In briefA multi-objective aligned small language model for generating dialogue from patients with substance use disorder
SummaryThis paper proposes a cognitively grounded framework for generating patient responses in substance use disorder (SUD) counseling. It first detects latent cognitive components, then generates dialogue aligned with patient histories and counselor questions. Small language models are trained using knowledge distillation from larger teacher models, preference optimization from human annotations, and attention-guided reward shaping. Reported evaluations show that cognitively informed fine-tuning improves cognitive realization and alignment over generic instruction-tuned baselines and mental-health-specific small models.
Thushara Manjari Naduvilakandy, Hyeju Jang, Mohammad Al Hasan
In briefAccuracy-Information Tradeoff for Loss-Difference Conditional Mutual Information
SummaryThis paper studies loss-difference conditional mutual information (ld-CMI), which measures how much a learner's loss differences reveal about which candidate in each pair it was trained on. The authors show that, for linear predictors with smooth convex losses and suitable regularizers on specified distributions, any proper learner achieving small excess risk at the optimal sample size must have ld-CMI of order n bits. As a result, range-scaled ld-CMI bounds cannot vanish there, even though these learners' generalization gap is O(n^{-1/2}).
WhyShows accuracy forces ld-CMI growth with order-n lower bounds, a substantive theoretical advance with rigorous proofs.
In briefDo Vision Models Learn Physics or Rendering Shortcuts? A Counterfactual Test for Physical Consistency
SummaryThis paper examines whether image editing results violate scene physics, such as missing shadows or floating objects. The authors build a counterfactual benchmark of 5,500 Mitsuba 3 renders from 500 scene families, covering five violation types. Three vision-language models are tested as baselines, reaching 64.0–67.8% category macro-F1 on standard scenes but dropping to 40.8% for LLaVA-1.5-7B under intervention shift, suggesting part of the in-distribution accuracy relies on rendering cues.
M. Moein Esfahani, Sepehr Salem, Mohammed Alser et al.
In briefOptimal information complexity of VC learning: recovering the best PAC guarantee via conditional mutual information
SummaryThis paper studies the information complexity of learning algorithms using the Conditional Mutual Information (CMI) framework, focusing on the evaluated CMI (eCMI). The authors construct a randomized Majority-of-5 learner whose eCMI is O(d) in the realizable case, where d is the VC dimension. This shows that the optimal PAC guarantee for VC classes can be recovered from algorithm-dependent CMI analyses.
WhyConstructs a randomized majority-of-5 learner achieving O(d) eCMI, recovering the optimal VC PAC guarantee via CMI analysis.
In brief"Few Bits, One Law: Toward W2A4KV2": a unified training framework, CanonQ, for extreme low-bit LLM compression
SummaryThis paper introduces CanonQ, a quantization-aware training framework that compresses the weights, activations, and KV caches of large language models together. It maps tensors from different sources into canonical coordinates using fixed rotations and normalization, then jointly trains the model to adapt to the coupled quantization errors. Under W2A4KV2 compression on LLaMA3 models, the authors report lower WikiText-2 perplexity and higher zero-shot accuracy than prior baselines.
In briefStyleFields: Multi-Scale AdaIN Implicit SDFs for Coarse-to-Fine 3D Shape Editing
SummaryStyleFields is a DeepSDF-based 3D reconstruction architecture that injects latent codes through multi-level Adaptive Instance Normalization at several decoder depths, with auxiliary heads trained on a coarse-to-fine schedule. This lets the coarse structure of one shape combine with the fine details of another, without part labels or adversarial training. The authors also apply a learned surrogate drag predictor as a differentiable objective to optimize reconstructed cars.
Ehsan Garaaghaji, Nicolas Talabot, Pascal Fua et al.
In briefFreeEvolve: A Method That Learns to Evolve Beyond Fixed Loops
SummaryFreeEvolve automates the optimization of prompts, skills, and workflows for language model agents, replacing hand-designed fixed search loops. Within limits set by the environment, the evolver itself decides what to test, how much evidence to gather, which candidates to pursue, and when to stop, following an editable evolution skill refined through meta-evolution. On four benchmarks, the authors report an average gain of 13.6 points on the primary held-out metric, matching or exceeding hand-designed evolvers.
In briefExact dynamics and finite-sample recovery of a simple linear model that learns data features step by step.
SummaryThe paper studies linear recursive feature machines (RFMs), which learn data representations by alternately fitting a predictor and updating features via the average gradient outer product. It analyzes their dynamics in noisy multi-output regression with a low-rank teacher matrix. The authors show the learned feature matrix stays close to its infinite-data counterpart at every iteration, with error decaying as O(√(d/n)) with high probability. Experiments on real text and single-cell gene-expression data illustrate the learned features.
In briefStableGrasp: Recovering physically stable hand grasps from a single image
SummaryThis paper presents StableGrasp, a method that reconstructs physically stable human hand grasps from a single RGB image. It uses differentiable simulation-based optimization that separates the visual hand pose from the control target governing grasp forces. The approach jointly optimizes hand geometry and control by minimizing grasp kinetic energy, with regularization to keep the result visually consistent and geometrically plausible. The authors report that the grasps are substantially more stable than those from alternative hand-control strategies under rigorous physical simulation.
In briefPatient, Place, Prior (P³): What Counts as Personalization in Medical World Models?
SummaryThe paper examines whether longitudinal imaging forecasts truly use a patient's own history rather than relying on population averages. It introduces the P³ audit, which tests use of patient history, patient-matched lesion maps, and predictive gain over population-average predictions. The authors also propose Cancer JEPA, which forecasts future breast DCE-MRI representations during neoadjuvant therapy. In a validation cohort, patient-history inputs reduced forecast error, but the 95% interval versus the population-average correction includes zero.
In briefBookkeeping, Composition, or Unreachable Gold? Checking MemoryAgentBench's Conflict-Resolution Scores Against a Frozen Last-Write Resolver
SummaryThis paper examines the Conflict Resolution split of MemoryAgentBench, which is read as measuring selective forgetting. The authors implement the benchmark's own rule, where the newest statement about a fact wins, as a frozen resolver with no learning, tested on one of four fact lists. Under the official metric it answers 80.25% of questions (74.5% on the three held-out lists). The authors argue that scores should be read per item rather than in aggregate, since 67 items have gold answers that the last-write graph cannot reach but overwritten statements can.
In briefR-CNN-Based Chess Position Recognition: a method that reads a chess position from a single photo of a board.
SummaryThe paper proposes an R-CNN-based framework that recognizes a chess position from a single image of a three-dimensional board. It separates piece recognition from board geometry estimation and combines both to rebuild the position. On a synthetic dataset, mean average precision for piece detection rose from 61.59% to 90.14%, and the full framework recovered 76.61% of test positions exactly.
In briefFrom Probabilities to Decisions: Search and Multi-Teacher Distillation with Jev
SummaryThis paper studies Jev, a fast model that outputs calibrated probabilities for fixed choices without generating text, applied to decision tasks under tight constraints. In bullet chess, placing Jev's judgment inside Stockfish search lets the bot climb above a 2200 Lichess bullet rating against other bots. Because live model calls are too slow, the authors distill pairwise judgments into a compact evaluator, and averaging Jev and Qwen3-32B labels beats spending the whole budget on Qwen alone.
