In briefSNACK: A sequence-based notation for generating random graphs with language models
SummaryThis paper introduces SNACK, a string representation that lets language models generate graphs with discrete node and edge features. It linearizes the nonzero entries of the lower-triangular adjacency matrix and masks invalid tokens to enforce constraints such as chemical valence, guaranteeing fully valid outputs. Experiments on graph generation and molecular benchmarks report competitive sample quality and substantially higher GFlowNet training throughput than graph-based generation.
In briefLabOPBench: A Benchmark for Evidence-Based Lab Decisions Using Text and Images
SummaryThis paper introduces LabOPBench, a benchmark of 756 expert-written, independently reviewed questions in organic synthesis and materials science. Each question pairs laboratory evidence and procedural context with an open-ended judgment, scored with question-specific rubrics. Eight multimodal models scored between 28.8% and 45.8% overall. On a 100-question subset, the strongest models beat undergraduates on average but stayed below PhD participants, and error analysis found Evidence Misreading and Evidence-Decision Conflict to be the most common failures.
In briefCompressive imaging reconstruction using a simulated-annealing Ising solver
SummaryThe paper proposes a physics-driven framework that formulates imaging inverse problems as Ising models and recovers objects through the self-evolution of Ising machines toward a Hamiltonian's ground state. The authors demonstrate a simulated-annealing-based Ising solver on single-pixel imaging and a lensless diffuser camera at sampling rates from 1% to 25%. With foreground-sparsity and smoothness constraints added, reconstruction quality is comparable to or better than mainstream algorithms. The approach aims to reduce the electronic computational burden of inverse reconstruction.
In briefQ-MAGIC: Autoregressive generation of quad-only meshes using singularity-aware serialization
SummaryThis paper presents Q-MAGIC, an autoregressive Transformer that generates quad-only meshes by predicting tokens for one face at a time. It introduces a serialization scheme for quad faces that starts wiring from singular points to shorten sequences, along with a circular relative offset tokenization that encodes geometry mainly through local offsets. The authors report that compared with state-of-the-art methods, Q-MAGIC yields more compact sequences and better geometric fidelity and mesh quality.
In briefFrom Global to Tail Convergence: Stable-Process Generative Models
SummaryThis paper studies score-based generative models driven by α-stable Lévy processes and gives a non-asymptotic convergence analysis for compactly supported and power-law targets. In p-Wasserstein distance, it quantifies the effects of early stopping, initialization, score approximation, and time discretization, and shows that the initialization error decays exponentially. Compared with Brownian dynamics, global Wasserstein convergence alone does not separate the two models, but their tails differ sharply: stable dynamics yields power-law tails with index α at any positive generation time, while Brownian dynamics stays light-tailed. For a target with tail index ν, asymptotically correct extreme-quantile scaling is possible precisely when α equals ν, offering a principled explanation for stable models' ability to capture extreme events.
WhyProvides non-asymptotic convergence analysis for alpha-stable generative dynamics and tail-index matching versus Brownian models; strong theory, impact mostly within the subfield.
Ianis Hammani (ENS Paris Saclay, LMO), Claire Boyer (LMO et al.
In briefWhen Does the Dependency Graph Help Code-to-Architecture Mapping?
SummaryThis paper addresses mapping source files to architectural modules, a task relevant to architectural technical debt. The authors frame it as semi-supervised node classification on a directed multi-relational dependency graph, comparing GCN, GAT, and a graph-free MLP against a Naive Bayes reference. Across 100 runs on ten open-source Java and C/C++ systems, all three neural models outperform the reference, and on Chrome (18,343 files) GAT and GCN exceed MLP by 22 and 19 macro-F1 points.
Rakhshanda Jabeen, Morgan Ericsson, Jonas Nordqvist et al.
In briefLooped transformers: learning to compute does not mean retaining or reusing answers
SummaryThe paper studies overthinking in looped transformers, where running more loops than during training turns correct answers into wrong ones, and separates failed computation, lost retention, and states that later computation cannot reuse. On in-context Mealy-machine execution, models trained at 32 loops kept only 22% of their answers at 64 loops. Changing the training supervision raised retention substantially, but holding intermediate states fixed showed that retained states are not reliably reusable by later computation.
In briefLikelihood Is Not Authorship: Measuring Human Contribution in AI-Assisted Writing
SummaryThe paper shows that a common method of estimating human contribution to AI-assisted text, based on how well a person's prompt predicts the final text, can be easily manipulated. The authors introduce SPECIFICITY PADDING, where asking a model to copy random IDs from the prompt makes a title-only request appear more human, and show that changing only which IDs the evaluator sees shifts scores substantially for the same text. They report this effect across news and scientific abstracts and LLMs ranging from 3.8B to 753B parameters. The authors conclude that likelihood-based scores cannot reliably establish human contribution without evidence of what the person added and how the text was produced.
WhyShows prompt-predictiveness contribution scores are easily manipulated, gives an identity and lower bound, with broad experiments across models and datasets.
In briefPhysics-Shaped, Not Physical: Auditing Physics-Guided Underwater Image Restoration
SummaryThis paper audits PhysGuide-Net, a physics-guided underwater image restoration network, using controlled experiments to test whether its physics module actually helps. With full training data the module gives no zero-shot gain, and its predicted transmission maps do not behave like a true physical quantity. The authors conclude that any benefit comes from a physics-shaped inductive bias rather than from the physics itself.
