In briefCan Natural Language Autoencoder Explanations Be Trusted for Oversight?
SummaryThe paper studies Natural Language Autoencoders, which use an Activation Verbaliser to explain activation vectors in English and an Activation Reconstructor to map explanations back to vectors. The authors argue that training and oversight are different referential games, so explanations may identify an activation without discriminating its true account from closely matched false ones. On one released checkpoint, true accounts beat other activations' accounts in all comparisons but beat single-claim false rivals only 60.2% of the time.
In briefDetecting Code-Switching in Speech Using Language-Selective Dimensions of Pretrained Encoders
SummaryCode-switching, mixing two languages within one utterance, is common in multilingual speech but rare in pretraining data. The paper proposes a lightweight method that detects code-switching directly from speech by finding language-selective hidden dimensions in pretrained speech encoders using monolingual recordings. On CS-FLEURS across 14 language pairs, the Whisper-small encoder reaches a mean F1 of 0.798, above a CNN at 0.776 and a HuBERT linear baseline at 0.751.
Jie Chi, Zakaria Aldeneh, Masha Fedzechkina et al.
In briefLearning Source Structure: Bayesian Prediction with Transformers
SummaryThis paper studies transformers trained on sequences whose source parameters are redrawn each time, showing they can learn only what the sequences share: the source family and its prior. The authors define an informed Bayes predictor, computable from the family alone, as a quantitative reference. Experiments on Markov sources with up to 64 symbols show it predicts how trained transformers behave in several tests, and small models match it when the shared structure is visible in the tokens.
WhyUses informed Bayes as a falsifiable reference, predicting four trained-transformer behaviors on Markov sources up to 64 symbols, with quantitative matches.
In briefAIRRI: A human-readable, machine-resistant reading interface for assignments
SummaryThis paper presents AIRRI, a framework that generates educational materials resistant to automated text extraction, along with a higher-education assignment dataset for evaluation. It combines image transformations such as spacing, opacity changes, and noise with decision-based adversarial attacks (RayS, SurFree, ADBA). Tested against Tesseract, EasyOCR, and GOT-OCR2, AIRRI substantially increased character and word error rates, with ADBA producing the largest degradation. The authors suggest this could make AI-assisted assignment completion harder.
In briefCollusion-free teaching dimension for planar graph neighborhoods
SummaryThis paper studies non-clashing teaching dimension (NCTD), a measure of how many examples a teacher needs under a collusion-avoidance rule, and compares it with VC dimension, an open question in learning theory. For closed neighborhoods in planar graphs, the authors prove a tight upper bound of two on NCTD, supporting the conjectured inequality for this class. For the positive-only variant NCTD+, they give a tight bound of four. They also show that deciding whether NCTD+ or NCTD is at most one is NP-complete.
In briefSteertuning: fine-tuning robot policies from only a few human videos
SummaryThis paper presents Steertuning, a system that lets robots learn new tasks in new environments from 5 human demonstrations. It pretrains a 400M-parameter diffusion policy on about 100 hours of egocentric human interaction data, then fine-tunes it for about 2 minutes and steers sampling using the demonstration whose observation is closest to the deployment scene. The authors report a 71% success rate across 12 tasks in 12 new environments.
In briefTraining-free method extracting token-level rewards from a reward model
SummaryThe paper proposes a training-free method that extracts token-level reward signals directly from an existing deep reward model. It estimates each token's influence using a first-order Taylor approximation and gradients, requiring only one forward and backward pass. Experiments on reasoning benchmarks report gains over outcome rewards, including +4.9% on MathVista and +11.5% on AIME24.
In briefUniW(A)M: A Unified Mixture-of-Experts Model for Predicting and Acting in Robot Worlds
SummaryThis paper addresses world models, which predict future states, and world-action models, which generate robot actions, that are usually trained separately. The authors propose UniW(A)M, a diffusion-Transformer mixture-of-experts model that unifies action-conditioned world modeling with several world-action variants. A plug-in router, PAIR, selects a suitable action interface for each execution chunk while the backbone stays frozen. Reported results include 88.3% average real-world success versus 76.7% for the strongest baseline.
In briefCryoTracer: a synthetic-to-real method for tracing whole filaments in cryo-EM images
SummaryThis paper addresses filament tracing in cryo-electron microscopy, where standard particle detection with bounding boxes fits filaments poorly. The authors propose CryoTracer, which pretrains on procedurally generated micrographs, predicts filaments as ordered centerline polylines with a transformer, and fine-tunes on a few annotated experimental images. It optimizes an orientation-invariant discrete Fréchet objective. On established and newly annotated datasets, the authors report improved instance-level filament recovery over existing particle-detection approaches.
Rados{\l}aw Kuczba\'nski, Wojciech Koz{\l}owski, Mateusz Grzesiuk et al.
In briefRefresh Timing Matters: Learning-Rate-Dependent Schedules in Kronecker-Preconditioned Language Model Training
SummaryThis paper studies periodic refreshes of cached curvature eigendecompositions in Kronecker-preconditioned optimizers, showing that at a fixed number of refreshes, their timing changes held-out loss at equal compute. Training GPT-style models with the Mousse optimizer on FineWeb, the two placements swapped rank as the learning rate rose, so a single-learning-rate comparison can give either ranking. The authors recommend reporting refresh schedules in full and comparing them across the learning rates that will be used.
In briefGeoReader: Teaching vision-language models to use geometry they can read
SummaryThis paper addresses the difficulty vision-language models have with metric and multi-view spatial relationships implicit in RGB images. It introduces GeoReader, which trains lightweight Geometric Readers to produce continuous tokens supervised by geometry descriptions decoded by a frozen VLM, then fine-tunes the VLM on spatial QA. The authors report 94.3 on MindCube-Tiny with Qwen3-VL-4B and an average of 72.5 across six additional spatial benchmarks with Qwen3.5-9B.
In briefEstimating Cryo-EM image orientations for symmetric molecules via semidefinite programming
SummaryThis paper extends the non-unique games framework of Bandeira et al. (2020) to symmetric structures, recasting orientation estimation under symmetry elements as a convex program. It uses a semidefinite programming relaxation solved with the alternating direction method of multipliers. Unlike methods tailored to specific symmetry groups, it offers one computational framework for all point group symmetries. Tests on simulated and experimental data show it can produce high-quality ab initio models from only a few experimental class averages.
In briefJust Respect the Semantics: A Semantic Image-Token Compression Method (Varest)
SummaryThis paper introduces Varest, a method for compressing image tokens in multimodal models while preserving semantics. Before the language model, it merges tokens at variable resolution, guided by object and text localizers that keep finer detail at boundaries and text. Learned [REL] registers then prune visual tokens after early decoder processing. Across three backbones and 13 benchmarks, it reportedly outperforms 15 baselines, retaining 86.10-92.33% of full-token performance with 5.56% of the layer-averaged image-token count.