Mohamad Yazan Sadoun, Sarah Sharif, Yaser Mike Banad
In briefPattern Recognition vs. Step-by-Step Reasoning in Language Models
SummaryThe paper argues that pattern recognition and step-by-step reasoning lie on a spectrum, depending on how much preceding context the next token relies on. It models reasoning traces as paths on a De Bruijn graph and reports that fine-tuning Qwen2.5-1.5B-Instruct shows frequent states give higher accuracy but more fragility to perturbations. The authors also state that Qwen3-14B and Qwen3-32B keep over 75% accuracy on several benchmarks under restricted attention windows.
In briefOne-Frame 4D Cardiac MRI Synthesis Without ECG Gating
SummaryThis arXiv paper presents PhaseFlow3D, which generates a full 4D cardiac cine MRI sequence from a single end-diastolic 3D volume without ECG. It models the cardiac cycle as a piecewise linear phase and uses Radial Contraction Decomposition to produce 3D displacement fields that warp the volume into each frame. On the combined ACDC and M&Ms benchmark, it reports the lowest ejection fraction mean absolute error, and downstream tests cover segmentation, pathology classification, label propagation, and strain analysis.
WhyNovel ECG-free 4D cardiac cine synthesis combining piecewise phase and radial decomposition; leads multiple metrics on ACDC and M&Ms with ablations and downstream validation.
In briefQ-PACE: dynamic precision allocation for training language models with quantization
SummaryThis paper proposes Q-PACE, a method for quantization-aware training of LLMs that assigns precision to each layer. It uses a second-order sensitivity model that predicts loss increase from per-layer quantization noise, and periodically recomputes these coefficients during training. In pretraining and supervised fine-tuning experiments on models up to 4B parameters, Q-PACE reportedly outperforms existing mixed-precision recipes and reaches comparable loss with lower total memory. The authors also note that quantization sensitivity is predictable by depth and layer type and stable during training, allowing infrequent, cheap recalibration.
Alexandra Volkova, Matin Ansaripour, Erik Schultheis et al.
In briefLayerRoPE: a method that adjusts residual-stream normalization gains by layer depth
SummaryThe authors examine 16 pre-trained LLMs and find that the growth of hidden-state norms with depth reflects a positional encoding carried by the normalization weight γ, which grows in magnitude and rotates in direction as depth increases. They propose LayerRoPE, which replaces all per-layer γ vectors with one shared vector and depth-conditioned scalars. The paper reports fewer parameters, a FLOPs change under 0.02%, better results than several normalization baselines, and improved learning-rate sensitivity.
WhyReinterprets residual-norm gain as depth-positional encoding and proposes a RoPE-like method; novel, validated across 16 models and multiple scales.
In briefAttention-based Gaussian mixture model for predicting vehicle trajectories with context
SummaryThis paper proposes a Context-Aware Attention-based Gaussian Mixture Model (CAA-GMM) that predicts several possible future vehicle paths along with their uncertainty. The model conditions on scene context and agent dynamics, using interpretable Gaussian components to represent diverse behaviors. A lightweight attention mechanism fuses rasterized environment cues with motion history. On the nuScenes and Argoverse 2 datasets, it reports accuracy competitive with or better than raster-based baselines at low computational cost.
Arash Raftari, Babak Ebrahimi Soorchaei, Yaser P. Fallah
In briefRDGSplat: Rendering-Dedicated Geometry for Novel View Synthesis
SummaryRDGSplat is a framework for generating new camera views from a frozen 3D foundation model. It duplicates the pretrained decoders and trains the copies using only photometric supervision, producing geometry dedicated to rendering while keeping the model's original metric predictions unchanged. A target-pose conditioned adapter reshapes the representation the decoders read. On RE10K, WM2.0 rises from 20.918 dB to 24.266 dB across three feed-forward backbones and four benchmarks.
In briefLower bounds on sequential rounds needed for parallel diffusion sampling, even with many simultaneous queries
SummaryThis paper studies whether diffusion model sampling can be sped up by running many score queries in parallel. The authors prove that sampling smooth, near-isotropic Gaussian mixtures requires about d^{1/3} sequential rounds, and uniform sampling from anisotropic axis-aligned boxes in the unit ball requires d rounds. The bounds apply to arbitrary randomized algorithms under approximate score oracles with inverse-polynomial error.
WhyProves the first polynomial parallel-round lower bounds for diffusion sampling, including a tight linear bound, a substantial theoretical advance with impact mainly in theory.
In briefTraining Language Models to Make Consistent Decisions: Keeping Beliefs Stable, Choosing by Utility, and Recognizing When Information Is Missing
SummaryThis arXiv paper studies whether language models can learn, through supervised fine-tuning, a decision procedure that preserves elicited outcome probabilities and selects the action maximizing expected utility. The authors train on 20 datasets and test transfer to unseen domains, held-out framings, and different payoff structures. They also add incomplete-information settings where required utilities are replaced with irrelevant text, and report that targeted fine-tuning improves coherent decisions and helps models recognize when an action cannot be justified.
In briefToolRACER: A robust conversation simulation resource for training and testing AI agents
SummaryTask-oriented conversational agents often fail in real conversations, especially when users are uncooperative, while existing function-calling benchmarks mostly cover cooperative interactions. The paper presents ToolRACER, a pipeline that coordinates user, assistant, and tool emulation models to generate validated multi-turn dialogues. Using it, the authors build ToolRACERBench, a corpus of 5.6K trajectories across six domains, about 66% of which contain failure-prone scenarios. Models trained on this data show improved end-to-end agentic accuracy on τ²-bench and ACEBench.
Arkajyoti Chakraborty, Aryan Tayal, Ishika Agarwal et al.
In briefSpecGuard: Proving a Task Is Broken Before an Agent Cheats
SummaryThis paper presents SpecGuard, a system that detects conflicts between a coding task's intent and its tests before an autonomous agent acts. It formalizes the task description into a Lean 4 specification and the tests separately, then uses the Lean kernel to check whether any implementation could satisfy both, producing a machine-checked certificate when none can. On conflicted SWE-bench tasks, it detects up to 72.8% of conflicts and certifies up to 51.1%, with a conflict miss rate nearly five times lower than model-based judgment.
WhyAutoformalizing task intent into Lean 4 specs with kernel-checked conflict certificates is novel; SWE-bench results give quantified detection and certification rates, though scale is limited.
In briefsk-bench: A Slovak-first benchmark for evaluating large language models
SummaryMultilingual LLM benchmarks often omit Slovak or cover it only through machine translation. This paper introduces sk-bench, a native-first Slovak benchmark with 30 datasets across ten skill categories, including several resources newly introduced or packaged for generative-model evaluation. The authors evaluate 55 open and closed models under one harness and compare native versus translated data. Findings cover model rankings, the effects of continued Slovak pretraining, and the impact of test-time reasoning.
WhyBuilds a native-first Slovak benchmark, evaluates 55 models, and derives design lessons for under-resourced languages.
Marek \v{S}uppa, Ivan Vykopal, Andrej Ridzik et al.
