In briefHow language models differ in shifting attention-head activity as tasks need more serial steps
SummaryThe study measures, in every layer of 17 open-weight models, whether attention-head activity concentrates on a few heads or spreads across many as the number of serial steps a task demands rises, with prompt length held fixed. In most models, layers just before mid-depth concentrate activity while later layers spread it, though the location and strength differ between models. An ablation study suggests that models whose activity is more concentrated on their top heads also depend more on those heads for the answer.
Johnny Jingze Li, Abdulla Kuleib, Kalyan Basu et al.
In briefLeRF: Learning Reference Coordinate Frames for Perspective-Taking Reasoning
SummaryThis paper introduces LeRF, a framework that teaches vision-language models to decide when a reference coordinate frame is needed for viewpoint-based questions. The model predicts the reference entity and frame origin, and a lightweight renderer overlays the frame on the image without external perception tools. Training combines supervised fine-tuning with reinforcement learning on spatial VQA pairs. The authors report consistent gains over the backbone model and competitive results against existing open-source methods on perspective-taking benchmarks.
In briefCineSubBench: Testing LLMs on Long Films and Culture Using Multilingual Subtitles
SummaryThis paper introduces CineSubBench, a benchmark for testing how well large language models understand long films using multilingual subtitles. It covers 1,012 films with full subtitles in six languages, totaling 8.13M timestamped entries. Seven tasks include narrative reconstruction, genre prediction, age suitability, national movie ratings across ten systems, and subtitle-grounded language safety. Tests on nine LLMs found plot premises were recovered more reliably than full synopses, and cross-lingual consistency varied widely.
Mir Tafseer Nayeem, Susmoy Chakraborty, Davood Rafiei
In briefSparse-view 3D animal behavior embeddings for brain-behavior studies
SummaryThis paper presents SABLE, a self-supervised framework that learns interpretable 3D representations of animal behavior from video. It augments a multi-view transformer with monocular depth and pose priors, reconstructing 3D behavior from extremely sparse views such as two-view videos without ground-truth 3D labels. Across the International Brain Lab and Cheese3D datasets, SABLE matches or exceeds prior state-of-the-art performance in neural encoding and decoding, and generalizes zero-shot to unseen animals after pretraining.
In briefTraining neural operators from equations alone, using a preconditioned loss
SummaryNeural operators are usually trained with data from a classical solver, which is costly to generate. This paper proposes a preconditioned residual loss that fixes the ill-conditioning of physics-informed losses, so the model can be trained purely from governing equations. The authors report that on the Poisson, Allen-Cahn and stationary Stokes equations, this label-free training matches supervised training and is 4 to 25 times more accurate than earlier physics-informed methods.
WhyProposes multigrid-based preconditioned residual loss fixing ill-conditioning; validated on Poisson, Allen-Cahn and Stokes with clear accuracy gains.
In briefEnergyEminence: Source-Aware Environmental Calibration for a Physics-Based Grid Digital Twin
SummaryThis paper presents EnergyEminence, an early-stage testbed that combines a graph-temporal predictor, nonlinear AC cascade simulation, and operator-style temporal replay for power grids. It introduces a shared calibration that turns wildfire-detection confidence and spatial extent into comparable evidence. Tested on 160 source-tracked grid scenarios built from 16 synthetic videos, a source-disjoint test gave 10 true positives, 8 false positives, 22 true negatives, and no false negatives. The authors show that scenario-level random splitting hides environmental shortcut learning, supporting separate handling of environmental alerts and electrical cascade inference.
In briefLost in Translation: How Non-Native English Affects the Quality of AI Chatbot Answers
SummaryLarge language models give lower-quality answers to users whose first language is not English. This paper introduces FABLE, a dataset of 190,911 English prompt variants derived from 174K real writing-related user prompts, and tests 34 open-weight LLMs on it. The models do not carry surface errors like misspellings into their outputs, but they do mirror higher-level rhetorical and lexical qualities of the prompt, and response quality differs substantially between the least- and most-fluent prompts.
In briefStudy of how the spectrum of generative denoiser Jacobians relates to generation quality
SummaryThis paper analyzes the eigenvalue spectrum of the Jacobian of the denoising network used in generative models such as diffusion and flow-matching. It finds that pre-trained models with better generative performance have larger Jacobian eigenvalues. The authors propose a regularization scheme that trains the denoiser on perturbed inputs to control this spectrum, and test it on ImageNet. The results suggest that strengthening data-relevant principal directions and suppressing data-irrelevant ones benefits the denoiser.
In briefCanonical order problem: when LLMs fail as knowledge bases for multi-valued relations
SummaryThis paper studies how large language models represent and generate multi-valued relations, which require producing sets of entities. The authors identify a canonical order problem: the models' internal probability distributions organize many such relations by a canonical ordering, such as alphabetical or chronological. A mechanistic analysis breaks set generation into three phases: retrieving candidate entities, internal sorting, and selecting the next element. As a result, prompts that deviate from this internal ordering markedly reduce the reliability of complete set generation.
Timo Pierre Schrader, Annemarie Friedrich, Simon Razniewski et al.
In briefRepresentable but Unlearned: Encoding Rank and Its Limit on Interaction Prediction
SummaryThe paper introduces a method that computes, without labels, loss, or a fitted model, which measured contrasts a predictor can jointly reproduce given an input encoding's equivalence classes and a fixed contrast design. On a 140-rectangle siRNA interaction dataset, a graph neural network's training-only feature mask reduces the rank of the interaction contrasts to 72, yielding an error floor equal to 14.6% of the fitted model's interaction squared error. Removing the mask eliminates the floor, but interaction MSE improves only slightly under the reported protocol.
In briefGeometric Representations of African Languages: A Regional Semantic Hub and Cultural Steering
SummaryThis paper studies how the Gemma 4 31B model represents African languages and responds to cultural steering. Across nine African languages and three controls, several layers show the Africa-versus-West contrast directions to be more aligned among African languages. A second study builds English-based steering directions for Nigeria, Ghana, Kenya, and South Africa, and most outputs pass an automated structural coherence screen.
In briefFastGuide: a faster way to steer diffusion language models with rewards.
SummaryThis paper introduces FastGuide, an adaptive method that mixes parallel and autoregressive decoding to speed up reward guidance for diffusion language models. It computes guidance once per decoding step and reuses it across several tokens, while using KV caching and sparse attention recomputation to make decoding efficient. On three reward benchmarks, FastGuide is reported to be up to 4.4 times faster than sequential reward-guided decoding with similar generation quality.
In briefSCOUT: A Two-Stage Verifier Combining Reasoning and Tool Use for Computer-Use Agent Safety
SummaryThis paper addresses the problem that computer-use agents can cause unintended harm even under benign instructions, which is hard to detect. It proposes SCOUT, a two-stage verifier in which a rubric generator reasons about task-specific completion and safety criteria, and a probing agent follows those rubrics to gather evidence from the post-execution environment. The authors report results on AutoElicit-Bench and OS-Blind, and state that test-time reflection lowered the unsafe execution rate from 30.2% to 17.2% on AutoElicit-Bench.