In briefRSIgen: A Mutual Self-Improvement Framework for Agentic Image Generation
SummaryThis paper proposes RSIgen, a framework for agentic visual generation in which an agent and a generator role of one multimodal model improve each other. Dynamic Responsibility Routing decides which role should learn, and the agent's action supervision is refreshed after generator updates before the next curriculum is built. On BAGEL, Qwen-Image and Wan-Image-7B, it improves AgentGen-Bench Full Overall9 by 3.2 to 3.3 points over a three-cycle SearchGen adaptation. In blind evaluation on 200 real-user requests, it reaches 70.0% preference for task completion and 64.9% for visual quality.
In briefRadarGSOcc: Doppler-aware 3D semantic occupancy prediction from 4D radar tensors
SummaryThis paper proposes RadarGSOcc, a framework that processes 4D radar tensors directly by extracting geometric and semantic features, including Doppler information, for 3D semantic occupancy prediction in autonomous driving. Its key component is a deformable attention-based geometric refinement module. The authors also introduce K-Radar 3DSemOcc, a large-scale benchmark built on the K-Radar dataset with refined LiDAR-based annotations. The method outperforms RadarOcc in weighted mIoU while using fewer parameters and less computation.
Christofel Rio Goenawan, Dong-Hee Paek, Seung-Hyun Kong
In briefKnowing the User Is Not Knowing What Was Agreed: Grounding Awareness for Personal Agents
SummaryThis paper argues that personal-agent memory research has focused on what users know, prefer, and want, but remembering a user does not mean knowing what the user and agent have jointly established. The authors introduce Grounding Awareness and a controlled benchmark that tests whether agents distinguish accepted, unacknowledged, revised, and scoped arrangements. They find that strong memory performance does not guarantee this awareness, and that persistent memory can cause unacknowledged proposals to be treated as user instructions. They also propose Grounding Layer, a training-free extension that tracks grounding status and checks consequential actions.
In briefProAS: A progress-aware method for fine-grained credit assignment in multi-turn AI agent training
SummaryThe paper notes that critic-free reinforcement learning methods such as GRPO assign the same trajectory-level advantage to every action, which can reinforce poor actions. The authors propose ProAS, a plug-and-play framework that uses a dynamically evolving Agentic Reference Graph with masking and asymmetric penalty redistribution to assign credit more finely. Experiments on ALFWorld and WebShop with Qwen2.5 models report improved success rates and faster training convergence.
WhyProposes progress-aware advantage shaping with a dynamic reference graph, validated against group baselines on ALFWorld and WebShop with convergence speedups.
In briefModel-based online decisions via generative trajectory planning
SummaryThis paper proposes Generative Trajectory Planning (GTP), a model-based framework for online decision-making. It uses a generative model as a proposal distribution over action sequences, sampling candidates from a diffusion prior and refining them with learned dynamics, reward, and value models. The authors report that GTP matches or exceeds model-free, diffusion-policy, and model-based baselines on continuous control and long-horizon manipulation tasks, and stays stable where strong baselines degrade.
In briefShared Attention-Head Interventions Reduce Harmful Context Copying in Language Models
SummaryThis paper studies a robustness issue in which large language models reuse irrelevant tokens from earlier prompt context when answering later queries, a behavior called contextual entrainment. The authors ablate attention heads that reduce this effect and test how well the recovered heads transfer across exact relations, semantic families, and mixed relation sets. They find that many heads are rediscovered across relation subtypes, and that intervening on these shared heads reduces harmful copying while preserving ordinary relation completion and simple in-context abilities.
In briefSMI: Ontology-based semantic IDs that build hierarchy-aware representations of medical concepts
SummaryThis arXiv paper (cs.CL, cs.AI) proposes SMI, a training-free framework. It assigns each diagnosis, procedure, or medication a coarse-to-fine semantic ID along an expert-defined medical ontology and aggregates frozen biomedical language model embeddings across those levels. The authors report improvements on clinical prediction tasks using MIMIC-IV and eICU, including mortality, readmission, and drug recommendation.
In briefPerceiving Beyond Visibility: Incompleteness-Aware Scene Understanding for Robots
SummaryThis paper presents a research vision for robot perception that accounts for incomplete observations caused by occlusion, unfamiliar objects, and degraded sensors. It organizes the problem around three challenges: incomplete evidence, incomplete knowledge, and incomplete sensor reliability. The author describes research directions including amodal perception, semantic occupancy forecasting, open-world scene understanding, label-efficient learning, and uncertainty-guided multi-modal fusion. The paper also outlines future work on belief-centric 4D scene representations and on active sensing and safe adaptation to reduce uncertainty through interaction.
In briefCloned Known Voices and the Limits of Personality Transfer to Virtual Characters
SummaryThis paper examines whether personality impressions transfer when a virtual character speaks with the cloned voice of a familiar person. Two studies were conducted: one at a university using a familiar lecturer's voice, and one online using voices of social media creators. The authors found limited personality transfer, with selective effects that did not generalize across traits or contexts. The results suggest cloned voices do not straightforwardly lead people to perceive artificial characters as capable social actors.
Johanna Magdalena Kuch, Maud van Lier, Daksitha Senel Withanage Don et al.
In briefPlanPath: an exploration-guided framework for answering questions about a patient's pathology slides
SummaryThis paper presents PlanPath, an agentic framework for answering visual questions about a patient's full set of pathology slides. It converts each question into an exploration plan, selects the slides likely to hold the needed evidence, and works from tissue structure down to individual cells, then combines the counts and proportions across slides into a traceable answer. The authors built renal and gastric cancer datasets averaging 15 slides per patient, and report accuracy gains of 9.4% and 12.1% over the strongest baseline.
In briefVirtual-Alloy Physics-Aware Semantic System: extrapolative prediction for perovskite solar cells
SummaryExisting machine-learning models struggle to predict perovskite solar cells made from new materials. This paper proposes V-PaSS, which interpolates relaxed crystal graphs of 23 pure ABX3 endpoint compounds via Vegard's law and couples layer sequence, interfacial band offsets, and a scientific language model. Tested on devices from unseen laboratories, its untuned default configuration achieves the lowest mean absolute error among fourteen models, at 2.861%. Its advantage is largest where test devices differ most from training data.