In briefSampling 2D SU(N) gauge theory by drawing plaquettes independently and reweighting corners
SummaryThe paper studies sampling two-dimensional SU(N) lattice gauge theory with the Wilson plaquette action using generative methods. It rewrites the action in terms of holonomies and turns the link consistency condition into a conditional sampling problem over group commutators, so plaquettes can be sampled independently and the results reweighted. Closed-form commutator densities are derived for SU(2) and SU(3). Tests on 2×2 and 32×32 lattices report per-plaquette acceptance above 99% and effective sample sizes typically above one half.
In briefFrozen AI Models That Learn From Use: A Model-Agnostic Framework for Multimodal Medical AI
SummaryLanguage and vision-language models are usually frozen after deployment, so they cannot learn from the cases they handle, which matters in medicine where guidelines and evidence change. This paper proposes a framework that lets frozen models learn through three kinds of external expertise, a Skill, a Knowledge Memory, and a Multimodal Knowledge Base, without changing model weights. An update is kept only if it helps new cases without hurting earlier ones. Tested on six benchmarks and four base models, the authors report gains of up to 34.2% over the base model during online deployment on medical tasks.
WhyProposes model-agnostic external experience framework, validated on six benchmarks and four models, gaining up to 34.2% on medical tasks.
In briefAdding Noise to Prompts: Improving Continuous Diffusion Language Models by Noising Conditioning Tokens
SummaryThis paper challenges the common practice in continuous diffusion language models of keeping conditioning prompt tokens clean during training. The authors add noise to these prompt tokens during training as well, and report better results on Sudoku, N-Queens, and Gigaword summarization, though gains do not appear in open-ended dialogue generation. The change is a single line in the training objective and adds no inference cost by default.
In briefDIVA: Dual-Space Intent-Aware Visual Attenuation for Robot Vision-Language-Action Policies
SummaryThis paper introduces DIVA, a module for vision-language-action policies that estimates how relevant each visual patch is to the task, combining high-level intent with low-level visual evidence. It attenuates low-relevance visual features both before entering the backbone and inside it, while keeping the full visual token sequence and needing no external grounding supervision. On LIBERO, DIVA raises OpenVLA-OFT's average success from 96.6% to 98.0% and its zero-shot LIBERO-Plus score from 69.6 to 72.6. Real-world experiments show consistent gains under task-irrelevant visual perturbations.
In briefFinding hidden side-effect gaps in AI task reward checks
SummaryThe paper examines whether reward verifiers used in reinforcement learning can catch runs that complete a task but also make an unrequested change. Using mutation testing, the authors found every false pass came from such unrequested changes, not from missing or wrong work. In some benchmarks, verifiers passed duplicated writes, and three recent RL methods give those duplicates full reward. A simple count assertion per template rejected all duplicates while passing the correct runs.
In briefBreaking the Space Barrier and Its Application to Language Model Inference
SummaryThe paper studies automata that enforce structured outputs such as schema-following JSON for language models, observing that their graphs have few paths between any two points. The authors prove that deciding reachability in such graphs, while verifying low path multiplicity, can be done in O(log²n/log log n) space, below the classical O(log²n) bound of Savitch's theorem. They build an inference engine from these constructions and compare it with MLX and llguidance on an Apple M2 Pro, reporting faster completion and smaller grammar size.
In briefDECO: Using Directional Evidence to Shrink the Search Space for Exact DAG Learning
SummaryThis paper addresses learning directed acyclic graphs (DAGs) from observational data, where exact methods face exponential growth in candidate parent sets. The authors propose DECO, a non-parametric hybrid framework that uses dependency and directional evidence to remove unsupported parent sets before exact optimization. Experiments on benchmark Bayesian networks and synthetic data show substantial search-space reduction with competitive structure-recovery performance.
Upala Junaida Islam, Abdelmonem Elrefaey, Rong Pan
In briefHow Likelihood Tempering Affects Limiting Predictive Moments in Variational Bayesian Linear Networks
SummaryIn wide Bayesian neural networks, mean-field variational inference can suffer from prior dominance, where the KL term outweighs the likelihood and the predictive distribution collapses toward the prior as width grows. This paper studies likelihood tempering, which scales the KL term, and asks how quickly the temperature must decrease with width to counter this. For single-hidden-layer linear networks with isotropic Gaussian priors, the authors derive limiting predictive distributions under temperature schedules and compare them with the untempered NNGP posterior. They report phase transitions in the predictive mean and variance at different scales.
In briefCADFather: Automatic CAD Rebuilding from 3D Shapes Using Coordinated Tools
SummaryThis paper introduces CADFather, an autonomous agent system that recovers editable parametric CAD programs from 3D meshes. A vision-language assistant inspects renders and chooses which tools to invoke, while learned and algorithmic tools propose CAD operations and numerical optimization refines parameters. The system keeps alternative candidates per part and preserves the best valid result, using pretrained models without additional training. Reconstruction quality and execution validity are evaluated on DeepCAD, Fusion360, MCB, CADENA-Bench, CADBench, and BenchCAD, along with computational cost analysis.
Gennadiy Savrasov, Maksim Elistratov, Nikita Gavrilov et al.
In briefAdaptive sampling method for generalizable physics-informed neural networks in metal additive manufacturing
SummaryThis paper addresses poor generalization of physics-informed neural networks (PINNs) for thermal modeling in metal additive manufacturing, caused by static collocation sampling. It proposes a two-stage framework: a conditional flow matching model learns high-residual distributions across process conditions, which are mixed with a domain-informed base distribution to generate adaptive collocation points. On metal AM numerical benchmarks, the method reduces relative L2 error by an average of 62.1% under the same collocation budget.
In briefDepth-to-RGB: reusing a frozen depth model to keep object geometry consistent when compositing images
SummaryThis paper presents Depth-to-RGB (D2R), a framework for inserting or replacing objects in a scene using a reference image and a mask. It adapts a frozen depth estimator to predict the depth of the intended composite scene, which then guides RGB synthesis. The authors also introduce AnyInsertion++, with splits that test generalization to categories unseen in training. The reported results show D2R outperforming open-source and closed-source baselines on geometry and image quality metrics.
In briefSpatial Induction Heads: Learning Multidimensional Cellular Automata In Context
SummaryThe paper studies how transformers learn unknown local rules of multidimensional cellular automata when trajectories are flattened into token sequences. The authors propose spatial induction heads, two-layer circuits where the first layer gathers the relevant spatial neighborhood and the second matches it against earlier occurrences. Trained two-layer transformers generalize to unseen one- and two-dimensional rules, with near-perfect deterministic rollouts and less than 0.005 nats KL divergence from the Bayes-optimal predictor on stochastic rules.
In briefAre Parameter-Efficient Fine-tuning Methods Really Different? A Study of Six Methods' Performance, Forgetting, and Weight Geometry
SummaryThis paper compares six parameter-efficient fine-tuning (PEFT) methods on language and diffusion models, examining how their parameterizations relate to task performance, forgetting, and changes in pretrained weight geometry. The authors find that the selected LoRA-family methods also approximately preserve pretrained geometry, and restoring their slightly drifted singular-value spectra largely keeps task performance, questioning whether explicit geometric preservation is necessary. They also report distinct adaptation-retention trade-offs: LoRA most consistently limits forgetting, DoRA often achieves higher mean task scores than LoRA, and PiSSA often incurs greater retention costs.