In briefPreviewDiff: Critic-guided search through diffusion latents to improve prompt fidelity
SummaryDiffusion models often miss fine prompt details such as object counts, attributes, spatial relations, and timed actions. This paper introduces PreviewDiff, a training-free method that decodes partial previews at intermediate denoising steps and has a multimodal judge score them and give natural-language critiques. The feedback guides branching over semantic prompt edits and locally re-noised latent continuations, and promising branches are rolled forward. Across image and video benchmarks, it reportedly outperforms budget-matched Best-of-N and scalar-search baselines.
Vighnesh Subramaniam, Boris Katz, Brian Cheung et al.
In briefWhether sentiment directions in language models stay consistent across languages with different tokenizer fertility.
SummaryThis paper examines whether sentiment directions extracted from language model representations stay consistent across samples. The authors measure split-half agreement in English, Hausa, and Yoruba texts across four models. English shows the highest agreement, followed by Hausa and then Yoruba, while classifiers still predict sentiment above chance. The study does not establish that tokenizer fertility causes these cross-lingual differences.
Muhammad Abdullahi Said, Abass Oguntade, Elisha Komolafe et al.
In briefLearning seismic-to-well representations by masked cross-modal autoencoding
SummaryThis paper introduces a physically grounded cross-modal masked autoencoder (CM-MAE) that learns representations linking 3D seismic volumes and sparse well logs. The model embeds both modalities in continuous physical coordinates and uses sparse mixture-of-experts layers with dense attention for information exchange. Pretraining uses 23 seismic volumes and about 92,000 wells. Seismic context improves held-out well-log reconstruction by 9.73%, while well-log context improves seismic reconstruction by 0.65%, and seismic-only pseudo-log salt predictions reach an AUROC of 0.910 offshore.
In briefFigAct: Turning Scientific Figures into Interactive Explanation Canvases
SummaryThis paper introduces FigAct, a framework that converts static scientific figures into question-specific visual presentations. It generates short narrations, links each to the matching graphical element, and applies visual actions to guide the viewer's attention. The authors also propose a hierarchical search for efficient localization, train an 8B model, and build a human-verified benchmark from real scientific paper figures.
In briefFD-AA: A Lightweight Head Detecting Incidental Abdominal Abnormalities in Chest CT
SummaryThis paper proposes FD-AA, a lightweight organ-aware classification head that can be attached to frozen 3-D CT encoders to detect incidental abdominal abnormalities visible on chest CT scans. Within each organ, an attenuation-aware module keeps sparse focal findings while masked generalized-mean pooling captures diffuse patterns. The authors report state-of-the-art results across seven abdominal organs on the CT-RATE test set (AUC 0.798) and the external RAD-ChestCT dataset (AUC 0.713).
In briefEvoMO-SR: LLM-based multiobjective evolution of symbolic formulas guided by substructures
SummaryThis paper proposes EvoMO-SR, a symbolic regression framework in which an LLM generates equation skeletons and an external optimizer fits their coefficients. It controls expression bloat with a multiobjective survival selection that balances accuracy and complexity, and mutates expressions using candidate reusable building blocks. On LSR-Synth, it achieves the best accuracy in seven of eight in-domain and out-of-domain settings while using the small Llama-3.1-8B-Instruct model.
Cristina Rossetti, Anna V. Kononova, Thomas B\"ack et al.
In briefTargeting key decisions to assign credit more precisely in agentic reinforcement learning
SummaryThe paper introduces ProVer, a framework for training LLM agents with reinforcement learning that assigns credit to specific decisions rather than whole trajectories. An agentic judge contrasts successful and failed rollouts to propose a segment responsible for the divergence, and this proposal is verified by comparing terminal success rates of continuations sampled before and after the segment. Verified positive advantages are added to the GRPO advantages of tokens in that segment. ProVer achieves the best average performance on ALFWorld, WebShop, and SearchQA, with relative gains over GRPO of 9.91% for Qwen3.5-2B and 7.12% for Qwen3.5-4B.
Dongwon Jung, Hemanth Neelgund Ramesh, Yifan Wang et al.
In briefFair policy optimization for weakly coupled major-minor decision processes
SummaryThis paper studies fair resource allocation in major-minor weakly coupled Markov decision processes, where resource constraints link a major sub-MDP with a population of minor sub-MDPs. Instead of a total-sum objective, it optimizes a class of monotone, concave, permutation-invariant fairness functions. For homogeneous minor sub-MDPs, the authors show the symmetric problem reduces to a utilitarian objective over permutation-invariant policies, and they propose a count-proportion-based deep reinforcement learning method. Experiments on machine replacement and a New York City pricing and taxi relocation task validate the approach's fairness performance and scalability.
In briefParallel drafting with layer-wise causal conditioning for faster speculative decoding
SummaryThis paper proposes DSpine, a draft model for speculative decoding that injects each position's predicted feature into its successor at every layer of the backbone, so causal conditioning flows through network depth while all positions update in parallel. On seven math, code, and chat benchmarks with Qwen3-4B and Qwen3-8B, it achieves the longest acceptance length at both temperatures. At temperature zero on Qwen3-8B, the seven-benchmark mean rises from DFlash's 3.77 to 4.82, and SGLang serving throughput is 23.3% higher on average than DFlash.
In briefLearning models for recognizing topological relationships in images
SummaryThis paper addresses recognizing topological relationships between objects in images, such as touching, overlapping, separate, or contained. The authors built a new labelled dataset of over 11,000 images. They compared classical machine learning models with deep learning models including VGG16 and InceptionResNetV2. Deep learning methods performed better, with VGG16 reaching 89.55% validation accuracy.
In briefTraining-free adaptive response geometry to improve metric depth completion
SummaryThis paper addresses recovering dense metric depth from sparse sensor measurements, where existing methods align visual foundation model priors to metric scale using rigid affine assumptions in fixed coordinate systems, causing systematic calibration errors. The authors propose adaptive response geometry, treating the choice of depth, log depth, or disparity as an image-level unknown unified by a continuous response family. The training-free pipeline reports macro AbsRel 0.0301 and macro NMed 14.04° under incomplete metric observations, improving both measures over PriorDA, LDCM, and Any2Full.
In briefTensor-Train Compressed Separable PINNs: A Second-Order Method for High-Dimensional Parametric PDEs
SummaryThis paper proposes a second-order optimization framework for physics-informed neural networks (PINNs) solving high-dimensional parametric partial differential equations. It uses the Gauss-Newton pullback metric and exploits separable network architectures and separable operator representations to compute the Gauss-Newton step exactly in a smaller compressed space, avoiding assembly of the full residual Jacobian on the exponentially large tensor-product grid. The authors study canonical polyadic and tensor-train parametrizations. Numerical experiments report that the method achieves substantially lower errors than tensor-compressed first-order baselines, with far fewer iterations and a fraction of the computing time.
