF, a framework that uses mechanistic interpretability tools to reformulate counterfactual fairness constraints in terms of transformer activations. It unifies previous approaches into a single pipeline for fairness analysis, without requiring a prior causal model or a specific transformer architecture. The authors evaluate MICF on real-world tabular data and report that counterfactual fairness improves across different prediction tasks.
In briefTeach, Abstract, Apply: Teaching cobots new tasks through dialogue
SummaryThis paper presents the Teach-Abstract-Apply (TAA) framework, which lets human operators teach collaborative robots new workflows in natural language, abstract them into reusable skills, and apply those skills in different settings. The authors test the framework on a collaborative structure-building task using the open-weight Qwen3-Coder and the closed GPT5.2-Codex models in self-play and cross-model setups. Skill abstraction improved task success rates and reduced interaction cost. A preliminary human-in-the-loop study suggests the pipeline also works with human instructions.
Chalamalasetti Kranti, Sherzod Hakimov, David Schlangen
In briefSeeing the Unseen: Unified Visible–Invisible Motion for Physically Consistent Video Generation
SummaryThis paper presents Seeing the Unseen, a unified framework for generating videos that follow physical dynamics. It first uses a lightweight Particle-based Graph Dynamics Simulator (PGDS) to learn physical interactions from data and predict 3D motion trajectories. It then introduces a Visible–Invisible Motion Field (VIMF) that represents both motion visible in the input frame and motion emerging over time, and feeds these trajectories into a diffusion-based generator. The authors report improved motion accuracy, interaction consistency, and fewer physical artifacts than strong baselines.
In briefTADA: A Memory-Technology-Aware Estimator for DNN Accelerator Performance and Energy
SummaryThis paper presents TADA, a estimator that predicts latency and energy of deep neural network accelerators for a given workload, hardware configuration, and memory technology. It is trained on a synthetic dataset spanning memory technologies, using technology signatures of read energy, write energy, and capacity as inputs. On 21 unseen technology curves, it achieves 1.5–4.4% mean absolute percentage error for energy and 0.7–1.2% for latency. It also reports evaluation throughput over 10^4 times higher than the simulator.
Tanvi Sharma, Surya Selvam, Anand Raghunathan et al.
In briefTaxonomy of failures when AI agents act for users
SummaryThis paper examines how AI agents that write code, manipulate user interfaces, and make transactions for users can fail. The authors analyze 695 documented incidents from two public repositories and derive 12 failure modes across three interaction stages, coding the autonomy level at which each arose. They find that all 12 modes appear at both low and high autonomy, and that increasing autonomy changes the form each failure takes.
In briefMeasuring How Well Nonlinear Embeddings Preserve Local and Global Structure
SummaryThis paper examines how well dimension reduction methods such as t-SNE, UMAP, and PCA preserve local and global data structure. The authors derive a local fidelity score L and a global fidelity score G from one shared locality curve, and propose an LG reversal test to check whether two maps trade off locality and globality and at what scale to switch between them. Across 156 maps of twelve datasets with up to 1.38 million points, G agrees with established global scores at 0.89 and 0.96. The test confirms the usual claim on seven datasets and finds neighbor embeddings more global on two retail datasets and two rolled-up manifolds.
WhyUnifies local and global fidelity as one curve statistic and proposes an LG reversal test, which the authors validate on 156 maps across twelve datasets.
In briefLocating Photos from Narrow Views: Cross-View Geolocalization Using a Global Dataset
SummaryThis paper studies cross-view geolocalization, which finds a ground photo's location by retrieving the matching satellite image, useful where GNSS signals are unavailable. The authors introduce CVW500k, a globally distributed dataset of over 500,000 ground-satellite image pairs that uses limited field-of-view ground images, closer to real cameras. They also propose GeoQueryNet, a query-based cross-view fusion transformer that uses a CLIP-based encoder with LoRA adaptation and learnable query tokens to align features across the two viewpoints.