In briefBridgeTwist: Training-Free Conditional Sampling by Twisting Schrödinger Bridges
SummaryThis paper proposes BridgeTwist, a training-free conditional sampler for pretrained Schrödinger bridge generative models. It uses a tempered twisted sequential Monte Carlo scheme, with a closed-form twist built from the velocity and score fields of the pretrained model, and a time-varying exponent that downweights the twist near the source. A bridge-shaped inference-time noise schedule and telescoping importance weights yield an asymptotically exact posterior sampler. The authors report that it outperforms training-free baselines on class-conditional and text-prompted image generation and is competitive with specialized solvers on inpainting, deblurring, and super-resolution.
In briefBayesian test-time inference of task-aligned similarity from weak feedback
SummaryThis paper introduces BASIS, a training-free Bayesian method that uses weak test-time feedback (scalar scores, pairwise preferences, or verifier outputs) to infer which similarity criterion in a supplied task-specific library governs a decision. It maintains a joint posterior over targets and criteria and selects probes by expected information gain. Across concept recovery, reasoning exemplar selection, and response refinement, it outperforms tuned static mixtures. BASIS only selects among the criteria it is given and does not discover new ones.
In briefTexture interpolation by matching similar local appearances instead of pixels at the same position
SummaryThis paper studies how the choice of which local elements to pair between two textures affects the quality of blended in-between textures. Instead of mixing features at the same spatial locations, the authors match textons (local appearance vectors) by appearance while keeping the encoder, representation, and decoder fixed. Across tested texture pairs, appearance-aware matching improved the distributional quality of midpoints over positional blending, with soft matching giving the strongest results. The same approach also improved a second pretrained interpolator without retraining.
In briefKwai Summary Attention V2 Technical Report: A Method for More Flexible KV Cache Compression and Cheaper Attention
SummaryThis paper presents Kwai Summary Attention V2 (KSA V2), an improved version of the summary-token architecture KSA. The original KSA compresses distant context into a learnable summary token per chunk and interleaves these layers with full attention layers. The authors identify two bottlenecks: all layers must use the same compression chunk size, limiting flexible KV cache compression, and full attention layers still incur quadratic prefilling cost. KSA V2 is reported to allow more flexible KV memory compression and to reduce computation in full attention layers, based on the finding that summary tokens can serve both as KV memory and as a sparse indexer.
In briefAMemOps: A Benchmark for Updating Memory in Long Audio Streams
SummaryThis paper introduces AMemOps (Audio Memory Operations), a benchmark that tests whether audio memory systems can retrieve the latest attribute of a recurring sound source, keep unchanged attributes, and reject sources that are absent in long audio streams. Paired controls vary recordings, training identities, and earlier events to examine how each affects updating. Training on other recordings of the test instruments raised a reference delta-rule memory's update advantage over matched attention by 13.4 percentage points.
In briefDoes pretraining actually teach chemical language models chemistry?
SummaryThis paper systematically studies how pretraining affects chemical language models (CLMs), evaluating them on 64 molecular property prediction datasets. Downstream gains from pretraining do not scale with size, and performance saturates below gradient-boosting baselines. Only supervised pretraining on traditional physico-chemical descriptors consistently helps, while some pretraining methods perform worse than a randomly initialized encoder. The authors conclude that CLMs mainly learn descriptor-like properties and low-level token statistics, with little chemistry beyond classical molecular features.
In briefSPELUNKA: A Framework for Finding Hidden Associations in Vision-Language Models
SummaryVision-language models often learn spurious or harmful associations from their training data, and methods for finding unexpected associations in retrieval tasks are less studied. This paper presents SPELUNKA, an automated inference-time framework that needs no predefined labeled subgroups, human supervision, or fine-tuning. It generates captions for related images, extracts entities, and uses a frequency-based rule to flag attributes over- or under-represented in top-ranked outputs relative to the full dataset. Across four datasets, the authors report it recovers synthetic and natural spurious correlations, reaches an average AUROC of 0.93 in identifying images strongly associated with a query, and outperforms baselines.
Cassandra Parent, Isha Agarwal, Kumail Alhamoud et al.
In briefSmall recursive message-passing networks that solve algorithmic reasoning tasks
SummaryThis paper applies deep supervision, truncated recursion, and test-time procedures from recursive transformers to message passing neural networks. The authors report about 99% test accuracy on Sudoku-Extreme with a 16k-parameter model. They also show zero-shot transfer to larger N-queens and Boolean satisfiability instances, outperforming popular neural baselines with far fewer parameters.
Tommaso Salvatori, Matteo Forasassi, Alexis Toumi et al.
In briefST-GRU PredRNN-V2: A simpler GRU-based recurrent cell for video prediction
SummaryThis paper replaces the spatiotemporal LSTM cells in PredRNN-V2 with a simpler spatiotemporal GRU cell for video prediction. The new cell uses two gates instead of three and merges the temporal memory into the hidden state, reducing parameters by about 5%. On the Moving MNIST benchmark it reports MSE 55.1, SSIM 0.883, LPIPS 0.065 and PSNR 19.53dB. It does not beat the original PredRNN-V2 on MSE, SSIM or PSNR, but achieves lower LPIPS and performs on par with PredRNN, and the authors suggest ST-GRUs are a promising building block.
Augusta van Haren, Hanna Hoogen, Luca Pattavina et al.
In briefMultiAlloc: Budget-Aware Precision Allocation for KV Cache Quantization
SummaryKey-Value (KV) caches use a large share of memory when large language models run inference, especially for long contexts. This paper introduces MultiAlloc, a framework that distributes a fixed quantization budget between layer-wise precision enhancement and token-wise preservation. It uses lightweight offline profiling of layer sensitivity and token importance, with no online allocation search. Experiments on reasoning and long-context benchmarks across Mistral, Llama, DeepSeek-Qwen, and QwQ show gains over layer-wise, token-wise, and recent multi-granularity methods under matched memory budgets.
Yangchengyu Zhou, Zhiping Wu, Runchang Jiang et al.
In briefTo Align or Not To Align: Check Your COMPASS Before You Train
SummaryThe paper proposes COMPASS, a diagnostic score that estimates whether two pretrained models will be compatible before they are aligned or merged. It combines a pair-specific k-NN overlap between the encoders with a ceiling set by their training paradigms. Tested on 72 cross-modal alignment experiments and 392 task-arithmetic merges, COMPASS estimates correlated strongly with downstream success.
WhyProposes a k-NN-overlap compatibility diagnostic, validated on 72 cross-modal alignments and 392 merges with strong correlation to outcomes.