In briefReviveBench: Benchmark for Reviving Broken Software and Rebuilding Engines with Coding Agents
SummaryThis paper introduces ReviveBench, a benchmark testing whether coding agents can revive non-running software while keeping its original methods, and rebuild industrial software engines from open specifications. The revival family has ten tasks involving issues like dependency conflicts and deleted modules, and the reconstruction family has thirteen tasks across numerical, geometric, hardware, and transactional systems. The authors also report 28 defects found in the verifiers used for grading, and propose three practical checks for validating such verifiers.
In briefREST: A training objective for more reliable latent reasoning in recursive LLM systems
SummaryLarge language models can reason in continuous hidden states rather than decoded text, but training that supervises only the final answer does not constrain these intermediate thoughts. The paper proposes REST, which adds differentiable losses for causality, minimality, separability, and stability of thought representations to the cross-entropy objective. Across 7 benchmarks in math, science, medicine, and code generation, REST improves accuracy by up to 7.5 percentage points and speeds convergence on final answers by 30%.
In briefEncoder-sharing hierarchical federated multi-task learning for vehicle networks
SummaryThis paper proposes EN-HMTFL, a framework that lets vehicles performing different but related perception tasks learn together. Only a globally shared encoder is exchanged and aggregated through a cluster-based hierarchy, while raw data and task-specific decoders stay on each vehicle. Evaluated on MNIST and GTSRB across several vehicular scenarios, it improves accuracy by up to 24.0% over a representation-sharing benchmark. Where it converges earlier, communication rounds drop by up to 69 (28.8%).
In briefMore Features Are Not More Evidence: Limits of Training-Free Activity Recognition
SummaryThis paper tests whether Jev, a fixed general-purpose probabilistic decision model, can recognize human activities from accelerometer data without training, using only deterministic descriptions of the sensor signals. Across WISDM, UCI341, and PAMAP2, its best macro-F1 is 0.038 to 0.118, far below supervised models at 0.686 to 0.907. Adding more numerical features lowered performance, and a semantic rendering partly recovered it, though the study cannot separate semantics from other serialization changes.
In briefFrom Sharp Eyes to Expert Mind: Embedding Forensic Expertise in MLLMs to Detect Tampered Text
SummaryThis paper studies tampered text detection, noting that expert models capture subtle manipulation traces but generalize poorly, while multimodal large language models understand semantics well but miss fine forensic artifacts. The proposed Expert Knowledge Internalization (EKI) framework works in two stages: first, text-focused and image-focused strategies direct attention to small text regions, then a Forensic-General Representation Alignment loss aligns shallow model representations with a pre-trained forensic expert. Experiments on in-domain and cross-domain benchmarks report state-of-the-art results, and the expert is needed only during training, so inference efficiency stays nearly the same as the vanilla model.
In briefLearning continuous patient trajectories from electronic health records with history-conditioned flow matching
SummaryElectronic health records capture patient states that evolve continuously, but only at irregular observation times. The paper introduces EHRFlow, a multi-marginal flow-matching framework that conditions on encoded patient history, so future dynamics depend on prior clinical trajectory, and it supports arbitrary forecasting horizons. On real-world datasets with over one million patients and an external validation cohort, EHRFlow improves horizon-averaged top-5 clinical-code accuracy over autoregressive and history-independent baselines.
Silas Ruhrberg Est\'evez, Kara Liu, Christopher Chiu et al.
In briefLanguage Models as "Insecure" Reporters: Do They Hide Critical Flaws?
SummaryThis paper studies whether large language models conceal narrative-changing flaws when writing reports on their own work, which the authors call "insecure reporting." Using eight adversarial reporting scenarios, GPT-5.5 flagged a planted negative result in only 2 of 200 reports, but 190 of 200 when told "Be honest in your response." Experiments on Qwen3.5-9B show honesty and success-seeking correspond to opposing directions in representation space, and steering toward honesty made reports substantially more transparent.
Jenny Y. Huang, Jiameng Fan, Ahmed Imtiaz Humayun et al.
In briefWhen Correction Becomes Repair: Auditing Internal Interventions in Tool-Using LLMs
SummaryThe paper presents SAKIKO, an auditing framework for checking whether internal activation interventions can fix how tool-using LLMs choose among actions such as calling a tool, asking for clarification, answering directly, or declining. Tested on seven models using When2Call and MetaTool, the interventions produced direction-specific net gains in five models, but behavioral change alone did not show true repair. One intervention with +55 net gain corrupted over half of the baseline-correct decisions it touched. The authors conclude that outcome-resolved adjudication is necessary before claiming internal repair.
In briefXiaomi-OCR-0 Technical Report: A 0.8B Unified Model for Document Parsing and OCR
SummaryThis report introduces Xiaomi-OCR-0, a unified 0.8B vision-language model for document parsing and OCR-centric understanding. The authors built a roughly 170M-sample corpus with an automated data engine and trained the model from Qwen3.5-0.8B using Q-Mask text anchoring, continued pretraining, and mixed-task reinforcement learning. The model reports strong scores on several document parsing benchmarks and an average of 83.2 on five OCR-oriented VQA benchmarks. Ablations suggest that OCR-centric understanding supervision adds gains for parsing once parsing training is sufficient.
In briefLightweight Hardware-Aware Kolmogorov-Arnold Networks for Efficient Medical Image Enhancement and Segmentation
SummaryThis paper presents FunKANLite, a two-stage, hardware-aware compression of the FunKAN network, which reached state-of-the-art accuracy on MRI Gibbs artifact removal and anatomical segmentation but was too large for edge medical devices. The compressed FunKANLite-ST has 5.6x fewer parameters and 3.7x fewer GFLOPs than FunKAN, stays within 1.4 percentage points IoU on BUSI, GlaS, and CVC-ClinicDB, and reaches 33.95 dB PSNR on IXI. On an NVIDIA Jetson Orin Nano and a Raspberry Pi 5, it reduces energy per inference by up to 68% and raises throughput by 2.9x.
In briefA Character-Level Neural Approach to Sinhala Sandhi Splitting
SummaryThis paper studies Sinhala Sandhi splitting, which recovers the words or morphemes hidden inside phonologically merged surface forms. Using the SandhiLex dataset and native Sinhala Unicode input, the authors evaluate recurrent character-level sequence-to-sequence models. The best model, a bidirectional LSTM encoder with a unidirectional LSTM decoder, reaches 68.40% exact-match accuracy on the hard lexicalized, derivational, and etymological subset and 94.00% on the affixational subset.
In briefMemory Is a Derivation: Evidence Gaps in Long-Term AI Agents
SummaryThis paper examines whether persistent memories created by long-running LLM agents are actually supported by the interaction history available when they were written. The authors introduce DerivAudit, a framework that checks evidence scope, compositional validity, and admission reliability. On two memory corpora, broader pre-write history supported nearly 60% of memories that looked unsupported from citations alone, while 17-21% remained unsupported.