In briefVerusWeave: A Benchmark for Repairing Proofs in Concurrent Verus Programs
SummaryThis paper introduces VerusWeave, the first benchmark dedicated to repairing proofs for concurrent code written in Verus, drawn from four open-source projects. It contains 75 focused tasks and 17 full-module variants, where selected proof annotations are removed and coding agents must fill the gaps while executable code and specifications are kept intact. An initial evaluation of four general-purpose coding-agent harnesses reached 75–85% success on focused tasks and 29–47% on full-module variants.
Pedro Ribeiro, Martim Afonso, Andr\'e Vasconcelos et al.
In briefA method for quantizing large models that handles weight outliers and sparsity
SummaryPost-training quantization reduces the memory needed to deploy large models by storing their weights in fewer bits, but this causes errors that hurt performance. The paper splits the trained weights into a quantized matrix, a low-rank component, and a sparse matrix that captures outlier rows, using simulated annealing to search for good decompositions. Tests on Llama 3.2 1B and 3B, Llama 3.1 8B, and Qwen2.5 7B with 4-bit NF4 and FP4 quantizers show it can clearly outperform direct quantization with minimal extra memory.
Leonidas Tsigkounakis, Abraham Jaeger Mountain, Yilang Zhang et al.
In briefOCELOT: Planetary-Scale Species Distribution Modeling with Satellite Embeddings
SummaryThis paper presents OCELOT, a model that predicts where species occur across the globe by fusing spatial coordinates with satellite foundation embeddings. It is trained on about 2 billion GBIF records covering 100,000 taxa. On spatial holdouts testing extrapolation to remote areas, satellite embeddings improve performance by 51% over coordinate-only models and outperform classical bioclimatic rasters. The model also matches or exceeds existing baselines on interpolation benchmarks.
Bart van Merri\"enboer, Lauren Harrell, Keith Anderson et al.
In briefDELBERT-2: Pretrained Fingerprint Language Models for DEL Protein Binder Prediction
SummaryThis paper presents DELBERT-2, a pretrained language model that converts sparse molecular fingerprints into token sequences using a ModernBERT encoder. It is trained with masked language modelling on 2.5 million molecules from the AIRCHECK DEL corpus. The model is evaluated on six targets under three out-of-distribution protocols, where it improves PR-AUC and NDCG@1000 over LightGBM baselines and transformers trained from scratch. The largest gains appear in the most stringent out-of-distribution settings.
In briefConstructive Automated Theorem Generation in First-Order Logic
SummaryThis paper proposes a constructive framework for generating theorems automatically in first-order logic, built from minimally unsatisfiable clause sets formed from selected literals. Some clauses serve as premises and the negation of a clause or conjunction of clauses as the conclusion, so generated theorems are correct by construction without separate proofs. The authors extend their earlier Δ1 generator with methods Δ2 to Δ8, implementing and releasing Δ1 to Δ4, and present a general Δ-class method. Combinatorial analysis indicates that the approach can systematically produce large numbers of distinct theorems.
In briefOutstrumental Variable Regression: Identifying Causal Effects Using Outcome Mechanism Shifts Across Environments
SummaryThis paper introduces Outstrumental Variable Regression (OVR), a method for estimating the causal effect of a treatment on an outcome from observational data when unobserved confounding is present. It shows the effect can be identified when the outcome mechanism varies across a finite set of environments, even if the treatment distribution is unchanged. The procedure computes conditional expectations per environment, forms a contrast vector against a reference environment, and regresses the outcome on the treatment while adjusting for the contrast and environment. The authors prove conditions under which OVR recovers the true effect, show a finite-sample version is consistent under regularity conditions, and demonstrate it on synthetic and real-world data.
Felix Schur, Linus K\"uhne, Julius von K\"ugelgen et al.
In briefA generative foundation model for brain imaging analysis (BrainGFM)
SummaryThis paper proposes BrainGFM, a reusable and reconfigurable generative foundation model for computational brain imaging. Pretrained on about 1.3 million MRI and CT slices, it supports task-aligned data synthesis, low-data predictive learning, and image reconstruction. It improved segmentation and classification across 48 downstream tasks and was validated on 4,796 stroke patients from six centers. In a human-AI study, BrainGFM assistance raised radiologists' stroke-subtyping accuracy by 14% in relative terms.