Fabian Gr\"oger, Petar Damjanovi\'c, Maria Brbi\'c
In briefGrid or Nodes? Testing Object Memory in Robot World Models When Views Are Blocked
SummaryA robot's world model must remember objects that leave the camera's view. The paper compares FloWM, which stores objects in a spatial feature grid with five built-in velocities, against a node memory that stores each object's position and velocity as numbers, using a Kalman tracker fed with oracle data. How the grid is read changes its score. On diagonal motion, which falls outside FloWM's five velocities, FloWM finds only 11% of blocks even after fine-tuning, while the tracker keeps 93%.
In briefDTScribe: Generating digital twin reporting models with local LLMs
SummaryThis paper presents DTScribe, an approach that uses local large language models to automatically turn digital twin documents into a conceptual architecture visualization and a reporting page. It extracts details aligned with a 21-characteristic Digital Twin Description Framework and generates an Ontological Modeling Language model. Refinement loops and ontological reasoning are used to improve quality, and the authors share practical lessons such as using medium-sized local models.
In briefGeometric Calibration of Supervised Fine-Tuning: A Training-Free Post-Hoc Correction
SummaryThis paper proposes SPACE, a lightweight post-hoc method that, after supervised fine-tuning (SFT), selectively reverts SFT update components within the principal subspace of the pre-SFT weights while preserving those in the orthogonal complement. It is data-free and training-free, requiring only the pre- and post-SFT checkpoints. Evaluated on Qwen, Llama-3.1 and Olmo-3 models from 2B to 14B, SPACE reduces out-of-domain capability regression while improving in-domain performance.
In briefBreakdown of Conversational Involvement: Joint Multi-Modal Analysis of Conversation Floors and F-formation Detection
SummaryThis paper studies conversational involvement in free-standing conversations by comparing two operationalizations: F-formations and conversation floors. The authors estimate floors from speaking status and analyze F-formation detections using different feature sets. They then jointly combine spatial features with floor patterns. The paper concludes that F-formations alone are insufficient to capture single conversations and that a more detailed theory and framework is needed.
In briefFalsification-proof mechanisms for non-market resource allocation
SummaryAllocating resources based on manipulable characteristics leads to falsification and inefficient allocation. The paper characterizes optimal falsification-proof rules across a broad range of falsification technologies. It also provides an honesty condition under which these rules solve the unconstrained problem of a planner maximizing a weighted sum of allocative surplus and agents' welfare.
In briefMajorTOM:Elliot Pretrain, a multi-sensor, multi-temporal Earth observation pretraining dataset
SummaryThis paper introduces MajorTOM:Elliot Pretrain, a dataset of 279,166 10 km tiles for pretraining Earth observation foundation models. Each tile combines Sentinel-2, Landsat-8/9, Sentinel-1, elevation, land cover, cloud masks, OpenStreetMap entities and ERA5 weather on a shared grid, organized into monotemporal, monthly and burst temporal regimes and released as 6.3 TB of Cloud Optimized GeoTIFFs. Short pretraining experiments with OlmoEarth-Base test design choices such as data subset selection, adding Landsat and Sentinel-1 inputs, using terrain and land cover as prediction targets, and rule-based captions.
Carmen Oliver Huidobro, Mikolaj Czerkawski, Kevin Monsalvez-Pozo et al.
In briefBest achievable accuracy for estimating treatment effects at one point, and how covariate design affects it
SummaryThis paper studies how the smoothness of the covariate design affects minimax estimation of conditional average treatment effects at a fixed covariate point. For design densities bounded above and away from zero, the authors derive upper and lower risk bounds whose convergence-rate exponents differ by an arbitrarily small gap. Under uniform design with low nuisance smoothness, they obtain matching bounds at interior points. The results show that rough covariate designs can yield slower rates than uniform designs under the same smoothness assumptions.
In briefDeterministic Prediction as Global Velocity: Flow Matching for Probabilistic Forecasting
SummaryThe paper addresses uncertainty quantification in time series forecasting by proposing DV-Flow, a unified flow-matching framework. It treats the deterministic predictor as a parallel velocity component rather than the endpoint of the transport path, which enables stable end-to-end training and subsumes several existing formulations as special cases. Experiments on six real-world datasets show consistent improvements in probabilistic forecasting over strong baselines.
In briefTwo Ways an Agent Prioritises Bodily Needs: Precision Versus Preference in Active Inference
SummaryThe paper compares two ways an embodied agent can respond to an urgent bodily need in active inference: increasing the precision of information about the need, or increasing preference for outcomes that relieve it. Experiments in a foraging gridworld with three lethal needs examine effects on survival, perception and the learned body model. Precision improves survival, while preference reweighting keeps bodily states further from lethal boundaries and changes the states the agent learns from.
St John Grimbly, Nicolas Kuske, Evert A. Boonstra et al.
In briefKnowledge graphs as files: agents that write their own graph-search tools
SummaryThe paper proposes Direct Knowledge Graph Interaction (DKGI), a framework where language model agents write Python code to create graph access tools directly over complete node and edge files. On benchmarks across three domains, the training-free method beats the strongest published baseline by at least 6.7 Hit@1 points on every graph. The authors argue that letting tool creation and tool use co-evolve during graph search is valuable.
WhyAgents write graph-access tools over full files and co-evolve them with use; controlled ablations beat strong baselines, with impact mainly within a subfield.
In briefVideoJev: A Single Model for Video Understanding and Interaction
SummaryVideoJev is a family of models that treat video understanding and interaction as structured decision tasks. It adapts pretrained multimodal large language models with a shared head that scores temporal intervals and answer options against video evidence, and reuses visual computation through cached and joint inference. On three temporal grounding benchmarks, the 4B and 35B-A3B versions improve mean mIoU over their backbones by 6.7 and 10.8 points, with 2.8 to 3.1 times faster inference. On four video QA benchmarks, both reach performance comparable to their backbones in Thinking mode, with reported speedups above 100 times.
In briefCan an LLM Agent Learn to Predict Its Own Failure? Outcome-Grounded Confidence Calibration
SummaryThe paper proposes calibrating an LLM agent's per-step verbalized confidence against the final outcome of a multi-step episode, rather than the correctness of the current step. Under GRPO, the rollouts for each task provide success labels for this calibration. In experiments on ALFWorld, TextWorld, and WebShop, step-confidence AUROC against episode outcome rises from near chance to 0.68 to 0.85, and online ALFWorld expected calibration error drops from 0.42 to below 0.12.
In briefDoes self-distillation improve reasoning? A re-examination of SDPO and OPSD
SummaryThis paper re-examines two self-distillation methods, SDPO and OPSD, on code, science, tool use, and math reasoning tasks. The authors find that the reported gains do not consistently generalize to unseen problems and are largely explained by overfitting or transfer of existing reasoning behavior rather than effective use of privileged information. Experiments on Qwen3 models (1.7B-14B) and OLMo-3-7B suggest leveraging privileged information for reasoning remains an open challenge.