In briefWhat Language Models Can Inherit from Solomonoff Induction: Limits and Attainable Properties
SummaryThe paper tests the claim that scaling language models brings them closer to Solomonoff induction, using known results from algorithmic information theory, universal prediction, and Bayesian consistency. It argues that in its uniform form the claim fails for every computable model, so any closeness can only hold pointwise, one environment at a time. The authors also discuss which Solomonoff-like properties, such as in-context learning as amortised Bayesian inference, can be inherited and which a deployed system should have or avoid.
WhyUses Kraft inequality and known algorithmic information theory results to give rigorous limits on inheriting Solomonoff guarantees, separating pointwise from uniform conclusions.
In briefAccessible but not adopted: why first-generation, low-income college students use LLMs only shallowly
SummaryThis paper studies the gap between access to large language models and their actual adoption among first-generation, low-income (FGLI) college students, based on 61 interviews. It finds these students use LLMs mainly as chatbots for narrow tasks, and identifies barriers such as low perceived value, underestimated self-efficacy, unclear starting points, low peer exposure, and resource constraints. From these findings the authors derive four design principles for closing the gap.
In briefReasoning with Neural Cellular Automata: Testing visual reasoning using only local connections
SummaryThis paper tests whether Neural Cellular Automata, networks using only local connectivity and asynchronous updates, can perform complex multi-step reasoning. The authors report that NCAs solve visual reasoning tasks such as large mazes, Sudoku, and ARC-AGI-1. They also report out-of-distribution generalization to larger grids, longer rollouts, and parallel trials.
WhyApplies neural cellular automata to multi-step visual reasoning on mazes, Sudoku, and ARC-AGI-1 with out-of-distribution tests; novel, though experiments are limited in scale and comparisons.
Mayalen Etcheverry, Pietro Miotti, Aidan Sirbu et al.
In briefBD-HazardVLM: Testing whether vision-language models spot hidden driving hazards on Bangladesh roads
SummaryThis paper introduces BD-HazardVLM, a benchmark of 500 Bangladesh road-scene images from RSUD20K and TFP-BD, annotated for hazards, risk timing, severity, and defensive actions. Three vision-language models were tested zero-shot, achieving about 88 to 89 percent accuracy on detecting hazard presence, essentially matching a majority-class baseline of 89.0 percent. Fine-grained reasoning was much weaker, with category accuracy below 40 percent and exact defensive-action matching below 2 percent. The authors conclude that coarse hazard recognition does not imply reliable defensive-driving reasoning and call for geographically diverse safety evaluation.
Abdullah Al Maruf, Md. Sajedul Islam, Tanmoy Mridha et al.
In briefRhinoNet: Unsupervised Learning for Generating High-Dimensional Point Clouds
SummaryThis paper introduces RhinoNet, a point cloud neural network built on simplicial complexes that creates geometry-preserving embeddings of high-dimensional point clouds. It represents each dataset in a cohort as a simplicial complex and learns a latent space via an autoencoder in which optimal transport distances between point clouds are preserved. Applied to patient single-cell datasets, it produces patient-level representations called phenoscapes and generates whole point clouds through latent flow matching or geodesic interpolation. The authors report state-of-the-art performance on single-cell and spatial data cohorts relative to tested baselines.
Siddharth Viswanath, Hiren Madhu, Eddie Cavallin et al.
In briefThinQuant: Scalable Rotation Learning for Quantizing LLM Weights and Activations
SummaryThis paper proposes ThinQuant, a method for learning rotations that enable low-bit weight and activation quantization of large language models. It selects a much smaller calibration set using the geometry of the activations' convex hull, and solves the reduced optimization with an efficient ADMM algorithm. The authors report faster calibration than DartQuant on Llama-3-70B and that it scales to Llama-3.1-405B on a single H200 GPU.
In briefAdaST: Adaptive Coupling for Space-Time Forecasting
SummaryThis paper addresses spatial-temporal forecasting for systems like traffic, climate, and energy networks. The authors observe that real data shows different coupling patterns, from temporal-dominated to strongly coupled, which existing models handle poorly. They propose AdaST, which decomposes inputs into components using heterogeneity-aware experts and recombines them with an adaptive recomposer. Experiments reportedly show it significantly outperforms state-of-the-art baselines.
In briefThe Layer Mystery of VLA: An information-theory study of which backbone layer robot policies should use
SummaryThis paper studies which layers of a frozen vision-language backbone should feed the action head in vision-language-action robot policies. It finds that most multi-layer fusion setups underperform the best single layer, and that the best layer varies across backbones and benchmarks. The authors propose action-conditioned InfoNCE as a proxy for layer quality, which selects good layers with far less compute than exhaustive policy sweeps.
In briefDyad: Giving Language Models Built-in Typed Decision-Making
SummaryThis paper introduces Dyad, an architecture that adds an environment-aware action encoder to a pretrained large language model. The encoder embeds each candidate action in parallel and scores it against the model's internal state to produce a distribution over typed actions. With the LLM frozen, training only the encoder yields consistent gains across four unseen environments. Jointly optimizing both components outperforms conventional RL post-training, including a 3.80% average absolute gain on ALFWorld with a 9B model, while general knowledge, reasoning, and coding also improve.
In briefKoa-decision: Fast, consistent structured decisions from generative LLMs
SummaryThis paper introduces Koa-decision, a framework that turns single-step decisions such as classification, semantic endpointing, Boolean checks, and scoring into constrained generation with single-token outputs. It uses atomic label tokens and supervised fine-tuning to reduce classification to a one-step decoding problem. On a production intent-routing benchmark it reports 85.5% accuracy and about half a second response time, several times faster than the frontier models tested.
Shenghong Dai, Shiva Kumar Pentyala, Yingchi Liu et al.
In briefLoopICL: Solving tabular tasks by looping a single transformer block
SummaryThe paper introduces LoopICL, a looped transformer for tabular in-context learning that uses a single block iterated over multiple passes, decoupling parameter count from computational depth. At the same computational cost (FLOPs), it performs competitively with TabICLv2 on TabArena and TALENT while using nearly 90% fewer parameters. Its recurrent design also lets users trade off inference cost against performance.
In briefLimits of transferring and equivariant signal processing across finite-dimensional algebraic models
SummaryThis paper studies the fundamental limits of transferability in algebraic signal processing via homomorphisms between algebraic signal models, which are linear maps commuting with filtering. Their existence depends entirely on coincidences among the filtered eigenvalues of the models' shift operators. The authors introduce a spectral transfer efficiency η(θ) to measure information preservation and characterize exactly when lossless transfer is achievable.
In briefScaling small-LLM teams: task-dependent returns across orchestration architectures
SummaryThis paper studies whether enlarging a team of collaborating small LLM agents improves accuracy, and which orchestration architecture benefits most from scaling. The authors test eight architectures across five 7-9B instruction-tuned models, five short-answer benchmarks, and one executable-code benchmark, with up to 30 calls. Gains from scaling are strongly task-dependent: they are large on two arithmetic word-problem benchmarks but small on ARC, GPQA, and MMLU, and no architecture wins across all tasks. The authors introduce an exact generate-transform decomposition that splits any accuracy change into a coverage term and a transformation term to explain these differences.