WhyThe authors propose a reusable generative brain-imaging prior, validated across 48 downstream tasks and multicenter stroke cohorts with a human-AI study.
In briefKINDER: A kernel-based method for fair representation learning by erasing sensitive information in prototype space
SummaryThis paper introduces KINDER, a method that directly removes sensitive-attribute information from the intermediate representations of deep models. It operates in a random Fourier feature space, estimating a sensitive subspace from protected-group prototypes and projecting representations onto its orthogonal complement. The authors state the method works with any model containing an intermediate feature space and report experiments across supervised and self-supervised settings, multiple tasks, and visual, textual, and tabular data.
Abtin Mogharabin, Alp Toykan Kaplan, Jiaee Cheong et al.
In briefModel collapse can be avoided even as real data shrinks, via a sharp criterion for iterative Gaussian estimation.
SummaryThe paper studies whether repeatedly training on a mix of real data and a model's own generated samples causes model collapse, focusing on iterative estimation of a multivariate Gaussian's mean and covariance. Each iteration mixes real samples with samples from the previous estimate without distinguishing them. The authors give a single equivalent condition for consistency of the estimators and show collapse is avoided for any fixed positive real-data fraction, or a vanishing fraction that does not vanish too quickly. However, the mean estimator's convergence slows as the real-data fraction approaches zero.
WhyThe authors give a single equivalent condition for consistency of iterative Gaussian mean and covariance estimation with mixed real and generated data, supported by CIFAR-10 diffusion experiments.
In briefSketchForge: training-free method turning images into editable vector sketches
SummarySketchForge is a framework that converts a raster image into a resolution-independent, stroke-editable SVG sketch without any additional training or fine-tuning. It optimizes Bezier curves for each input image at test time, separating structural contour strokes from free-form texture strokes and compositing them via multiply-blend layering. The paper reports that it outperforms existing vector-based baselines in structural fidelity and generation quality.
Quang-Thinh Nguyen, Duy-Khang Do, Hien Trinh et al.
In briefTaurus: LLM-in-the-loop CT annotation matching coding agents at half the cost
SummaryThe authors present Taurus, a framework in which an automated pass labels aortic CT scans and a large language model reviews the result. They evaluated it by blind label recovery on six public datasets (939 volumes), scoring each approach against each dataset's own classes. The final version smooths label surfaces with a fixed rule and asks the model multiple-choice questions when automated checks flag large errors. It showed no difference from a free-form Opus agent on the primary metric, at $1.22 per case against $2.71.
In briefSketchBloom: Training-free vector sketch outpainting with stroke-level style control
SummarySketchBloom is a training-free framework that extends sparse sketches and outputs an editable SVG made of Bezier strokes. It combines pretrained raster outpainting and style-transfer guidance with per-instance vector optimization to keep contours coherent and match a reference style. On LineArtOut-1K, with 1,024 input and style combinations, it reports better perceptual, semantic, and structural fidelity than the evaluated vector baselines. Runtime is about six minutes per pair on a single NVIDIA H200 GPU.
Quang-Thinh Nguyen, Duy-Khang Do, Tam V. Nguyen et al.
In briefMEMENTO: Learning to Manage Context for Long-Input Agents
SummaryThis paper presents MEMENTO, a long-context agent built on Qwen3-8B that manages its fixed context window by compacting its history into self-written notes. It is first trained on synthetic trajectories with hindsight-generated notes, then refined with reinforcement learning using a reward covering answer correctness, note quality, and confidence calibration. On book-level question answering across four benchmarks, it raises Qwen3-8B's accuracy from 36.9% to 67.8% while never holding more than 20K tokens in context.
Yunfan Zhang, Taehoon Hwang, Kathleen McKeown et al.