Jason Ken Adhinarta, Raina Wu, Ashutosh Sharma et al.
In briefLegal-R1: An agentic retrieval framework for evidence-based legal reasoning
SummaryThis paper presents Legal-R1, an agentic retrieval framework designed to ground legal question answers in valid, citable evidence rather than one-shot semantic matching. It splits legal reasoning across three agents: a Lawyer agent that decides when to retrieve and what evidence to seek, a Legal Assistant that selects normative sources and precedent cases by authority and diversity, and a Judge agent that writes the final answer. Only the Lawyer agent is post-trained, using masked supervised fine-tuning and trajectory-aware reinforcement learning. The authors report improvements over general-purpose, legal-specific, and retrieval-augmented baselines on multiple Chinese legal benchmarks.
In briefJev-ToM: Evaluating Theory of Mind in Typed Decision Models
SummaryThis paper introduces Jev-ToM, a benchmark comparing five typed decision models with seven generative language model baselines on 11,117 English questions from nine Theory of Mind benchmarks. TypeSafe Jev achieves 73.2% source-balanced accuracy and ranks first or second across all six capability groups. The authors also examine confidence-based selection, candidate expansion, and related-question cues, and suggest integration priorities for assistant systems.
In briefAdapting a Diffusion Transformer into a one-step image decoder for extreme compression
SummaryThe paper presents PrismCodec, which turns a pretrained Diffusion Transformer into a one-step decoder driven by the compressed bitstream. The decoded representation serves both as the generation starting state and as a source of hierarchical semantic and structural guidance. A progressive adaptation strategy bridges stochastic generation and deterministic decoding, so no extra conditioning bits are needed. On Kodak, CLIC, and DIV2K, it reports BD-rate reductions of 70-75% in LPIPS and 75-81% in DISTS over StableCodec.
In briefBeyond Answer Voting: Fine-Grained Correction of LLM Reasoning via Graph Alignment
SummaryThe paper introduces DDRG, a framework that selectively corrects reasoning from a fixed set of sampled reasoning traces rather than relying only on majority voting over final answers. It builds source-grounded claim graphs, sends localized claim conflicts to an answer-blind verifier, and performs one focused repair only when a verifier-supported refutation reaches the majority answer's supporting path; otherwise it returns the self-consistency vote. Across five benchmarks, zero-shot DDRG improves macro-average accuracy over direct answering by 8.65 points for Qwen2.5-7B and 15.83 points for Llama3.1-8B.</summary>
In briefFrom Hindsight to Foresight: How LLMs Struggle to Gather Key Clinical Evidence
SummaryThis paper examines why large language models fail to acquire diagnostically critical evidence during dynamic clinical diagnosis. It traces failures through evidence identification, utilization, and diagnostic updating, measuring accuracy, evidence coverage, trajectory dynamics, probability, and information entropy. Verification and generalization experiments indicate these limitations are widespread. The authors argue the central challenge is whether models can identify, acquire, and use the right evidence, not merely whether they hold relevant clinical knowledge.
In briefLLM2R: Using Large Language Models to Reassemble 3D Parts
SummaryLLM2R treats 3D part assembly as sequence modeling, using a pretrained decoder-only large language model to predict each part's rotation and translation in a single forward pass. Learnable summary tokens appended after the last part let every part's representation aggregate information about the whole assembly at a constant token cost. The authors report the highest part accuracy and success rate on PartNet categories, strong results on PartNeXt and Breaking Bad, and advantages that grow with the number of parts, including zero-shot transfer to unseen categories.
In briefAgentStripes: A Maturity Model Grading Evidence for AI Agents
SummaryThe paper proposes AgentStripes, a maturity model for rating how well AI agents are engineered, with 0 to 7 stripes, 12 process areas covering 40 capabilities, three gates, and evidence grades E0 to E5. An exploratory demonstration on 30 public Python agent repositories, using static evidence only, capped final stripes at 2 and showed that re-judging changed results for some repositories. The parameters are expert conventions without outcome-based calibration, and the stripes have not been shown to predict reliability or safety.(Sixty-plus words)
In briefChecking whether time-limited language agents use future memories
SummaryThis paper examines whether language agents restricted to historical evidence still recall post-cutoff facts and whether that affects how they gather evidence. Across 48 historical company-comparison tasks, the authors used closed-book revenue probes to predict the better-remembered company before running paired agent trials. Qwen and DeepSeek both recalled future facts, but neither showed the predicted shift toward the better-remembered company, and a DeepSeek replication showed an inverse pattern. The authors recommend assessing memory accessibility, search allocation, and final outputs separately, alongside cutoff checks on external evidence.
In briefRethinking projector training in multimodal LLMs, with a training-free alternative
SummaryMultimodal LLMs usually train a projector on large image-caption data to align a vision encoder with a language model before joint training. The paper finds that using only 10 to 20% of the data recovers most gains, and projectors trained on different datasets can be linearly interpolated with largely preserved performance, suggesting the stage mainly performs coarse alignment. The authors propose PORTAL, a training-free initialization using only per-modality summary statistics and an optimal-transport map on a shared subspace. Without paired data, PORTAL matches the standard pipeline trained on 558k pairs across six LLM backbones, two vision encoders, and 16 benchmarks, outperforming it in 3 of 7 settings.
WhyNovel training-free projector initialization via optimal transport on summary statistics; matches standard pipeline across six LLMs, two encoders, and 16 benchmarks.
Fabian Gr\"oger, Na\"el Ouerghemi, Shuo Wen et al.
In briefPhysicsSliders: Learning physical controls to generate impact sound effects
SummaryThe paper presents PXS, a generative model for Foley contact sounds controlled by physical parameters such as hardness, size, speed and damping. It is trained on simulated contact sounds and accepts either direct parameter inputs or numerical captions. On 18 impact classes from five real-audio datasets, it follows targets with a mean grouped Spearman correlation of 0.976.
In briefACCRUE: Continual off-policy RL for robot foundation models
SummaryThis paper proposes ACCRUE, a framework that lets robot foundation models carry lightweight actor and critic networks from earlier reinforcement learning into new tasks. The critic is updated only on current-task data, while prior behavior is preserved by distilling the actor toward cached targets on a few retained episodes. The authors report, on contact-rich assembly in simulation and on real hardware, faster subsequent-task learning, near-zero forgetting, and 31.0% to 68.5% less real-robot training time.