In briefDeceptive Bandit Problem: How Leaked Exploration Signals Can Destabilize Multi-Agent Learning
SummaryThe paper studies how randomized exploration in multi-agent learning can be exploited for security purposes. A deceptive player uses leaked signals correlated with a victim's exploration and couples them with its own exploratory actions, steering learning toward a new steady state called the deceptive Nash equilibrium (DNE). The authors prove that the resulting dynamics converge to an arbitrarily small neighborhood of the DNE while retaining optimal convergence rates, and illustrate conditions and effects on the deceiver's cost in a resource-allocation game.
In briefWhich Buildings Are AI-Ready? A Measurement-Based Framework for AI Question Answering and Control
SummaryThis paper proposes a framework for assessing how ready a building is for agentic AI systems that answer questions and control operations. A building's knowledge graph sets two ceilings: the share of operational questions its data can answer, and the share of control actions it exposes. A reference AI agent's accuracy on a fixed question set shows how much of these ceilings is realized. In a simulated office, the agent realizes 0.62 of a 0.64 answerable ceiling, and across 37 public real-building graphs the median answerable ceiling is 0.16.
In briefTopoCurve: Geometry-aware Topology Reasoning with Bézier Curves for Self-driving
SummaryThis paper addresses topology reasoning for autonomous driving, which detects 3D lanes and traffic elements from multi-view images and infers how they connect. The proposed TopoCurve models each lane as an endpoint-fixed cubic Bézier curve, giving smooth geometry with exact endpoints and analytic tangents. It reports 50.6 OLS on the OpenLane-V2 benchmark without post-processing and higher endpoint detection scores than prior methods.
In briefWorkhorse: Teaching Humanoid Robots Whole-Body Manipulation from Human Motion Data
SummaryThis paper presents Workhorse, a method that teaches humanoid robots contact-rich whole-body manipulation using robot-free human demonstrations. A visual planner predicts target poses for the torso, both wrists, and both feet, and a reinforcement-learning tracker executes them on the robot. On a real Unitree G1, the robot carries boxes, kicks, catches a thrown box, and climbs a suitcase, and it recovers when pushed or when the box is taken. In simulation, box transport succeeds in 77% of episodes and 64% under 40 N.s pushes.
In briefSPLATIFY: An AI framework that turns 3DGS research papers into trainable code
SummarySPLATIFY is a multi-agent framework that converts 3D Gaussian Splatting papers into trainable gsplat-based implementations. It uses a constrained code grammar, citation recovery, graph-based synthesis, and visual feedback to reproduce methods faithfully. The system can also compose complementary regularizers, losses, and densification strategies to improve results. On the 30-paper SPLATIFY-Bench, it matches expert implementations for papers without public code and improves PSNR by up to 2.4 dB.
In briefFrom Uncertainty to Action: Learning to Steer LLM Agents
SummaryThis paper studies how to decide whether, where, and how to correct an LLM agent's trajectory. The authors intervene at each non-terminal step with four mechanisms and run every continuation to completion, building a table of about 82,000 counterfactual outcomes. They find uncertainty can flag failing trajectories but no single signal reliably identifies the step where steering helps. They propose VoS, a monitor that learns the value of steering at each step from this table and uses a harm-budgeted trigger to limit disruption to successful trajectories.
In briefGeoNatureAgent (GNA): A Testing Framework for AI Tools on Geospatial and Environmental Tasks
SummaryThis paper introduces GeoNatureAgent (GNA), a framework for checking how well AI agents that use tools perform on geospatial and environmental tasks before they are deployed. It provides a fixed set of sixteen tools served through a Model Context Protocol server, plus a 103-task benchmark covering three environmental indicators in Spain and Portugal. Nine language models were tested, with Claude Sonnet 4 scoring highest at 61.7%, while DeepSeek V3.2 reached 93% of that capability at much lower cost. The authors note that part of the gap versus general GIS benchmarks comes from stricter scoring rather than task difficulty alone.
WhyOffers a reusable pre-production evaluation framework and 103-task benchmark, comparing nine LLMs across multiple seeds with substantive empirical analysis.
Gabriel Diaz-Ireland, Diego Prieto-Herr\'aez, Mario Garc\'ia Peces et al.
In briefSame Text, Different Prediction: Serving Conditions Changing Text Classifier Outputs
SummaryThis paper systematically studies how text classifiers change their outputs under different serving conditions, such as batch shape, hardware, and inference engine, with the checkpoint and text held fixed. The authors train 180 models across discriminative, pseudo-generative, and fully generative formulations. They find that stable labels do not imply stable scores, and that fully generative classifiers change more labels than discriminative ones. The paper also gives sufficient conditions for label stability and a mitigation for each mechanism.
In briefPreventing Adversarial Attacks from Transferring to Fine-Tuned Speech Recognition Models
SummaryMany organizations fine-tune public pretrained speech recognition models and deploy them as black boxes, assuming limited access offers protection. This paper shows that adversarial perturbations crafted on the public base model transfer effectively to fine-tuned models, severely degrading performance. The authors propose TransferBreaker, a fine-tuning framework combining three components to suppress adversarial transfer. Evaluated across three languages and four large ASR models, it reduces adversarial word error rate from 92.6 to 27.8.
WhyProposes a unified fine-tuning framework with theory, validated on three languages and four large ASR models, cutting adversarial WER from 92.6 to 27.8.
Mojtaba Nafez, Aref Mousavi, Mohammad Ebrahim Mahdavi et al.
In briefConvex-Concave RL: Rewriting Policy Learning as a Solvable Difference-of-Convex Program
SummaryThis paper shows that maximizing expected return in reinforcement learning, though non-convex, has a difference-of-convex structure in log-density-ratio coordinates. The authors solve each iteration with sequential convex programming, prove convergence under mild conditions, and show that CPI, NPG, TRPO and AWR arise as special cases. Experiments report that the multi-step method performs well on diagnostic MDPs with long dependencies, is competitive with tuned PPO on classic control, and converges faster in a stochastic healthcare domain.
Shripad V. Deshmukh, Yaswanth Chittepu, Dhawal Gupta et al.
In briefConstraint Tree Search (TRACE) for Learning from Language Feedback
SummaryThis paper studies how an agent can learn user intent from natural-language feedback that points to violated requirements. It proposes TRACE, which organizes candidate constraints in a tree and tests each one by generating actions that satisfy it, committing only when feedback does not contradict it. The authors give high-probability coverage bounds and evaluate TRACE on six language-feedback tasks, reporting higher final-output success on RecMovie than the compared prompting baselines.
In briefComputational complexity of deciding whether hidden Markov models can be uniquely identified from data
SummaryThe paper studies whether a hidden Markov model's parameters can be recovered from sampled data, a property called identifiability. Existing conditions are often sufficient but not necessary, leaving many models undetermined. The authors show that the decision problems for several identifiability notions are decidable in PSPACE, and that deterministic variants are already coNP-hard for simply parameterized families.