In briefOmniPiano: A benchmark for dexterous robotic piano playing under standard, robust, safe, and multi-agent reinforcement learning.
SummaryThis paper introduces OmniPiano, a benchmark for dexterous robotic piano playing that supports one to five Shadow Hands. It covers standard, robust, safe, and multi-agent reinforcement learning within a shared task family, with at least 912 task settings and 36 baseline algorithms. The benchmark also includes LLM-based agents, and experiments show that current state-of-the-art methods still struggle on challenging piano-playing tasks.
In briefRE-SciVQA: Better Answers About Scientific Figures Using Relevant Retrieval and Light Model Tuning
SummaryThis paper presents RE-SciVQA, a method for answering questions about figures in scientific papers. It expands queries with relevant section headers and adjusts retrieval scores using figure references and caption alignment, while fine-tuning small language models with LoRA. On the SciVQA benchmark it reaches a ROUGE-1 F1 of 73.9%, ahead of chart-focused models such as UniChart, ChartGemma and Matcha-ChartQA. Compared with a standard RAG baseline, it improves exact match by 11.9% and fuzzy match by 29.5%.
In briefIsoAlign: Isoform-aware DNA-protein alignment using joint-embedding predictive architectures
SummaryThis paper introduces IsoAlign, a multimodal framework for pretraining genomic foundation models by aligning transcripts with their corresponding protein isoforms to learn DNA representations. It uses two Joint-Embedding Predictive Architectures: DP-JEPA, which predicts protein representations from DNA via contrastive alignment, and IM-JEPA, which models relationships among alternative isoforms of the same gene through masked protein prediction. The learned representations were evaluated on tasks including pathogenicity, fitness, gene expression, and TF binding prediction, and the authors report improved downstream performance from isoform-aware cross-modal alignment.
Niousha Sadjadi, Arman Seyed-Ahmadi, Tony Tu et al.
In briefReal-time source-free polyp segmentation using directional dual-head supervision
SummaryThis paper presents MedRT-SFSeg, a real-time source-free domain adaptation framework for polyp segmentation built on a dual-head YOLO architecture. It introduces Mask-DHF to fuse precision- and recall-oriented predictions, DURR to use dual-head disagreement as a directional cue for adjusting masks and recovering missed instances, and Seg-MARD for mask-guided, region-stratified feature regularization. In bidirectional adaptation experiments between Kvasir-SEG and CVC-ClinicDB, the method reports Dice scores of 80.43% and 80.77%, improving on the non-adapted baseline while running at over 62 FPS end-to-end.
Hai-Dang Nguyen, Nhat-Khoi Le, Huu-Thuan Huynh et al.
In briefProLand: a new optimizer for efficient LoRA fine-tuning of language models
SummaryThis paper introduces ProLand, a retraction-free optimizer for LoRA fine-tuning that keeps low-rank factors near the Stiefel manifold. It combines coordinated factor updates with Muon-style spectral directions, separating objective optimization from orthogonality restoration. Experiments cover four GLUE tasks on DeBERTa-v3-base and seven language benchmarks on Qwen2.5-3B, Llama-3.2-3B, and Moonlight-16B-A3B across several adapter ranks. The authors report that ProLand achieves the highest average score among compared methods in every tested model-rank configuration.
In briefLocalized anomaly detection using a calorimeter normalizing flow
SummaryThis paper uses a normalizing flow trained as a fast calorimeter simulator to compute each shower's likelihood, which serves as an anomaly score. Because the global likelihood sums over all voxels, a signal confined to a few voxels can be lost in the rest of the shower's fluctuations. The authors define a regional score by summing the squared latent coordinates of the same flow within small detector regions around where the particle enters the calorimeter. At the same background efficiency, this regional score selects more than 95% of the signal, compared with less than 5% for the global CaloFlow likelihood used in prior work.