In briefFairness Under Prediction Error: Testing Demographic Parity in Predict-then-Optimize Pipelines
SummaryThis paper studies whether a demographic parity constraint enforced on predictions still holds for true outcomes in predict-then-optimize pipelines. Using a budget-constrained allocation problem solved as a linear program, the authors find the true-outcome fairness gap grows with prediction error and then saturates, and they derive theoretical bounds. On US Census income data, enforcing parity on predictions leaves roughly one-third of the true income disparity intact, undetectable from the predictions alone.
In briefPredicting annotator disagreement on harmful content across languages using interpretable questions
SummaryThe paper proposes predicting annotator disagreement on harmful content by having an LLM answer binary questions about content characteristics, borderline severity, missing context, and local slurs. These answers feed a classifier, which outperformed direct prompting and strong text baselines across four English, German, and Turkish datasets. The authors report that questions about derogation, missing context, and decision boundaries were the strongest predictors.
Irem Ecem Yelkanat, Faeze Ghorbanpour, Alexander Fraser
In briefGeoMER: A Joint Euclidean-Hyperbolic Framework for Multimodal Emotion Recognition
SummaryThis paper introduces GeoMER, a framework that recognizes human emotions by combining information from modalities such as speech and text. It fuses cross-modal token-level interactions in Euclidean space and uses hyperbolic emotion-core routing to capture more abstract, hierarchical emotion representations. On the IEMOCAP benchmark, the method reports 83.54% weighted accuracy and 83.35% unweighted accuracy, outperforming Euclidean baselines including MemoCMT.
In briefComparing Answers to Fix Voting: Pairwise-Guided Logit Adjustment for LLM Test-Time Scaling
SummaryThis paper addresses how to aggregate multiple reasoning paths from a large language model at test time. It proposes Pairwise-Guided Logit Adjustment (PGLA), which compares confidence between paths supporting different answers, inspired by the Condorcet principle, and combines this with voting support via a KL-regularized objective. The authors report that the method can correct voting-based ranking errors and evaluated it on 10 benchmarks with 7 models ranging from 1.7B to 32B parameters.
In briefReads Correctly, Acts Wrongly: Why Fast Agent Models Fail
SummaryThe paper studies why fast, single-pass models (System One) that choose an agent's next action make mistakes. The authors pair each next-action question on a frozen benchmark with factual probes about the same state, answered independently. Across seventeen fast systems and one reasoning model, most wrong actions occur after the model answered the factual probes correctly, and more accurate systems show a larger share of such errors. These errors cluster on the same states across different systems.
In briefWhen extra objectives make finding the best option easier, depending on preferences and feedback
SummaryThis paper studies best-arm identification in multi-objective bandits, aiming to find optimal options with as few samples as possible across several criteria. For Gaussian bandits, it establishes a characteristic-time lower bound for fixed-confidence identification and compares Pareto, linear, and lexicographic preferences. The authors show that multiple objectives can sometimes require fewer samples than single-objective problems. They also propose the Order-Aware Track-and-Stop (OATS) algorithm, which is correct at any prescribed confidence level and asymptotically matches the lower bound.
WhyReveals three phenomena where multiple objectives reduce sampling needs, with a matching characteristic-time lower bound and algorithm OATS.
In briefDetecting unreliable predictions from quantized models without labels
SummaryThe paper proposes a label-free way to check quantized models by comparing each prediction's output probabilities before and after INT8 quantization using cosine similarity. This requires only a single forward pass per sample. Across ResNet-50, VGG-16-BN, Wide ResNet-50, and ViT-B/16, the agreement similarity separates lower-accuracy predictions and outperforms a confidence-only baseline on all four architectures.
In briefVaried AI Morals: Caution Bias in North American News Compared Globally
SummaryThis paper analyzes 12,924 English-language news articles about AI published since 2022, comparing storytelling across North America, China, India, and the Middle East. Using large language models, the authors extract the implicit story morals each article conveys. North American coverage is distinctively marked by cautionary narratives and less emphasis on innovation and collaboration. However, this pattern is largely explained by a broader negativity bias in North American news sentiment rather than heightened concern about AI itself.
In briefLLM Agents That Automatically Design and Optimize Post-Training Recipes
SummaryThis paper proposes Structured Recipe Optimization (SRO), a method that automatically builds and tunes a post-training recipe for large language models, covering data sources, data processing, cross-domain mixing, and training settings, under a limited experiment budget. SRO uses an LLM-based designer to construct a recipe space adapted to the target tasks and a multi-domain optimizer that uses structured trial results to guide candidate choices and Bayesian trial selection. The authors report that SRO achieved the highest average score across evaluated model and agent combinations, for example 41.1 versus 15.4 for the strongest baseline on Qwen3-1.7B.
In briefParameter-Efficient Brain Tumor Segmentation Using Shared Adapters Across MRI Modalities
SummaryDifferent MRI modalities show different parts of a brain tumor, and segmentation models combine them to outline tumor regions. The authors propose the Block-Wise Shared Adapter (BWSA), which merges compressed features from each modality encoder in a shared path and returns adapted features to each stream, plus a trainable-gated variant (TG-BWSA) that learns separate weights per modality and stage. Both were evaluated on BraTS 2020 and BraTS 2023 with four backbones. Both variants add roughly 0.85 to 3.73 percent more parameters.
Milad Rostamian, Mohammadreza Mohammadi, Ender Konukoglu et al.
In briefOrder-consistent latent geometry for reliable short-horizon planning in world models
SummaryLatent planners often score action sequences by the Euclidean distance between their endpoint and a goal latent. The paper argues that planning only requires that latent distance not reverse true transition progress, and shows this condition is the classical self-approaching property. It proves that an acute bend in the interpolated latent path can rank a transition-worse candidate above a better one, and a controlled six-branch experiment validates this ordering failure.
In briefCREATE: Testing Vision-Language Models' Creativity with Visual Association Puzzles
SummaryThis paper introduces CREATE, a benchmark of 155 visual Remote Associates Test problems that ask solvers to find a hidden connection among three seemingly unrelated images. The authors validated the benchmark with a human study of 25 participants and evaluated 11 vision-language models on it. Most models performed worse than humans, and their associative ability depended on the concepts given in text cues rather than extending reliably to visual puzzles.
In briefConfiguring RAG Systems: a study of what breaks, why, and how to fix it
SummaryThis paper presents the first large-scale empirical study of 654 real-world configuration issues collected from 10 diverse retrieval-augmented generation (RAG) systems. It categorizes the symptoms, root causes, relevant pipeline stages, and fix patterns of these issues. The authors also offer nine actionable insights for practitioners on testing, localizing, and fixing RAG configuration problems.