WhyReduces HMM identifiability decision to the theory of the reals, proving PSPACE decidability and coNP-hardness, going beyond prior sufficient-only conditions.
In briefAre Forecasting Models Fairly Benchmarked? How Preprocessing Affects Time Series Accuracy
SummaryThe paper argues that current time series forecasting benchmarks mostly use simple scaling and ignore transformations such as differencing needed for nonstationary data, which biases results toward models with built-in preprocessing. The authors evaluate 11 forecasting models across 16 reversible preprocessing pipelines on 29,000 M4 time series. They report that optimizing preprocessing per series improves performance by roughly 27% to 87% across models, with architectures lacking built-in preprocessing gaining the most, making simpler models competitive with complex ones.
Guilherme Afonso Galindo Padilha, Paulo Salgado Gomes de Mattos Neto, Rafael Menelau Oliveira e Cruz
In briefMaRK: Adapting Frozen State Space Models to Iterative Generation via Context-Conditioned Recurrence
SummaryThe paper proposes MaRK, a framework that maps context vectors to bounded modulations of a frozen SSM's recurrence and related parameters, so each diffusion timestep reshapes the model's memory kernel. It is applied to a frozen 111M-parameter Hydra backbone with three adapter types, requiring only 6.3–11M trainable auxiliary parameters. The Chebyshev variant performed best, with an average validation loss of 2.55.
In briefFedRSPO+: A heterogeneity-aware federated method for decision-focused learning
SummaryThe paper proposes FedRSPO+, a federated learning framework for decision-focused learning across cross-silo clients whose objectives and feasible sets differ. It builds on RSPO+, a regularized predict-then-optimize surrogate that smooths the decision map through projection. Experiments on synthetic knapsack and shortest-path tasks and a real-world energy pricing task compare it against prediction-only federated learning and decision-focused learning baselines.
Konstantinos Ziliaskopoulos, Alexander Vinel, Jiaqi Wang
In briefTucker Bottleneck Attention: A Faster Attention Method for Multi-Dimensional Sequence Data
SummaryThis paper introduces Tucker bottleneck attention (TuBA), which uses low-rank tensor structure to project hidden tensors into compact Tucker cores, applies self-attention there, and writes results back to the original space, reducing the quadratic cost of attention. An autoregressive extension is also proposed. On video prediction and global weather forecasting, TuBA reports lower error and computation than standard self-attention, with speedups up to 4.27 times.
In briefRippleCP: Measuring Whether Checkpoints Help Coding Agents Recover Faster
SummaryThis paper introduces counterfactual checkpoint advantage, which measures how much future recovery cost is reduced by checkpointing a candidate state instead of skipping it. The authors test 12 SWE-bench Verified tasks with 106 real recovery branches under matched conditions. Checkpointing saves 49.4 seconds per task on average, but the benefit varies sharply: the first checkpoint is highly effective, while a second checkpoint one step later yields a slight loss.
In briefU-Space: Locating When and Why Language Models Become Uncertain
SummaryThe paper introduces U-Space, a low-dimensional subspace that makes a language model's uncertainty measurable and interpretable during reasoning. Semantic anchors for doubt and certainty define an orthogonal basis, and the U-Lens projects each token's state onto it to produce a token-level uncertainty map or a scalar score. The method needs no correctness labels, repeated generations, or training. The authors report that its confidence score outperforms established baselines on reasoning benchmarks under standard and length-controlled evaluation.
WhyProposes a training-free low-dimensional uncertainty subspace, validated against baselines under length-controlled evaluation, with impact mainly on LLM reliability research.
In briefAnaximander: Running Geospatial Deep Learning Models Interactively on Any Compute Backend
SummaryApplying deep learning to satellite imagery in GIS often requires custom deployment and georeferencing code because models use incompatible formats and run in different environments. This paper presents Anaximander, an open-source system that lets users choose a model source and compute location through one interactive interface. It pairs an inference server with a QGIS plugin that handles tiling, result reassembly, and georeferencing. The authors demonstrate a code-free side-by-side comparison of three models on an agricultural field delineation task.
Satej S. Soman, Akram Zaytar, Girmaw A. Tadesse et al.
In briefLEGO: Large-scale Repository Engineering Using Reusable Code Primitives for Agents
SummaryThis paper introduces Code Primitives, reusable executable components with interface contracts, dependency closures, validation tests, and provenance, organized into CodeFace, a searchable library of 1,424 validated primitives. It presents LEGO, a method that activates task-relevant primitives, integrates their adapted implementations, and revises results against executed tests. The authors also build LEGO-REPO, a benchmark of 522 executable repository reconstruction tasks, and report that LEGO improves 13 evaluated backbones by 0.1474 on average.
In briefDISRQAD: A Quality Dataset and Benchmark for Diffusion-Based Image Super-Resolution
SummaryThe paper introduces DISRQAD, a dataset of subjective quality scores (MOS) for 14,000 super-resolution outputs from diffusion and non-diffusion methods across four degradation conditions and x2/x4 upscaling. It evaluates 51 standard and 11 adapted quality metrics, finding much weaker agreement with human ratings on diffusion outputs (best no-reference SRCC 0.431) than on non-diffusion outputs (0.813). A pruned, distilled student model is presented as a benchmark use case, reaching 0.496 SRCC on diffusion outputs.
Nikita Kukuzei, Artem Borisov, Evgeney Bogatyrev et al.
In briefDetecting AI-generated music plagiarism as a version identification problem
SummaryThis paper tests whether music version identification methods built for human-to-human cover songs can detect plagiarism from AI-generated music. The authors introduce COPYCAT, a benchmark of 350,654 evaluation pairs built from real plagiarism cases and extended with generative re-synthesis and signal-processing obfuscation. They report that scalar distance thresholding fails under generative re-synthesis, while a supervised method using coordinate-wise embedding shifts raises overall F0.5 from 0.612 to 0.803.
Fotis Koutsikos, Ioannis Prokopiou, Spyridon Kantarelis et al.
In briefTAP: Trajectory-anchored pruning to make long-horizon AI agents cheaper to run
SummaryThe paper observes that long-horizon agent tasks require repeated model calls, making inference costly, and that existing pruning methods built for question answering severely degrade agent performance. It proposes Trajectory-Anchored Pruning (TAP), a structural pruning framework for reinforcement learning-trained agents that combines on-policy recovery anchored to teacher trajectories with iterative channel re-scoring. With 60% of FFN channels removed, the pruned 7B agents retain 99.2% and 88.0% of the dense model's task success on ALFWorld and WebShop, while GPU time per successful task drops by about 22% and 17%.
In briefOverLay++: A Dense-Overlap Layout-to-Image Generation Dataset
SummaryThis paper introduces OverLay++, a large-scale dataset of about 500K images with structurally complex scenes for layout-to-image generation. Each image averages 6.6 objects, giving 1.67 times the annotation density of existing datasets, and object captions are more than six times longer. The authors report that state-of-the-art methods trained on OverLay++ show consistent improvement across multiple benchmarks and faster convergence.
Shivansh Aggarwal, Shresth Grover, Divyansh Srivastava et al.