In briefSpectral and Conformal: Neural Network Methods for Calibrated Trajectory Prediction and Control-Point Intervals
SummaryThis paper presents a framework for quickly assessing the feasibility and safety of trajectories for autonomous systems. It combines a spectral-backbone neural network that learns trajectories and quantile bands with built-in constraints, a modified conformal quantile regression for calibrated joint trajectory coverage, and interval arithmetic to extend noise-free coverage guarantees to control points. On a Raspberry Pi Zero 2 W, quadrotor trajectory inference under a closed-loop LQR takes 3.45 ms, and the full pipeline takes 9.81 ms, or about 100 Hz.
In briefEditing Cell Nucleus Images Across Gene and Pathomics Data in Spatial Transcriptomics
SummaryThe paper examines how tissue sectioning geometry creates an apparent correlation between gene expression and nuclear pathomics features in spatial transcriptomics. The authors apply the Manifold-Aware Diffusion (MAD) framework to condition-driven editing of cell nucleus images, modifying one feature group while preserving the other. Evaluated on podocyte cells from CosMx SMI kidney tissue, MAD achieves MAE 0.6636 and FID 25.19 for pathomics editing, outperforming all compared baselines.
In briefOPINE: Orthogonal Projection in Invertible Nonlinear Embeddings for Turbulent Flows
SummaryThis paper introduces OPINE, a method that learns an invertible nonlinear coordinate transformation and then applies proper orthogonal decomposition (POD) in the transformed coordinates to compress physical fields such as turbulent flows. Because the transformation is invertible, all information loss occurs in the POD projection, so one trained model yields reduced coordinates of every smaller dimension by ordering its basis. The authors prove that their alternating training update converges locally to the exact POD subspace, and they evaluate the method on synthetic manifolds, Kuramoto-Sivashinsky and Burgers dynamics, two-dimensional Kolmogorov flow, and three-dimensional minimal channel flow.
WhyCombines an invertible nonlinear map with POD for single-model reduced coordinates; the authors report comparisons against POD and autoencoders on several turbulent datasets.
In briefGazeRIB: A Systematic Test of Using Multiple Radiologists' Eye Gaze to Train Chest X-ray AI
SummaryThis paper proposes GazeRIB, a method that trains medical imaging models by masking regions radiologists did not view, while using full images at inference. Eight experts read the same 80 chest X-rays, and reader panels of one, three, five, and eight were compared across architectures and seeds. Every panel improved macro F1 for four-class severity prediction on a separate cohort, while external validation results varied by dataset.
Nattakorn Kittisut, Roy M. Gabriel, Marly van Assen et al.
In briefContinuous-space latent reasoning for unified multimodal image generation (LARC)
SummaryThis paper addresses the difficulty of unified multimodal models in using language-based chain-of-thought reasoning for image generation, attributing it to a mismatch between discrete token reasoning and continuous image generation. It proposes LARC, which reasons over continuous hidden states through two stages: curriculum supervised fine-tuning that gradually shifts token-level reasoning to latent reasoning, and information-gain reinforcement learning that self-evolves without ground truth supervision. Experiments on text-to-image benchmarks report improved generation quality, prompt following, and compositional fidelity, with state-of-the-art GenEval results, a 7-point gain over BAGEL, and better test-time scaling than token-based generation.
In briefFlide: Faster Decision Inference via Exact State Forking
SummaryThis paper presents Flide, an inference engine that speeds up decision models used in LLM agent systems. It separates the logical tree structure of a decision request from its physical execution in a single flattened token stream, and uses ForkAttention with GPU-resident indices across standard attention and recurrent Gated DeltaNet layers. According to the authors, Flide answers queries up to 9.5 times faster than vLLM and SGLang on public hybrid and decision models, and stays closest to a float32 reference in total variation distance.
In briefLatent Sparse Attention: A Faster Attention Method for Ultra-Long Contexts
SummaryThis paper proposes Latent Sparse Attention (LSA), which clusters keys hierarchically in a learned latent space so queries score clusters instead of every key, reducing attention cost at ultra-long context lengths. Two variants are introduced: Distilled-LSA (D-LSA), applied to pre-trained models via distillation, and Native-LSA (N-LSA), trained only on the next-token loss. The authors report per-layer prefill and decode speedups for D-LSA at 10M tokens relative to the original sparse and full attention, and state that N-LSA matches full attention accuracy at 128k tokens with a 1.6x per-layer training speedup.