In briefObservation quality in DSM-assisted aerial localization: benchmark and calibrated uncertainty
SummaryThis paper studies how position observations from aerial map matching, which uses digital surface models, should be weighted for navigation. It defines five benchmark tasks and evaluates a base-anchored uncertainty calibration method that combines image and map sensitivities with a configuration subset's second moment, leaving the reported position unchanged. On two held-out query sets, the method reduces equal-location normalized position-update risk by 7.70% and 5.46% relative to a constant full-matrix baseline. The authors state that navigation benefit and real temporal accumulation remain untested.
In briefLinear last-iterate convergence of optimistic learning in harmonic games
SummaryThis paper proves that the last iterate of Optimistic Follow-The-Regularized-Leader (OFTRL) converges linearly to a Nash equilibrium in multiplayer harmonic games. The result holds under a mild game-regularity condition and for many regularizers, including entropy, Tsallis, and log-barrier. The analysis uses a dual view showing that an energy function decreases along the dual iterates, with tight quantitative bounds. The result also extends to the Extragradient FTRL variant.
WhyProves linear last-iterate convergence of optimistic FTRL to Nash equilibria in harmonic games, via a rigorous energy-function analysis.
John Lazarsfeld, Anas Barakat, Jacob Abernethy et al.
In briefDenoise X: a self-supervised U-Net that reduces noise in chest X-rays
SummaryThe paper proposes Denoise X, a lightweight self-supervised U-Net that suppresses quantum mottle and readout noise in low-dose chest radiographs. It is trained on the NIH ChestX-ray14 dataset without paired clean images. On unseen Montgomery County and Shenzhen datasets it reports PSNRs of 36.09 dB and 36.44 dB respectively.
In briefLast-iterate convergence of clipped SGD with linear learning-rate decay
SummaryThis paper studies the last-iterate convergence of clipped stochastic gradient descent under heavy-tailed noise with bounded central p-th moment (1<p≤2). For smooth convex objectives, a linearly decreasing learning rate yields a high-probability error bound. The authors extend the guarantees to nonsmooth composite objectives and smooth composite objectives with proximal updates. When a causal adversary replaces at most q gradients, a corrupted-setting bound is obtained, and lower bounds for clean and corrupted oracle models are also provided.
In briefUnbalanced W-Flow: Wasserstein drifting that stays stable under small-batch noise
SummaryThe paper proposes UW-Flow, which replaces the balanced Sinkhorn divergence in W-Flow with a debiased unbalanced version to reduce undesirable transport caused by finite minibatch fluctuations. It analyzes the velocity error between minibatch and population fields for balanced and unbalanced transport. On MNIST, CIFAR-10, and ImageNet-100, the authors report improved generation under matched training budgets, with the largest gains in unconditional generation.
In briefUnderstanding circulant permutation to extend the CMinHash estimator
SummaryCMinHash uses circulant permutations to design MinHash algorithms aimed at lower variance and reduced storage. Its original variance analysis is hard to extend to other MinHash estimators. This paper presents a new approach that yields explicit, easily comparable approximate variance expressions, and applies it to analyze the Minner estimator. Numerical experiments show that the derived variance expression matches the empirical mean square error of the estimates.
In briefInput-conditioned search for efficient test-time scaling of recursive reasoners
SummaryRecursive reasoners apply a learned update rule repeatedly to latent states, scaling inference by adding more updates or perturbation-generated trajectories. These methods can converge to incorrect attractors. The paper proposes Input-Conditioned Search (ICS), which searches over input conditions to induce alternative update maps from a pretrained reasoner without extra training. Across six reasoning tasks, ICS achieves the highest mean accuracy with up to two orders of magnitude fewer FLOPs than the baselines, and theory and numerical checks indicate it can locally eliminate incorrect attracting fixed points.
In briefObject pointers from a segmentation model as transferable tumour representations
SummaryThis paper studies object pointers from Segment Anything Model 2 (SAM2), compact object-level representations produced during segmentation, as features for downstream tumour prediction. In experiments on two PET/CT tumour prediction tasks, pointers learned from segmentation supervision alone already encoded tumour-discriminative information. The authors also introduce task-guided object pointers, which add an auxiliary objective so the pointer space is shaped for both segmentation and the downstream task. These improve cancer-type classification over encoder-based transfer and radiomics, while segmentation-only pointers perform best on recurrence prediction.
In briefEmpathy Is Not the Scapegoat for Factuality and Reliability Declines
SummaryThis paper examines whether making language models warmer and more empathetic reduces their factuality and reliability. The authors compare prompting, supervised fine-tuning, and reinforcement learning for inducing warm behavior, and evaluate capability preservation on an 11-dataset benchmark suite and EMOFACT under emotional contexts. They find that warmth effects depend on the task and model, and that task-aligned adaptive tone can improve empathy and targeted reliability, though some capability trade-offs remain.
In briefOn Symmetry for Characterizing Agent Heterogeneity in Multi-Agent Systems
SummaryThis paper studies agent heterogeneity in finite-horizon partially observable stochastic games (POSGs) by defining policy-evaluation symmetries, transformations that commute with complete policy evaluation. Agent orbits under these symmetries define structural agent types. The authors prove that every componentwise environment automorphism is such a symmetry, but the converse fails. Applied to cooperative Dec-POMDPs, the framework shows that agents sharing a common reward can still belong to different structural types.
In briefMeta-Video: Efficient Standardized Evaluation for Video Understanding
SummaryThis paper argues that video understanding benchmark scores depend heavily on evaluation choices such as frame sampling and input formatting, which are often left at framework defaults. The authors build Meta-Video, a balanced 1,000-question benchmark over 791 videos that largely preserves model rankings while running faster. They also propose a standardized evaluation protocol and release reproducible reference scores for five models across seven benchmarks.
Soumya Shamarao Jahagirdar, Anna Kukleva, Vishaal Udandarao et al.
In briefEuclidean distance-based flow matching: a more robust training objective for generative models on contaminated data
SummaryThis paper examines Flow Matching, a generative modeling technique whose standard squared-Euclidean regression objective is sensitive to contamination. The authors replace it with the plain Euclidean distance, interpret the new target as a conditional geometric median, and analyze its properties and the effect of a contaminated target distribution. Experiments on CIFAR-10 with a portion of samples contaminated show a clear advantage for the proposed method under severe contamination.
In briefHow Reliable Are Fast Medical Decision Models? Jev, Jev-like Models, and LLMs Compared
SummaryThe study compares various fast, intuitive-style decision models across 15 medical datasets with 8,742 instances under a unified protocol, assessing accuracy, decision stability, and recognition of unreliable decisions. Jev performs comparably to leading models on text-driven tasks but is weaker on numerically driven ones. Most evaluated models show limitations in recognizing missing correct answers and in using optional abstention. Normalizing option-wise binary scores yields higher retained accuracy than choice scoring.