In briefJoint modeling of multivariate ordinal preferences with covariates, and its link to comparison models
SummaryThe paper studies multivariate ordinal data with covariates, such as movie ratings with user demographics or human feedback on LLM responses across several attributes. Standard methods model attributes separately or convert the data into pairwise or listwise comparisons, which coarsens the observed information. The authors propose a covariate-dependent consecutive ratio Markov random field model, show that pairwise and listwise comparison models arise as restrictions of it, and develop a maximum likelihood procedure that works despite an intractable normalizer.
In briefGUARD: Reliability-aware cleaning and labeling of noisy 3D scans of hard disk drives for robot disassembly
SummaryThis paper addresses scanning artifacts in hard disk drive point clouds that can look like real components but do not match the overall geometry. It proposes GUARD, which estimates per-point geometric uncertainty with a transformer and Gaussian Process model, and combines it with predictive entropy to filter unreliable points. On 2,745 real HDD point clouds, PointNet++ segmentation mean IoU improves from 0.7739 to 0.8318. Experiments on ShapeNetPart and ScanNet further examine robustness across corruption types and backbones.
In briefIs Chatting with an AI Really a Conversation? A Case for Seeing LLMs as Artifacts
SummaryThe paper introduces the "artifactual stance," which treats interaction with large language models as exchanges of candidate texts generated for utility rather than meaningful utterances from a speaker. The authors argue that no one stands behind the output, so the apparent conversation is effectively the user's own solo activity. On this view, philosophical questions about referring to "I" and "you," or whether systems can lie or make promises, rest on a false premise. The authors conclude that the key questions concern the normative grounds of design, adoption, authorization, and usage.
James Ravi Kirkpatrick, Alexandru Radulescu, Rachel Katharine Sterken
In briefMulti-label topic tagging of e-commerce user content by distilling LLMs into generative versus discriminative small models
SummaryThis paper studies assigning multiple topics to user-generated e-commerce content, such as product reviews and buyer-seller conversations, using small student models trained on labels from large language models. It compares 1B, 4B, and 8B generative models against discriminative baselines such as DeBERTa-V3 and ModernBERT. The authors report that discriminative models do better on structured product reviews, while the smallest 1B generative model outperforms them on complex multi-turn conversations. Generative models also remain more robust as the label set grows to 112 topics and under long-tail distributions.
In briefEDiS: a framework that sparsifies graphs for training graph neural networks using reusable edge-disjoint subgraphs
SummaryThis paper introduces EDiS, a sparse training framework for graph neural networks. It decomposes a graph once into cacheable edge-disjoint subgraphs, then recombines them each epoch under edge-budget constraints without re-extracting structure. Across 19 node classification benchmarks under the same edge budget, compared with 17 baselines, EDiS reports the highest mean score and the lowest average rank among ranked methods.
Sai Karthik Navuluru, Siddhartha Shankar Das, Franck Dernoncourt et al.
In briefLearning explicit many-to-many layer maps between models' activations
SummaryThis paper studies how layers and features correspond between two independently trained large language models. The authors propose MATCHA, which factors the activation alignment into an inspectable layer map and a layer-shared feature map, learned jointly from prompts. Across 42 pairs of seven models from three families, MATCHA reconstructs target activations more faithfully and substantially improves retrieval-based metrics over prior methods. The learned layer maps are broadly monotone in depth but many-to-many, and the alignments allow activation-space interventions such as steering vectors and probes to transfer between models.
Alina Sudakov, Guy Bar-Shalom, Fabrizio Frasca et al.
In briefUncertainty-aware defect detection that abstains from risky predictions in medical device quality control
SummaryThis paper examines whether accounting for epistemic uncertainty improves the reliability of automated defect detection in medical device manufacturing. The method uses knowledge-graph representations of manufacturing and device-report data and abstains from predictions whose estimated epistemic uncertainty exceeds a threshold. On 266,170 real FDA MAUDE reports, abstaining on 10% of samples reduced classification error by 48%.
Raham A. Butt, Marco Romanelli, Roche C. de Guzman
In briefA geometry-based capacity theory for finite-feature associative memory
SummaryThis paper develops a geometry-based capacity theory for exact-key retrieval in compressed finite-feature Hebbian associative memory. It splits retrieval interference into finite-feature noise that decreases with feature dimension and structural interference set by squared kernel overlap among stored keys. The theory gives fit-free retrieval predictions, a geometry-dependent capacity ceiling, and the feature budget needed for a target quality. Predicted curves closely match empirical results on synthetic, visual, and medical-image representations.
Jianhai Zhang, Donghao Zhang, Pattarawut Charatpangoon et al.
In briefMimicX: Refining Training Supervision for Video-Based Humanoid Motion Tracking Using Execution Feedback
SummaryThis paper presents MimicX, a framework that uses feedback from physics-based policy execution to improve how humanoid robots learn to track motion taken from human videos. It identifies difficult motion transitions and affected body regions, then adjusts tracking objectives and the reset curriculum, keeping only refinements verified through repeated rollouts. Across four core video tasks, the authors report a 25.7% average reduction in body-tracking error and a 255.6% increase in execution horizon relative to a fixed reference baseline.
In briefJustice Beyond Identity: Can LLMs Offer a View from Everywhere?
SummaryThe paper asks whether large language models could help form fairer moral judgments by drawing on diverse human identities, rather than excluding them as Rawls's original position attempts. The authors compare human, base model, and frontier or fine-tuned model judgments on classic moral dilemmas while systematically varying identity relationships and Rawlsian identity constraints. The paper closes by speculating whether humans would actually accept thoughtful advice from advanced AI systems.
In briefBeatFlow-ECG: a rectified-flow method that rebuilds ECG heart signals from wearable PPG and motion data
SummaryThe paper presents BeatFlow-ECG, a conditional rectified-flow model that reconstructs single-channel ECG from synchronized PPG and inertial measurement signals. It is evaluated on PPG-DaLiA and WESAD using leave-one-subject-out protocols. The authors report that it achieves the best results among the compared baselines across all reported waveform and beat-timing metrics.
In briefA Self-Pruning Transformer That Compresses Its Memory Cache Extremely Efficiently
SummaryLarge language models store a big key-value cache during inference, which limits efficient deployment. This paper proposes a unified framework for new decay mechanisms and an architecture called Universal Attention that keeps RoPE embeddings and Softmax attention while being trainable end to end. Its composite decay acts as an adaptive pruning criterion that removes tokens contributing least to attention. Experiments report 10x compression on natural language and synthetic data, and 25x compression at a length of 16k..
SummaryThis paper studies representation learning for financial markets, noting that prior evaluations often cover single predictive tasks and narrow datasets. The authors introduce Market-1T, a dataset of nearly one trillion U.S. equity observations from 2008 to 2025 at 1 Hz resolution, and a rigorous evaluation protocol. Comparing 18 encoder-training strategies, they find that encoders with similar predictive performance can organize market state very differently.
In briefLearning Jump-Diffusion Transition Kernels with Conditional Diffusion Models
SummaryThis paper studies how to learn transition kernels of time-homogeneous jump-diffusion processes using conditional diffusion models, so that new sample paths can be generated from N independent trajectories observed on a high-frequency discrete time grid. Theoretically, it provides non-asymptotic bounds on the conditional score estimation error and on the KL divergence between the laws of true and generated discretely observed paths. Numerically, the method is tested on synthetic data against the approach of Gao et al. (2025) and then applied to real-world data for a probabilistic forecasting task.