Ales Kubicek, Andrea Caciolai, Liangzu Peng et al.
In briefHierarchical World Model with Implicit Dynamics for Long-Horizon Robot Planning
SummaryThis paper introduces ID-WM, a hierarchical world model that separates state-space reachability from low-level control. A high-level model samples feasible transitions across the state manifold, while a low-level world model acts as an inverse dynamics solver to turn these transitions into precise actions. The authors report that ID-WM synthesizes novel long-horizon plans from short segments, with a 38.5% improvement over state-of-the-art policy and diffusion-based planners and a 135% gain over single-level world model MPC across three navigation and manipulation suites.
In briefSemEllipsoid: A Method for Semantic-Aware 3D Shape Abstraction Using Ellipsoids
SummaryThis paper addresses shape abstraction, which represents 3D shapes with a set of primitives. Prior methods focus on geometric accuracy and often let primitives span several semantic parts. The authors propose SemEllipsoid, which uses ellipsoidal primitives and injects semantic constraints from large-scale foundation models so each primitive fits a single semantic part. The method is non-learning-based, requires no network training, generalizes to arbitrary shape categories, and supports adjustable levels of abstraction.
In briefRoboArchitect: Automatically Designing Robot Agents with Persistent Memory
SummaryThis paper presents RoboArchitect, a system in which a coding model observes a robot's rollouts, diagnoses failures, and iteratively revises the control harness around a frozen robot policy. Without updating any model weights, average success rises from 9.7% to 43.3% on unseen scenes and from 12.7% to 40.2% on unseen tasks across multiple benchmarks and policies. The evolved harness also develops persistent memory that carries learned experience into later decisions.
WhyThe authors automate harness design around a frozen robot policy, reporting large success gains on unseen scenes and tasks across benchmarks.
In briefReLift: Dense 3D Tracking via Lifted Trajectory Repair
SummaryThis paper addresses dense 3D tracking from monocular video and presents ReLift, a feed-forward framework that treats the task as repairing and completing geometric trajectories. It starts from lifted trajectories built with pretrained 2D trackers and 3D geometry models, and uses a trajectory corrector to fix corrupted positions and fill in motion under occlusion. The model is trained only on geometric inputs and can improve trajectories from different trackers and geometry providers without retraining. On WorldTrack, it reports state-of-the-art mean accuracy.
Assaf Singer, Maroun Khriesh, Noam Rotstein et al.
In briefA Better Critic Matters for Distribution Matching Distillation
SummaryThis paper studies what the online critic in Distribution Matching Distillation (DMD) needs in capacity and information to give a reliable training signal. Experiments on ImageNet across seven SiT model sizes show that enlarging the teacher while keeping the critic small can degrade students and cause collapse. The authors propose joint scaling, which grows the critic with the teacher, and a step-aware critic (STAC) that conditions on the student's generation step. Together they improve few-step distillation for ImageNet, SDXL, FLUX.2, and Wan2.1 generation.
Zongze Wu (1), Xianbang Wang (2), Zhengyang Geng (1) et al.
In briefExact information ledger for instance-dependent sequential limits
SummaryThis paper proves instance-dependent limits for bandit regret, sequential tests, best-arm identification and Markov decision processes through one shared chain of reasoning. The authors write this chain as an exact information ledger, showing that data processing loses exactly a conditional divergence. A one-parameter family of martingale identities extends the decomposition to every tilt between two instances. The ledger is applied to seven published sources to locate where standard algorithms differ from the classical limits.
In briefDecisionRank: On-Policy Distillation for Decision-Model Rerankers
SummaryThis paper introduces DecisionRank, a decision model that predicts document scores directly for reranking, avoiding autoregressive decoding. It is trained with item-space on-policy distillation, where a generative LLM teacher gives feedback on the ranked list built from the student's scores. Combined with offline ranking supervision, the authors report state-of-the-art results on TREC DL, BEIR, and BRIGHT, with substantially faster inference than the generative listwise counterpart.