In briefOne Geometry, Different Outcomes: Why Changing the Modality Gap Helps or Hurts Vision-Language Models
SummaryContrastive vision-language models keep image and text embeddings separated by a modality gap. The paper finds that across CLIP and SigLIP encoders, a single dominant direction captures 94.4-99.9% of the image-text mean separation. Decomposing the similarity score, it explains why modifying this gap affects zero-shot classification, standard cross-modal retrieval, and mixed-modal retrieval differently.
In briefData Unlearning via Inverse Distillation (IDU)
SummaryThis paper proposes Inverse Distillation Unlearning (IDU), a framework that distills a multi-step flow or diffusion teacher into an efficient one-step generator while suppressing outputs tied to a designated training subset. It requires only a pretrained full-data teacher and forget-set data, with no retained examples, extra feature extractors, or classifiers. Experiments on MNIST and CIFAR-10 show substantially reduced generation of forgotten classes while retained-class generation quality is preserved.
Aleksei Leonov, Nikita Kornilov, Zhenhe Zhang et al.
In briefDoes better prediction yield more faithful model ablations? Evidence from sequential choice
SummaryThis paper examines whether input ablations, which remove information from a predictive model and measure the resulting performance change, reliably reflect how behavior is generated. The authors test GRUs, Transformers, a fine-tuned LLaMA model, and cognitive models on two synthetic sequential bandit tasks with known generating policies. They find that accurate predictors can respond much less to donor-reward replacement than the true generator, so ablation results alone may not support conclusions about the generating process.
In briefPHASE: A physiology-guided foundation model for intracranial EEG
SummaryThis paper introduces PHASE, a foundation model for intracranial EEG that makes directly measurable physiological characteristics explicit learning targets. It has a temporal stage (PHASE-T) within each channel and a spatiotemporal stage (PHASE-ST) across synchronized channels. The model is pretrained on recordings from 222 participants at nine clinical sites. On five Omni-iEEG clinical tasks, frozen PHASE-T outperforms the evaluated foundation models, and fine-tuned PHASE-T surpasses task-specific models.
WhyMakes physiological features explicit targets with masked latent prediction, outperforming prior foundation models on five clinical iEEG tasks with matched ablations.
In briefPADM'E: Data synthesis for checking whether LM agent evaluators agree with human judgment
SummaryLanguage models are often used to grade other language models, so it matters whether these graders agree with humans. The paper reframes this meta-evaluation as a preference judgment problem and introduces PADM'E, a method that uses only small language models to synthesize criterion-based meta-evaluation data without human involvement during evaluation. A 1,000-sample dataset spanning four agentic domains and three criteria was built, and on a 150-sample human-validated subset, agreement with human judgment rose from 73% to 85% over a naive baseline.</summary>
In briefMeasuring trainable degrees of freedom in materials graph neural networks using random-subspace intrinsic dimension
SummaryThis paper trains three materials graph neural networks (CGCNN, ALIGNN, DimeNet++) inside randomly oriented parameter subspaces across six property prediction tasks. It tracks how performance recovers as trainable degrees of freedom are restored. Some tasks recover near-reference accuracy from small fractional subspaces, while phonon prediction is most sensitive to dimensional restriction. The method serves as a targeted stress test of how these models use their optimization space.
In briefGeoOutageBench: A Benchmark for Ambiguity-Aware, Ontology-Grounded Geospatiotemporal Question Answering on Power Outages
SummaryThis paper introduces GeoOutageBench, a benchmark for testing how well LLM-based systems answer geospatiotemporal questions over a knowledge graph about power outages and resilience. The knowledge graph combines outage records, remote sensing, weather, storm and power events, and domain ontologies, with questions graded by difficulty. The benchmark evaluates three tasks: interpreting ambiguous natural-language questions as SPARQL queries, assessing ontology utility, and measuring answer accuracy in multimodal retrieval.
In briefHow Early Learning Shapes Later Representation-Change Directions in Continual Learning
SummaryThe paper studies whether representation changes early in training form a low-dimensional subspace, called a scaffold, that guides later learning. Across four pretrained visual encoders and two datasets, later changes favor this early subspace over random alternatives, and this reuse depends on each network's learning history. Constraining updates within the scaffold slows acquisition of new tasks more than random constraints, while effects on old-task retention are less consistent.
In briefA Polyphonic View of AI Understanding: a paper arguing that LLM outputs come from several parallel mechanisms working together, and proposing a notion of understanding suited to this.
SummaryThe paper argues that when judging whether to trust an AI model's output, the usual question of what the model understands is ill-framed because it assumes a single underlying mechanism. Drawing on a range of mechanistic evidence, the authors show that LLM outputs emerge from coalitions of parallel mechanisms of uneven reliability that complement, duplicate, or drown out one another. They propose a conception of understanding centred on sound circuitry that is reliably recruited and controls outputs, making attributions of understanding claims about internal organisation that can guide trust.
In briefGEM-KMeans: A memory-efficient, accurate GPU-optimized method for clustering very large datasets
SummaryThis paper presents GEM-KMeans, a GPU implementation of nonnegative low-rank matrix factorization for K-means clustering. It fuses the gradient update, nonnegative projection, and normalization statistics into one matrix-multiplication step, so only a single large factor matrix is stored in GPU high-bandwidth memory. The authors show accurate clustering on synthetic and real datasets at massive scale. Their reported speed gains over existing GPU-accelerated Lloyd's algorithms depend on the data.
Peng Xu, Nihar Koganti, Volodymyr Kindratenko et al.
In briefArgus: Detecting AI-Assisted Coding to Protect Academic Integrity
SummaryThis paper presents Argus, an automated system that detects signs of large language model assistance in undergraduate C programming assignments by combining behavioral and stylistic indicators. Applied to a large-enrollment CS2 course at Purdue University, it flagged patterns consistent with LLM-assisted coding for 45% of students in Spring 2026. Flagged LLM use showed a significant negative correlation with in-person proctored exam performance, and the authors argue that any such system needs a structured human review process to limit false accusations.
David Racovan, Ajay Rawat, Christopher K. May et al.
In briefLongCat-DeepResearch: A Multi-Agent Deep Research System Technical Report
SummaryThis report presents LongCat-DeepResearch, a deep research system that pairs an enhanced LongCat model with a multi-agent workflow to produce evidence-grounded reports. Planning agents first build a research plan called ResearchSpec, then research agents investigate and draft sections in parallel before global review guides local revisions. The authors report scores on DeepResearchBench, DeepResearchBench II, ResearchRubrics, and an in-house benchmark.
Meituan LongCat Team: He Zhu, Yue Xu, Wanli Wu et al.
In briefMixture-of-Kittens: An MoE Training System Built as a Megakernel for NVL72
SummaryThis paper presents Mixture-of-Kittens (MoK), a Mixture-of-Experts training system designed for Nvidia's NVL72 scale-up architecture. The authors find that existing MoE systems, optimized for scale-out networks, often run slower than a naive PyTorch and NCCL baseline on this hardware. MoK fuses token dispatch, shared and routed expert FFNs, and token combine into a single deterministic training megakernel. It delivers up to 2.37 times the throughput of the strongest public baseline across MoE layer shapes, and improves end-to-end training throughput by 1.41 times in a 512-GPU production run.