In briefQuadratic weak-to-strong generalization in random feature networks, analyzed with random matrix theory
SummaryThis paper studies weak-to-strong generalization, where a stronger student trained on labels from a weaker teacher can outperform the teacher, in two-layer random feature networks. Using random matrix theory, the authors derive deterministic equivalents for the population errors of an optimally trained teacher and a gradient-flow-trained student. For ReLU activation and a pure spherical harmonic target, under a Gaussian universality assumption, they show the student error scales as the square of the teacher error, a quadratic improvement that attains a general lower bound from Medvedev et al. (2025).
In briefCareful Judge: Safe and Efficient Human-AI Collaborative Decision Making
SummaryThe paper presents CARE (calibrated adaptive rectification and escalation), a pipeline that combines AI models with human reviewers to produce safe, human-aligned decisions. It continuously learns from human feedback so that the system can become more automated while requiring fewer human queries. CARE works with any black-box AI model and includes an adaptive calibration module that aims to control risk at every time step. Experiments on four safety-critical datasets covering driving, language, and robotics report 25-81% fewer human queries than baselines.
In briefORACLE: Aligning Constrained Learning with the Optimizer's Own Update
SummaryThis paper proposes checking constraint compatibility after the optimizer has computed its parameter update, rather than only modifying the objective or gradient beforehand. The ORACLE method linearizes several constraint types jointly at the realized step, builds the correction in the optimizer's own geometry, limits its strength, and applies it only after validation. Across eight PDE benchmarks and four optimizers, the authors report it matches or improves on the native optimizer in 94% of configurations.
In briefPost-processing LLM-extracted customer intents into concise, traceable, reusable units
SummaryThis paper presents a dataset-adaptive post-processing architecture for Customer Intent Extraction, which reduces high-recall LLM outputs into stable, traceable intent units. The method deduplicates and standardizes candidates, clusters them in a shared embedding space with keyword-based explainability, and uses constrained language-model aggregation to produce one intent per cluster. The authors also describe two downstream uses: inferring likely objectives for customers lacking direct evidence, and mapping intents to business outcomes via semantic retrieval.
Mahesh Viswanathan, Joan Rossello, Leticia Fernandes et al.
In briefWhen the Governor Becomes the Disturbance: Cost-Aware Backoff for Tool-Using Agents
SummaryThis paper studies how a supervisory governor can interfere with the tool-using agents it regulates. In a controlled file-recovery environment, higher regulatory intensity triggers imposed tool failures, and a cost-blind governor turns these into persistent blocking that prevents task completion. The authors compare it with a backoff rule that lowers intervention probability based on a moving average of induced events. Across tested settings, including 576 episodes with Gemini 2.5 Flash, backoff reduced blocking and improved completion.
In briefA graph neural network method that automatically generates and evaluates shared roadmaps for multi-robot path planning.
SummaryThis paper proposes a scalable heterogeneous graph neural network framework for generating and evaluating shared roadmaps used in multi-agent path planning. Waypoints, agent locations, and task locations are modeled as distinct node types, and the network is trained on occupation density maps collected from expert solver trajectories to identify key points and prune redundant nodes and edges. The authors report that the approach reduces runtime and graph size by at least 40% for dense roadmaps.
In briefShape-Bayes: Probabilistic Inference of Structured Shapes When Images Are Ambiguous
SummaryThe paper argues that deterministic regression struggles to infer structured shapes such as human faces when visual evidence is ambiguous or heavily occluded, sometimes producing incoherent shapes or generic averages. It introduces Shape-Bayes, a probabilistic framework that combines uncertainty-aware landmark prediction with a Transformer-based prior over a PCA shape space and a differentiable Bayesian solver. On 2D face shape regression under severe occlusion, the authors report up to about 34% absolute IDR improvement and up to 12.5% relative error reduction over state-of-the-art deterministic models.
Mani Kumar Tellamekala, Tosh Brown, Michel Valstar
In briefBeyond Explanation: Debugging Medical Imaging Models via Concept Intervention
SummaryThis paper presents a plug-and-play framework for interpreting and refining medical imaging models. It aligns a single-modality encoder with BioMedCLIP to build a Concept Bottleneck Model that allows concept-level interventions, separating causal concepts from spuriously correlated ones and generating counterfactual samples for fine-tuning. Evaluated on a Mayo Clinic ultrasound dataset and the CheXpert 5x200 chest X-ray dataset, the approach enables reliable model diagnosis and, in some cases, improves predictive performance through guided fine-tuning.
Samrajya Thapa, Daniel J. Quest, Timothy L. Kline et al.
In briefBEACON-SP: An Ontology-Grounded GraphRAG Framework for Clinical Suicide Risk Assessment
SummaryThis paper presents BEACON-SP, an ontology-grounded GraphRAG framework that supports clinicians in suicide prevention by combining patient knowledge graphs with ontology-guided retrieval for multi-hop reasoning. It integrates three established suicide-related theories and models into one ontology covering risk factors and social determinants. On a 1,500-query benchmark across 100 patients and 15 clinical categories, it outperformed a vector-based RAG baseline in completeness, clinical relevance, and evidence grounding, with a small factual accuracy gain. GraphRAG was preferred in 76.4% of paired criterion-level comparisons.
Kemal Davaslioglu, Nathan Conger, Sastry Kompella et al.
In briefSPIN: A Shadow Predictive Indexer That Cuts Indexing Cost for Sparse Attention in LLMs
SummaryThe paper proposes SPIN, a method that reduces the overhead of indexer-based sparse attention, which otherwise must score the entire KV cache at every decoding step. SPIN uses lightweight, history-based prediction to identify important KV blocks instead. Across long-context and agentic benchmarks it reaches 30-40% sparsity while preserving task quality. In end-to-end vLLM serving, it improves output throughput by up to 14.9% and reduces median inter-token latency by up to 13.2%.
In briefNot every call needs a frontier model: testing small language models per call site in a home-automation agent
SummaryThe paper argues that an agentic system's LLM calls differ widely in difficulty, yet one model usually serves all of them. The authors tested 9 models, from 0.8B parameters to a frontier hosted model, across five call sites of the open-source Wactorz home-automation framework, using two real Home Assistant installations (280 cases, 2,520 scored calls). The best local model was statistically indistinguishable from both hosted models at four of five sites, with code generation the only clear separation. Routing each site to its best local model reached 91.8%, compared with 95.4% at no per-call cost.
WhyPer-call-site evaluation shows capability ranking varies across sites and per-site routing cuts cost, backed by 2520 scored calls and live deployment validation.
Panagiotis Kasnesis, Christos Chatzigeorgiou, Lazaros Toumanidis et al.