In briefTransferability of ICU mortality prediction inputs when measurement conditions change
SummaryThis study predicts in-hospital mortality at 48 hours for ICU patients using MIMIC-III, MIMIC-IV, and eICU data. It compares models built from latest-value snapshots, longitudinal physiological summaries, and measurement frequency and timing features. Physiological trajectories transferred more consistently to external data, while observation-pattern features were sensitive to the timing and availability of measurements. The authors argue that clinical prediction models should be evaluated under the measurement conditions that supply their inputs, not only in new patient populations.
In briefGradient Descent on Two ReLU Neurons: Global Landscape and Bifurcation Dynamics
SummaryThis paper studies how gradient descent learns in a two-neuron ReLU teacher-student model with isotropic Gaussian inputs, trained by population gradient flow from small initialization. The authors analyze the population squared-loss landscape and show that every critical point with both student weights nonzero is either a global minimum or a saddle. They also track the gradient-flow trajectory, finding that student neurons first collapse toward a bisector saddle, then escape and specialize to the teacher neurons.
In briefFaster privacy loss calculation for random allocation using multiscale FFT
SummaryThis paper studies privacy loss accounting for the random allocation sampling scheme, in which each user's data is used in k randomly chosen steps out of t. The authors introduce a multiscale FFT algorithm that reduces the running time's dependence on the accuracy parameter α from O(α^-2) to O(α^-1 log(1/α)), while keeping the accountant's guarantees. They also use a mean-preserving projection to shrink rounding errors. Experiments show speedups of several orders of magnitude at high accuracy and improved privacy bounds for DP-SGD with PREAMBLE.
In briefACCT: Asymmetric Cross-LLM Context Transfer, a method for reusing context computation across models of different sizes
SummaryThis paper proposes ACCT, a method for transferring KV cache between language models of different scales within the same model family. It converts the KV cache from a source model into sparse KV states that a target model can use directly, so the target does not need to reprocess the original context. Across five general-capability benchmarks and six context-dependent tasks, ACCT matches or exceeds the target model's original performance while cutting the KV cache size by 75%. At about 19K tokens of context, its KV conversion runs 6.5 to 21.2 times faster than target context prefill.
In briefSubquadratic attention using a hierarchical tree indexer for long contexts
SummaryThis paper proposes Zyphra Hierarchical Attention (ZHA), a sparse attention method that replaces the quadratic cost of standard attention with a tree-structured indexer. The method builds a tree of recursively compressed key summaries over the KV cache, which queries traverse top-down to retrieve a sparse set of keys. The authors report that ZHA can be applied to existing pretrained models with a short KL distillation stage and maintains performance on long-context and reasoning tasks.
WhyThe authors propose a hierarchical tree indexer for sub-quadratic sparse attention, distilled into pretrained models and evaluated on retrieval, reasoning, and coding tasks.
Leon Lufkin, Cutter Dawes, Alexander Speicher et al.
In briefSATFRONT: Machine learning method to detect weather fronts from satellite images
SummaryThis paper introduces SATFRONT, a machine learning framework that detects cold, warm, and occluded weather fronts over Europe using geostationary satellite observations. The model is trained on frontal analyses from five European national meteorological services and evaluated against UK Met Office analyses for 2024. Using satellite data alone, it achieves a Critical Success Index of 0.428 at a 150 km tolerance, which rises to 0.591 when ERA5 reanalysis data is added.
Ioannis Dravilas, Miruna Stoicescu, Roope Tervo et al.
In briefRegret-Optimal Wasserstein-Robust Linear Regression: examines how robust regression methods perform in excess risk (regret) under distribution shift, comparing several approaches.