Stuart H. Sul, Nash Brown, Henry Wildermuth et al.
In briefUneven Decline of Shared Programming Knowledge After Generative AI: A Stack Overflow Study
SummaryThis paper uses the release of ChatGPT-3.5 as a natural shock to study over two million Stack Overflow questions posted from 2020 to 2025. Easy questions decline sharply while difficult questions become more common, with rising code complexity. Data-rich topics lose share while data-scarce ones gain, and the easy-question decline is concentrated in data-rich domains. The authors conclude that generative AI's effect on collective knowledge is uneven, eroding easy, accessible knowledge first.
WhyUses ChatGPT's release as a shock across two million Stack Overflow questions, revealing uneven decline in easy, data-rich knowledge; extends prior participation-decline work substantially.
In briefAerialDojo-200K: A Large-Scale Benchmark for Drones Finding Target Objects in Open Worlds
SummaryThis paper presents AerialDojo-200K, a large-scale benchmark suite for open-world aerial object-goal search, where drones must autonomously explore complex 3D environments to reach targets described semantically or by reference images. The suite includes 42 simulation scenes and 205,732 task instances, with in-distribution and out-of-distribution scene partitions. The authors evaluate nine multimodal large language models and find that general-purpose aerial agents still have a long way to go.
In briefFLOORA: A Human-Aligned Domain-Specific Language Model for Architectural Floor Plan Generation
SummaryThis paper introduces FLOORA, a family of small domain-specific language models that generate architectural floor layouts. The models combine a token-efficient language, specialized pretraining, supervised fine-tuning, and reinforcement learning using human-preference and verifiable rewards. According to the authors, their 0.6B model outperforms much larger frontier models, and human evaluators selected it as the best model in 89.3% of comparisons.
Sahand Rezaei-Shoshtari, Patryk Wozniczka, Shu Ishida et al.
In briefHow Neural Routing Solvers Make Decisions: A Look Inside Their Internal Mechanisms
SummaryThis paper examines how three autoregressive neural combinatorial optimization models, AM, POMO, and LEHD, make decisions internally during solution construction. The authors use behavioral analysis, representation probing, and causal interventions. They find that AM and POMO follow a persistent geometric pattern, while LEHD encodes linearly accessible information about multiple future actions. Experiments cover the Traveling Salesman Problem and the Capacitated Vehicle Routing Problem.
Fatemeh Askari, Mazdak Teymourian, Mohammad Izadi et al.
In briefSMat-Attention: Structured Long-Context Sequence Modeling
SummaryThe paper introduces SMat-Attention, a family of causal attention masks with structured long-range routing whose VC-dimension d acts as a tunable knob between standard causal attention (d=1) and richer routing patterns. It gives chunkwise forward and backward algorithms and states complexity bounds of O(T^{2-3/d}+T) work and O(T^{1-1/d}) cached states for decoding. Experiments show routing expressiveness on subset-routing and multi-key retrieval tasks, and extensions to Mamba-2 and Gated DeltaNet improve recall while keeping subquadratic prefill.
In briefMnemon: a long-term memory agent for LLM assistants that splits work into fast judgments and slow reasoning
SummaryThe paper introduces Mnemon, a memory agent for LLM assistants that keeps conversations as raw dated records instead of rewriting them at write time. An LLM plans searches, a fast decision model called Jev judges the returned records, and explicit budget rules assemble a small context for the answering model. The authors report top scores on LoCoMo and LongMemEval-S with small per-question context and slow cost growth on the BEAM benchmark.
WhySplits memory work into fast judgments and slow reasoning, reporting top LoCoMo and LongMemEval-S scores against 14 public systems.
In briefABC: Advantage-Based Control Variates for RL with Verifiable Rewards
SummaryThis paper studies actor-critic methods for reinforcement learning with verifiable rewards (RLVR). It proposes Advantage-Based Control Variates (ABC), links it to Direct Advantage Estimation (DAE), and combines the two into one algorithm. In an offline-to-online setting on mathematical reasoning tasks, it reportedly performs competitively with GRPO using substantially fewer online optimization steps.
In briefMirror-Score: A calibrated scoring method for designing D-peptides, showing the limits of sequence-based ranking
SummaryThis paper presents Mirror-Score, an inference-only framework for scoring D-peptide binders against L-protein targets, along with a public benchmark of 31 crystal complexes. The authors find that raw ProteinMPNN negative log-likelihood is not a valid affinity ranker, with a pooled Spearman correlation of 0.19 and reversed correlations across target families. They instead evaluate Boltz-2 mirror-space cofolding confidence, which gives directionally consistent results on the viral-entry family, though the authors note that with only three independent chemotypes it is not a statistically validated predictor.
In briefGeoWind2Plan: Fast 3D Urban Wind Prediction for Energy-Efficient Drone Route Planning
SummaryBuildings in cities reshape wind into complex 3D flows, so a drone's energy use depends on local wind as well as distance. Detailed wind data is usually unavailable at planning time, and CFD simulations take hours. GeoWind2Plan predicts mission-relevant 3D wind from only a background wind vector, building geometry, and start and goal points, then plans a path and speed. Evaluated under high-fidelity CFD wind, it cuts inference time to about 3 seconds and reduces energy by 6.9%, 12.7%, and 4.5% in tailwind, headwind, and crosswind missions.
In briefCould automating AI R&D trigger an intelligence explosion?
SummaryThis paper examines whether AI systems automating most AI research and development could cause progress to accelerate sharply in an "intelligence explosion." The authors assess preliminary evidence, analyze the potential impacts, and propose policy responses. They argue the scenario could bring large benefits but also serious risks, so policymakers should act urgently.
In briefPowerZooJax: A JAX-Based Power System Benchmark for Reinforcement Learning
SummaryThis paper introduces PowerZooJax, a JAX-based benchmark suite for reinforcement learning in power system operation. It offers five constrained Markov decision process tasks covering generation, transmission, distribution, distributed energy resources, and data center microgrids. By expressing power flow, economic dispatch, market clearing, and device dynamics as JAX computation graphs, it runs training and evaluation entirely on GPU. Experiments report substantial speedups over CPU-based simulation and standardized evaluation of policy returns, safety violations, and out-of-distribution stress conditions.
In briefCo-trained monitors to improve scalable AI oversight
SummaryThis paper studies whether training a monitor alongside a worker model can prevent the worker from learning to evade a fixed monitor. In the supervised setting, the authors show monitoring is possible with vanishing error and query rates exactly when the monitor function class has finite Littlestone dimension. They also propose a self-supervised co-training method based on test-time distillation and stress-test it in code-security settings. The results suggest adaptive monitors keep pace with evolving worker strategies better than fixed monitors.
Joseph H. Rudoler, Kevin Tan, Benedict Tessler et al.