In briefRetrieval Disagreement as a Signal: Diagnosing and Adjusting Multilingual Dense Embedding Models
SummaryMultilingual dense retrievers map all languages into one shared vector space. The paper finds that the same question asked in different languages retrieves substantially different documents, because correct answers in each language get ranked far down under non-relevant documents in a shared index, not because of weak language understanding. The authors propose a disagreement-guided steering method that adds a language-specific direction to queries, which most strongly reduces same-language crowding while staying competitive on recall, though it trades off against mean removal on ranking quality.
Faeze Ghorbanpour, Shaghayegh Kolli, Jana Diesner et al.
In briefAPTree: Extended Behavior Trees for Automated Robot Task Planning in Construction
SummaryThis paper presents APTree, a model-driven framework that uses behavior trees as its execution model and combines automated task planning with modular task execution. At runtime, it repeatedly applies and coordinates several symbolic planners to generate and compose modular task plans, and includes a domain-specific language for specifying and validating those plans. The authors evaluated APTree on the real-world livMatS Biomimetic Shell pavilion, where it detected injected errors and generated large-scale task plans. A machine-checked formal analysis shows that composed execution and local replanning are safe under stated assumptions.
Shermin Sherkat, Thomas Wortmann, Andreas Wortmann
In briefLost in Implicit Routing: Why LLM Agents Fail to Load Relevant Skills as Libraries Grow
SummaryThis paper examines whether LLM agents can keep loading the right reusable skills as skill libraries grow. Holding tasks and relevant skills fixed while adding irrelevant ones, the authors find that loaded-skill recall drops from 73.6% to 27.4% with 1,000 irrelevant skills, while explicit preselection retains 73.9% recall. The authors attribute the decline to semantic interference from skill descriptions and instruction interference between skill choice and task execution.
In briefLanguage Model Logit Geometry: Common Patterns and Learned Deviations
SummaryThe paper finds that, when a language model's vocabulary logits are ordered by rank, they follow a consistent log-linear relation across contexts, architectures, and scales, emerging gradually during pretraining. Order-statistic baselines explain much of this geometry, but the trained slope α (about 1.4 to 1.7) is well above the initialization value (about 0.2). Because α correlates with predictive performance, the authors argue it reflects learned structure beyond generic extreme-value effects.
In briefSparse-view 4D CBCT reconstruction using spatial-temporal consistency learning to cut scan dose and time
SummaryThis paper presents STCNet, a framework that reconstructs 4D cone-beam CT images from only 16 projections per respiratory phase. It uses physics-driven positional encoding for geometric consistency and temporal encoding in latent space to capture respiratory motion across phases. Tested on the TCIA-LUNG and CIRS thorax phantom datasets, it outperformed conventional and recent deep learning baselines in image quality and artifact reduction.
In briefImprinting: How Knowledge Distillation Can Shift a Student Model's Behavior
SummaryThis paper studies whether a teacher model passes on behavioral tendencies, not only task skills, when its responses are used for knowledge distillation. The authors built two versions of one base teacher, one accepting harm and one avoiding it, and distilled their responses to unrelated reasoning and knowledge questions into different student models. Students acquired the task ability while their safety-related decisions moved toward the source teacher's disposition, an effect called imprinting. Post-distillation preference optimization increased harm avoidance but did not remove the difference caused by the source teacher.
WhyShows distillation on unrelated data transfers teacher safety dispositions via controlled teacher variants, a novel finding with moderate experimental scale.
In briefCurvature-aware pruning that also keeps influence functions intact
SummaryThe paper introduces CACTUS, an ADMM-based optimizer for pruning and sparse-plus-low-rank decomposition of pre-trained weights, using a curvature-aware objective that preserves gradient alignment between compressed and dense models. At 50% pruning of Qwen3-4B, it cuts the C4 perplexity gap to the dense model by 45% and runs over 28 times faster than LLM-Surgeon. The authors also show that preserving influence functions supports effective data selection.
In briefA Global View of How LLMs Value Human Time
SummaryThis paper examines how large language models value people's time by changing only the user's country and measuring the compensation they deem appropriate for a 60-minute appointment delay. Across 193 countries and 5 LLMs, results are compared with GDP per capita on a purchasing power parity basis, mean income, and a literature-based patience score. The authors find that LLMs evaluate time differently across regions, recommending the highest compensation for Europe and mostly the lowest for Africa, and underestimating the value of time for users in Asia and South America relative to GDP_PPP.
In briefApproximating optimal allocation in asynchronous multi-player bandits with mixed activation rates
SummaryThis paper studies multi-player bandits where devices are active only sometimes and each has a different activation probability. It shows that computing the optimal allocation is NP-hard. It then proves a polynomial-time greedy algorithm achieves a 1/e approximation, improving to 1/√e when players are processed in decreasing order of activation probability.
In briefSmart Buildings Challenge: Training general building-control models that work across many real buildings
SummaryBuildings account for nearly 40% of US carbon emissions, with HVAC systems responsible for a large share. Most existing building control methods work only for the specific building they were developed for. This paper proposes a challenge to develop building foundation models that adapt easily to new buildings, with participants training control policies on simulated buildings and evaluated on unseen ones.
Judah Goldfeder, Ayush Rai, Philippe M. Wyder et al.
In briefCoding agents on long clinical trial documents: where they fail when digitizing protocols
SummaryThis paper tests whether coding agents can reliably turn complete clinical trial protocol PDFs into accurate structured data. The authors build TrialPD-Bench, an expert-annotated benchmark of 150 public protocols, and evaluate nine models across five agent harnesses. Even the strongest agents fail to fully solve the task; they usually find the right evidence but make errors when transcribing and structuring it. The authors also propose TrialScribe, a domain-specific harness that improves an open-weight model and lets it match or surpass some proprietary models.
In briefSemi-Amortized Bayesian Inference with Reusable Local Posteriors: a method that trains neural estimators for parameter blocks and alternates them with standard Markov chain updates.
SummaryThis paper proposes a semi-amortized Bayesian inference method that splits model parameters into smaller blocks and uses conditional independence to determine the information each block needs. Neural estimators are trained for blocks whose updates are expensive or repeated, learning their full conditional posteriors from simulations. A Markov chain alternates direct draws from these estimators with conventional updates of the other variables, and neural updates need no likelihood evaluation or accept-reject step. The authors illustrate the approach on high-dimensional, computationally expensive hierarchical and spatiotemporal models.
Niek Stevenson, Paul-Christian B\"urkner, Andrew Heathcote et al.