SummaryThe paper studies distributionally robust optimization (DRO) for multivariate least-squares regression when the source and target distributions differ by an adversarial 2-Wasserstein perturbation of size ε. It characterizes the small-ε complexity with a problem-specific condition number κ, showing that simple methods including DRO attain the instance-optimal rate O(κε²). The authors identify settings where DRO is provably suboptimal by dimension-dependent factors, and propose minimax regret optimization (MRO) and spectral-norm regularization (SNR), which attain the minimax rate. Exact convex reformulation, an accelerated algorithm, high-probability guarantees, and numerical experiments are provided.
WhyThe authors show DRO can be suboptimal in regret and that MRO and SNR attain the minimax rate, with a convex reformulation and an accelerated proximal algorithm.
In briefEnforcing exact discrete physics in a neural network for the Burgers equation via a projection layer.
SummaryThis paper extends hard constraint projection (HCP) from linear PDEs to the nonlinear one-dimensional Burgers' equation. The equation is discretised with a finite difference stencil, and a physics-informed neural network (PINN) predicts velocity, which a final unlearnable layer projects onto the hyperplane of exact solutions to the discrete equation. The authors report that the HCP-PINN keeps reducing its loss after the PINN loss plateaus, reaches a data error about an order of magnitude lower on an unseen uniform grid, and is less sensitive to the choice and number of physics collocation points.
Miranda J S Horne, Peter K Jimack, Amirul Khan et al.
In briefStrategic Tanking in Stable Matching Markets: deliberately lowering one's appeal to others to improve matches
SummaryThe paper introduces strategic tanking, in which an agent intentionally makes itself less attractive to selected potential partners to improve its eventual match, without misreporting its own preferences. Cutoff tanking has exactly the same effect on the stable set as truncating one's preference list at that threshold. Arbitrary tanking can go further, producing equilibria outside the sincere stable set under proposer-deferred acceptance.
In briefVC-1K: A Diverse Video Test Set for Real-Time Video Calls
SummaryThis paper introduces VC-1K, an open-source test set of about 1,000 video sequences for evaluating video processing and encoding in real-time communication. The videos are user-generated video call content recorded on consumer mobile devices under varied conditions, mostly talking-head footage with a small set of rear-camera videos without faces, and each includes synchronized audio. The authors evaluate OpenH264, SVT-AV1, and libaom under constant bitrate encoding, along with RealESRGAN and GFPGAN super-resolution methods.
In briefHyperbolic Inverse Reinforcement Learning: Modeling Rewards with Hyperbolic Distance to Infer Behavioral Goals
SummaryInverse reinforcement learning infers the reward function behind observed behavior. This paper introduces Hyperbolic IRL, which parameterizes reward as a function of hyperbolic distance between a feature vector and a learned parameter vector, along with a Euclidean distance variant. The method performs comparably to state-of-the-art models on established benchmarks under deterministic and stochastic policies, and excels in the low data limit and dense environments. Applied to human navigation among multiple museum exhibits, it reveals reward structure that depends on non-separable conjunctions between exhibit-distance features, not just individual distances.
Iulia Rusu, Anoop Praturu, Sergei Gepshtein et al.
In briefEffects of plasticity loss on reinforcement learning in noisy environments
SummaryThis paper studies how loss of plasticity in neural networks affects deep reinforcement learning when environments are stochastic. The authors train SAC agents on three DeepMind Control tasks with controlled levels of observation and action noise. Plasticity-preserving methods improve robustness on the harder tasks, with the advantage growing as observation noise increases.
In briefMemU-Bench: A Benchmark for GUI Agents Handling Vague, History-Dependent Long-Horizon Tasks
SummaryThe paper introduces MemU-Bench, a mobile benchmark of 118 vague tasks built on 29 long-horizon user interaction trajectories, where agents must recall missing context from history. Task outcomes are verified against application databases, and metrics such as Interaction Efficiency and SPL are used to distinguish genuine recall from exhaustive search. The authors also propose AppMem, a training-free framework that builds a structured, app-centric memory bank, raising GPT-5.5's SPL from 15.7% to 66.9%.