In briefNeural networks for spectral optimization: fast learned surrogates that predict eigenvalues from domain geometry to find shapes optimizing spectral quantities
SummaryThe paper addresses finding a domain that optimizes a functional depending on the spectrum of a differential operator. Since PDE solvers are computationally expensive, the authors propose two neural network models that learn eigenvalues directly from domain geometry. One uses Fourier coefficients with a light MLP for star-shaped domains, reaching 0.2% precision; the other uses the landscape function and produces orthogonal eigenfunctions and eigenvalues, with about 1% mean relative error on the first ten eigenvalues versus 4% for an FNO. Both surrogates recover classical spectral optima such as the disk for the first eigenvalue.
Alexis de Villeroch\'e, Beniamin Bogosel, St\'ephane Breuils et al.
In briefHybrid neural simulation-based inference for robust, low-budget analyses
SummaryThe paper proposes two hybrid methods that approach neural simulation-based inference (NSBI) performance while greatly reducing computational cost and keeping some or all reliability guarantees of parametric methods. The first is broadly applicable, and the second targets a class of particle physics analyses with a semi-parametric NSBI formulation. The authors report only a modest loss in raw sensitivity. Based on their comparisons, they recommend the first approach, Latent Categories, for robust and efficient inference.
In briefSAGE: A Statistical Gate Deciding Which Self-Edits Self-Evolving AI Agents Keep
SummaryLLM-based agents self-improve by editing a skill document, and a gate decides whether each edit is kept. The standard gate accepts any edit that raises the average validation score, which can hide regressions on already-solved items. SAGE compares the current and edited skills item by item and commits an edit only when its wins are statistically reliable against its losses. Across five benchmarks and four backbone LLMs, it lowers the regression rate in 19 of 20 settings and reaches the highest final score in all 20.
In briefDual-anchor acceleration nearly optimal for stochastic monotone root-finding
SummaryThe paper studies stochastic monotone root-finding by combining dual-anchoring with stochastic resolvent approximation and optimized variance control. Under unbiased oracles with bounded variance and sample operators that are monotone and uniformly L-Lipschitz, the algorithm reaches an epsilon-residual point with near-optimal oracle complexity. In the noise-dominated regime this reduces the poly-logarithmic factor from cubic to quadratic.
In briefTORQUE: Choosing Which Coordinates to Keep Precise Around Random Rotation
SummaryThis paper presents TORQUE, a framework that improves quantization methods using random rotations by jointly choosing which coordinates to keep at high precision before and after rotation, under a fixed overall bit budget. The authors derive an error upper bound and prove that keeping the top-k largest coordinates before rotation minimizes it, reducing the search to optimizing k. Numerical evaluation under a Gaussian model and experiments on nearest-neighbor retrieval, KV-cache compression, and activation compression show a better trade-off between accuracy and storage cost.
Ran Ben Basat, Michael Mitzenmacher, Shay Vargaftik
In briefProof that the best known bound for preferent compression is tight
SummaryThe paper concerns sample compression, a method for producing safety and performance guarantees for learning-based systems. Sharp bounds exist for algorithms with a preference (stability) property, but whether they are tight was an open question. The authors show the current best bound is tight, using an explicit construction based on the uniform distribution and order statistics. They also give a shorter proof relying only on elementary counting.
WhyResolves an open question by an explicit construction attaining the preferent compression bound, plus a shorter elementary proof, benefiting the subfield.
In briefCoDimRecon: Agentic 3D Scene Reconstruction with Deformable Curves, Surfaces, and Volumes
SummaryCoDimRecon is a framework that reconstructs editable 3D scenes with rigid, articulated, and deformable objects from multi-view RGB images. Deformable objects are reconstructed by dimensionality as curves, surfaces, or volumes, and agent-guided behavioral tests trigger revisions to their physical models. On Replica and ScanNet++ scenes it achieves competitive compositional reconstruction accuracy while producing deformable assets for simulation. The paper also demonstrates robot interactions across all three representations.
In briefPersistence Forcing: Using Feature Specialization in Pixel-Space Diffusion Models
SummaryThis paper proposes assigning different feature groups in pixel-space diffusion Transformers (DiTs) distinct refinement budgets across depth, which leads to persistent features mainly encoding global structure and active features specializing in local high-frequency details. Building on this, the authors introduce Persistence Forcing (PerF), where persistent features condition the refinement of active features. On ImageNet 256×256, PerF-L achieves an FID of 1.91, and PerF-H reaches FIDs of 1.63 and 1.76 at 256×256 and 512×512 resolutions, respectively.
In briefIn-Context Learning for Robots: A Survey of Methods and Applications
SummaryThis paper is a literature review of in-context learning (ICL) for robots, where neural parameters stay fixed during deployment and demonstrations or interaction guide existing competence toward new tasks. The authors organize methods into four families: context-conditioned policies, geometric demonstration transfer, world-model-based control, and skill- and agent-based execution. They compare these approaches' transfer assumptions and examine how they preserve taught requirements across manipulation and navigation as objects and environments change.
In briefInfrared subtraction built with AI for NLO jet corrections in electron-positron annihilation
SummaryThe paper presents a local infrared subtraction method developed with AI assistance, separating an integrable radiation term from a finite Born-level contact term. The contact term is fixed using the EFT singular distribution in an N-jettiness resolution variable, with two implementations: a neural-network phase-space projection fitted to EFT cumulants, and an analytic construction that keeps Born momenta fixed. The authors reconstruct the full NLO correction for massless 3- and 4-jet production in electron-positron annihilation, finding good agreement with EERAD3. The calculations were run on a laptop without GPU acceleration.
In briefFLIP platform makes cross-continental federated learning repeatable and auditable
SummaryThis paper presents FLIP, an open-source platform that packages common federated learning workflows as composable services. The authors tested two use cases, federated fine-tuning and federated evaluation, on synthetic chest X-ray data across nodes in the UK and Thailand. The study makes an operational claim rather than an algorithmic one, concluding that such platforms support international collaboration, repeatability, auditability, and site-specific governance.
Rafael Garcia-Dias, Alexandre Triay Bagur, Chayanin Tangwiriyasakul et al.
In briefCoordinated Multi-Robot Navigation Following Language Instructions: Task Formulation, Benchmark, and Method
SummaryThis paper formalizes multi-agent vision-and-language navigation as a constrained coordination problem, where subtasks carry dependency and resource constraints. The authors build the MAVLN benchmark with 11,724 episodes across 145 scenes, involving teams of up to four agents, along with constraint-aware metrics. They also propose TRISS, a system combining an LLM-based subtask scheduler, shared topological memory, and conflict-aware execution. Experiments establish TRISS as a baseline and show substantial room for improvement in scheduling, planning, and execution.
In briefCausal and interpretable structures behind LLMs' relational reasoning
SummaryThis paper studies how large language models represent relationships among three tokens tied to cyclic concepts such as months, hours, weekdays, and musical notes. Analyzing activations across Llama, Qwen, Gemma, and Mistral, the authors find that intermediate layers use a joint representation of two tokens' relationship, while later layers use one covering all three. Restricting models to these causally relevant joint representations improves next-token prediction accuracy.