In briefNeighborhood Smoothing for Calibration: a training-time method that uses neighbor structure in learned representations to make neural network confidence more accurate
SummaryModern neural networks are often overconfident and poorly calibrated. The authors propose graph smoothing as a training-time calibration principle, encouraging neighboring samples in representation space to have similar predictive distributions, and derive bounds linking this to calibration error. They introduce a graph-based regularizer that penalizes Jensen-Shannon divergence between neighbors' predictions. Experiments on standard benchmarks show improved predictive quality, complementary to temperature scaling, with the lowest NLL among train-time methods in seven of eight image and tabular settings after temperature scaling.
In briefPAIR: Linking Perception and Action in Vision-Language-Action Models
SummaryThis paper proposes PAIR, a framework that learns a shared intermediate representation between perception and action in vision-language-action (VLA) robot models. During training, a masked action autoencoder and a Bridge Module produce Bridge Tokens that are injected into the initial action tokens to start action-expert refinement; the autoencoder is removed at inference. On LIBERO-Plus, VLA-Adapter's success rate rises from 59.1% to 64.2%, and across seven real-world tasks OpenVLA-OFT's rises from 51.4% to 65.0%.
In briefPersonalize at Test Time: Learning User Preferences for Image Generation
SummaryThis paper addresses how to align diffusion model image outputs with individual users' preferences from limited feedback. The method compresses many visual attributes into 50 preference dimensions with an autoencoder, then represents each user's preferences as a low-dimensional weight vector fitted under the Bradley-Terry model. The learned personalized rewards guide generation at inference while the diffusion model stays frozen. On real-user data, the approach reaches about 77% held-out pairwise preference prediction accuracy and improves alignment with individual preferences.
In briefMaking Looped Language Models Compute Dynamically
SummaryLooped language models are parameter-efficient and could save memory and compute on easy tokens, but the open Ouro models do not achieve this in practice because each loop needs its own KV cache. The authors propose a simple "best-available" KV caching strategy that works out of the box, reporting up to 30% reductions in FLOPs and KV memory while keeping full-depth performance. They also modify the early-exit prior objective so tokens exit at heterogeneous depths based on difficulty.
Aayush Mishra, Arnau Padr\'es Masdemont, Victor Conchello Vendrell et al.
In briefWhose Memory Is It? Scope-Aware Commit Rules for Long-Term LLM Memory
SummaryThe paper argues that when an LLM agent stores intermediate reasoning, such as rejected plans, simulated tool results, or another speaker's beliefs, in long-term memory, these possibilities can later reappear as facts. The authors propose CASK (Causally Anchored Scoping Keys), a commit rule that preserves the stable relations indicating which world, branch, or speaker a proposition belongs to. In controlled long-conversation and tool-agent experiments, CASK improved memory admission and kept provisional content from contaminating later answers.
In briefStep-by-step scratchpads and staged curricula for small Transformers learning arithmetic
SummaryThis paper studies how compact Transformers with about 10.6M non-embedding parameters learn multi-step arithmetic from synthetic data written as step-by-step scratchpads. It reports that sequence padding starves gradients and that continuous packing fixes this, and that language pretraining is essential. A digit-by-digit long division scratchpad with a four-stage curriculum raised single-digit division accuracy from 4.0% to 86.7%, while multi-digit multiplication remained difficult and accuracy fell to 0.00% on unseen four-digit operands.
In briefLearning to Report Unsafe Tasks in a Multi-Agent Game
SummaryThis paper studies when a team of agents sharing a task reward should report and stop unsafe work. The authors analyze when audits let policy-gradient updates reach universal reporting, and show that a fully shared policy can be pulled toward silence even when reporting is optimal. Experiments with PPO on 24 witness graphs find that grouping witnesses into separate policies substantially reduces unsafe task completion.
In briefPhysEvo: A Framework Letting a Frozen Model Improve Robot Manipulation Through Its Own Actions
SummaryPhysEvo is a framework for physical recursive self-improvement in robot manipulation, built around one frozen model without weight updates or a separately trained action policy. A task agent runs robot tasks, while a meta-agent diagnoses failures from the trajectories, revises tools and skills, and tests corrections, including improvements to its own diagnostic tools. Across 42 RoboDojo tasks with held-out layouts, it reaches 62.00% success, versus 47.17% for the compared one-shot Astra agent. On real-world AgileX PiPER tasks, it reaches 84.00% success across 25 trials.
In briefPhoneme-guided initialization to improve LLM-based speech recognition with scarce data
SummaryThis paper addresses the weak performance of speech LLMs for speech recognition when paired speech-text data is limited. The proposed phoneme-guided initialization pre-trains the audio encoder on speech-to-phoneme and the LLM on phoneme-to-grapheme tasks, then fine-tunes the connected model end-to-end. Experiments on Japanese, Chinese, Tatar, and Urdu show results that match or exceed a cascaded baseline and an end-to-end model without this initialization.
In briefVerify Less, Evolve More: idea-level critics that make ML evolving agents verify efficiently
SummaryThis paper addresses the high computational cost of empirically verifying machine learning modifications proposed by self-evolving agents for AI for machine learning (AI4ML). The authors train idea-level critic models, via supervised fine-tuning on critiques synthesized by Gemini-3.1-Pro followed by GRPO, to predict whether a proposed ML change will improve the current solution. The critics outperform Gemini-3.1-Pro in static idea evaluation and improve results in inference-time evolution, continual learning, and policy training.
In briefREFIT: Label-free fixing and testing of wearable sensor placement shifts
SummaryREFIT is an input calibration method for frozen activity-recognition models when wearable inertial sensors are worn differently at deployment, such as a watch moved to the other wrist or a strap sensor put on reversed. It describes such shifts with families of axis transforms and fits them so simple statistics match the training data, without labels or retraining. Experiments on real and simulated re-attachment show it outperforms label-free test-time adaptation methods and restores most lost accuracy. It also provides a label-free accuracy estimate used to decide when the user should re-wear the sensor.
In briefHow Much Synthetic Data Activation Probes Need, and Why Coverage Matters
SummaryThis paper studies how many synthetic training samples activation probes need to monitor deployed language models. It traces learning curves for three concepts (high-stakes situations, harmful replies, and instruction-violating replies) across fourteen held-out evaluation distributions and four probe models. The authors find the required data depends on the concept, with instruction probes needing several times more samples, and that the depth needed is set by coverage rather than per-kind difficulty.
In briefLASER: a latent-space adjoint matching method for offline RL that stays close to the dataset while optimizing policies.
SummaryOffline reinforcement learning optimizes policies from fixed datasets, but executing out-of-distribution actions remains a key risk. The paper introduces LASER, which performs entropy-regularized RL in the latent space of a flow-matching behavior-cloning policy using latent-space adjoint matching, avoiding backpropagation through time. Across 40 OGBench tasks, LASER reports state-of-the-art results using fixed hyperparameters, outperforming baselines that were tuned per task and dataset.
In briefA neural-network method for jointly analyzing multimodal TAIGA experiment data
SummaryThis paper proposes a method using autoencoders to extract essential features from multimodal data recorded simultaneously by several installations. It applies the approach to data from the TAIGA experiment, which is currently analyzed separately for each installation. Based on Monte Carlo simulation, the authors show the method enables effective multimodal data analysis and could also be used at other experimental complexes.
Alexander Kryukov, Julia Dubenskaya, Elena Fedotova et al.