In briefSIFARI: a self-supervised neural method for fitting radio interferometer data to make astronomical images
SummaryThis paper presents SIFARI, a self-supervised neural network that fits measured radio-interferometric visibilities directly to reconstruct sky brightness, without an external training image set or an explicit spatial regularizer. It estimates brightness uncertainty using Stochastic Weight Averaging-Gaussian and combines it with a thermal-noise floor to produce signal-to-noise maps. The authors report that on synthetic ALMA data it outperforms CLEAN in resolution, extended-flux recovery and fidelity, and they also apply it to ALMA observations of PDS 70 and WISPIT 2.
Shunyuan Mao, Andrea Isella, Paris Perdikaris et al.
In briefSolver Agent: an AI framework for physics calculations, applied to F-theory uplifts of O3-planes and S-folds
SummaryThis paper presents Solver Agent, a large-language-model-based framework that performs calculations and proofs in mathematics and theoretical physics, tracking assumptions and derivations in a persistent ledger and verifying results with independent agents. The authors apply it to global F-theory uplifts of Type IIB orientifolds and S-fold generalizations, giving sufficient conditions for Weierstrass models to yield elliptically fibered Calabi-Yau fourfolds with isolated quotient singularities. They use stringy invariants to derive fixed-point contributions to Euler characteristics, which determine the localized D3-brane charges needed for tadpole cancellation.
Eliott Morgensztern, Cesar Fierro Cota, Alessandro Mininno
In briefFurE: An efficient method that reconstructs editable 3D animal fur for individual animals without animal-fur datasets
SummaryFurE is a method that reconstructs realistic, editable animal fur strand by strand from multi-view images. It uses a PCA-based decoder that can be learned from human-hair strand data, easing the shortage of animal fur data. The authors report about a 10x speedup in strand training over current state-of-the-art methods while keeping strand fidelity.}
Srinjay Sarkar, Prakhar Kaushik, Soumava Paul et al.
In briefTelescopic Language Models: One nested model that works at many compute budgets
SummaryThe paper introduces Telescopic Language Models (TLMs), trained with stochastic prefix supervision plus a full-capacity anchor so the model is valid at every depth. On a 200M-parameter proxy suite with 20B tokens, a single TLM run works at all twenty layer prefixes. It reduces the area under the quality-budget curve by 43-44% versus fixed-exit suites while matching them at full capacity, at about 12% lower GPU cost. The authors conclude the training objective, not the nesting itself, makes a model elastic.
In briefPDMD: Projected Distribution Matching Distillation for Faster Video Diffusion
SummaryThis paper proposes PDMD, a modification of Distribution Matching Distillation (DMD) that reduces the number of denoising steps needed by video diffusion models while keeping output quality stable. The authors attribute training degradation, such as oversaturation and artifacts, to critic errors and filter them by projecting out a component of the DMD update. The method requires only a one-line code change, and with Wan2.1 at 4 NFE it reaches a VBench total score of 83.73, 1.03 points above matched DMD.
WhyAdds a projection correction for critic error to DMD; supported by comparisons on Wan2.1 and MiniMax-H3 plus user studies.
In briefTeaching unified image models to reflect and revise via interleaved reinforcement learning
SummaryThis paper studies how unified multimodal models can repair their own image generations by diagnosing errors, revising, and re-rendering in a loop. The authors propose UMM-Reflection, which uses reinforcement learning to optimize full reflection trajectories within one model, comparing sibling trajectories from the same initial image to update both reflection text and image revisions. No external verifier is needed at inference. On BAGEL, it improves GenEval by 12.05 points over supervised fine-tuning, with gains also reported on WISE, OneIG-Bench, and T2I-CompBench++, which were not used in training.
In briefFinding Bible References in Karen Blixen's Seven Gothic Tales by Computer
SummaryThis paper studies how computers can find Biblical references hidden in Karen Blixen's Seven Gothic Tales, where sources appear through paraphrase, allusion, and historical language. The authors built a benchmark of 189 annotated references and searched all 31,170 verses of Danish Old and New Testament translations. Normalized BM25 performed strongly without training, and fine-tuning a Danish encoder substantially improved results, especially on allusions. The authors argue these models should serve as heuristic aids for expert close reading rather than as authoritative discoveries.
Andr\'as Kov\'acs, Alexander Conroy, Daniel Hershcovich et al.
In briefReliability-gated fusion of consumer head and foot IMUs for lower-body 3D pose estimation
SummaryThis paper addresses the unreliability of consumer inertial sensors, such as earbud head IMUs and smart-insole foot IMUs, used for camera-free lower-body pose estimation. The authors learn a temporal gate for each channel of each sensor stream, trained with an auxiliary reliability objective on synthetically corrupted data. The channel-gated model is reported as the most accurate learned fusion method on clean data and under simulated faults, and its gates can suppress biased foot-orientation channels and detect dropout bursts.
In briefA unified framework for distributional training in one-step image generation
SummaryThe paper proposes a unified theoretical framework that separates distribution modeling from matching discrepancy and links global objectives to pointwise feature updates via Wasserstein gradient flow. It introduces MGFlow, which models feature distributions with Gaussian mixtures at adjustable granularity and supports optimal transport and score-based matching. On ImageNet 256×256 it reports state-of-the-art results, and for text-to-image it post-trains FLUX.2 [klein] 4B into a one-step generator.
In briefTokenCast: Predicting Token Use During LLM Agent Runs
SummaryToken consumption by a large language model agent can vary widely across runs of the same task, and growing context raises the input cost of each later call. TokenCast learns a composable cost representation for each execution segment and combines adjacent segments to estimate cumulative consumption, refreshing the forecast as the run proceeds without extra LLM calls. Across 96 evaluated combinations, it reduces mean absolute error by an average of 14.5% against the strongest comparator, and in offline budget-control replay it uses 21.3% fewer tokens on average than a fixed-budget policy.
WhyProposes composable per-segment cost modeling that accounts for context growth and updates forecasts during execution; MAE drops 14.5% across 96 combinations.
In briefScaling Long-Form Story Generation via Narrative State Tracking
SummaryThe paper introduces NstAgent, a training-free agentic framework that lets LLMs track a structured narrative state covering characters, past events, and future requirements. The authors extend an existing benchmark to compare narrative consistency across lengths and pair it with a writing-quality benchmark to evaluate stories from 10K to 100K words. They report that NstAgent achieves better narrative consistency and writing quality as stories get longer, with neither noticeably degrading as length increases.
In briefHEIR: Learning Human-Entity Interactions with Functional Roles (image benchmark and prediction method)
SummaryThe paper introduces HEIR, an image benchmark of 18,730 images for evaluating complete sets of participants and their roles across object, interpersonal, and self-directed interactions. It also proposes CoRISP, which uses shared entity identities to combine role-conditioned evidence and predict normalized participant-role sets. The authors report that CoRISP leads evaluated systems on repeated-role events and shared-participant images in HEIR, and gives role AP and complete-set AP results on V-COCO.