In briefPRISMATIC: A multi-agent system that checks drug prescriptions for conflicts in clinical decisions
SummaryThis paper introduces PRISMATIC, a multi-agent system in which specialized agents representing different clinical roles collaborate on a blackboard architecture to structure patient data from clinical notes and iteratively generate drug regimens. Two adaptive mechanisms govern coordination: Dynamic Self-Updating Weight Adjustment, which adjusts agents' influence over time based on historical outcomes, and Discrepancy Feedback Precise Calibration, which resolves disagreements for the current patient. Experiments on MIMIC-IV patient records report that PRISMATIC significantly reduces drug conflicts, improves prescription coverage, and enhances treatment consistency compared with strong LLM baselines.
In briefCAST: History-adaptive spectral clipping to stabilize LLM pretraining gradients
SummaryThis paper introduces CAST, a history-adaptive extension of soft spectral clipping (SSC) that adjusts clipping scales and thresholds using past optimizer statistics. The method is applied to pretraining a mixture-of-experts model with 19.08B total parameters on 83.89B tokens. CAST achieves lower training and mean validation losses than nine gradient-processing baselines and produces the fewest relative gradient spikes. It reaches the best baseline's final smoothed training loss in 7.2% fewer steps.
In briefQwen-MT-Lumen at WMT2026: Instruction-Specialized Translation via Quality-Aware Training and On-Policy Distillation
SummaryThis paper presents Lumen, an open-source translation model submitted to the WMT 2026 General Machine Translation Shared Task, focused on translating while following fine-grained instructions on terminology, style, and format. It uses a four-stage pipeline combining pre-training and fine-tuning, online preference optimization with an LLM judge, on-policy distillation from a larger teacher, and structure-aware inference. Across 23 translation directions, Lumen reaches an overall mean score of 4.60 on a 1–5 LLM-judge rubric, with performance reported as comparable to leading proprietary systems despite a 14B backbone.
In briefBeyond Reachability: A Relation-Semantic Framework for Biomedical Multi-hop QA
SummaryThis paper addresses biomedical multi-hop reasoning over knowledge graphs, where correct answers are retrieved but often ranked poorly and their supporting evidence may not match the required relation composition, which the authors call the reachable-but-invalid problem. The proposed ReSemReason framework shares a question-induced distribution over typed, directed relation schemas between graph search and evidence selection, using relation-conditioned attention and an absolute-support score. Reported experiments on BioHopR, PrimeKGQA, and MedReason show improvements in answer ranking and answer-set quality.
In briefDiscriminative Score Function: Reusing Pretrained Classifiers and Detectors as Image Generators
SummaryThe paper introduces the Discriminative Score Function (DSF), which extracts an image-space generative field from the loss-gradient geometry of a pretrained discriminative model without retraining. DSF generates unconditional images from noise and supports conditional guidance, editing, inpainting, and explanation through the same field. It is applied to ResNet, DETR, and DINOv2 models. DSF-generated samples also provide zero-shot quantization calibration when real data is unavailable, preserving feature geometry better than data-free baselines.
In briefHyperLoop: Using Hyperspectral Imaging to Improve Loop Closure in Planetary Rover Mapping
SummaryThis paper presents HyperLoop, a framework that adds hyperspectral imaging to the loop closure module of a 3D Gaussian Splatting SLAM system, aiming to help planetary rovers recognize previously visited locations on repetitive terrain. Two hyperspectral descriptor types are introduced and fused with RGB descriptors, then evaluated on a custom dataset from DLR's Moon-Mars Outdoor Test Site. The best hyperspectral fusion achieves 5.04 cm ATE RMSE, nearly identical to SIFT-VLAD alone (4.95 cm), so the gain comes mainly from SIFT-VLAD. The authors conclude that the 600-860 nm range offers limited discriminability for rock terrain and suggest SWIR sensors for future missions.
Annie Bhalla, Dennis Rotondi, Riccardo Giubilato et al.
In briefDoes Reasoning Language Style Matter? Classical Chinese as a Compact Reasoning Format for Chinese LLMs
SummaryThis paper proposes Wenyan Thinking, which has Chinese reasoning models write intermediate reasoning in Classical Chinese while still giving a short direct answer. The authors rewrite modern Chinese reasoning traces into Classical Chinese and fine-tune Qwen3-4B-Thinking-2507 on them, comparing against a compact modern Chinese control trained with the same data, teacher, base model, and recipe. Under constrained decoding on several Chinese benchmarks, the method substantially reduces reasoning length and missing-answer rates. Wenyan outperforms the control on the primary benchmark and stays competitive on the others.
In briefProtModernBERT: A Compute-Efficient Protein Language Model
SummaryThis paper introduces ProtModernBERT, a modernized BERT-style encoder pre-trained on amino acid sequences, aimed at low-compute research settings. It is compared with ProtBert, Amplify, and ESM2 on supervised structural tasks and the ProteinGym zero-shot functional benchmark. The authors report that it matches or closely tracks Amplify-120M and ProtBert while reducing pre-training FLOPs by up to 80-90%.
Alberto Mart\'inez, Natalia Flechas Manrique, Lloren\c{c} Espinosa-Portal\'es et al.
In briefMakSafe: a spatiotemporal framework for formally checking road vehicle safety rules
SummaryThis paper presents MakSafe, a formal framework that specifies and checks road vehicle safety rules using semantic zones, action-based forecasting, and trace-tree reasoning, without uniform time discretization. The authors evaluate it with a trace-dependent case study and a scalability benchmark varying traffic density, horizon, and branch factor. The framework gives branch-specific safe or unsafe verdicts from the same initial scene, and a single rule check runs well below a millisecond, with longer horizons slowing the time-to-collision check.
Maksym Labzhaniia (UniCA, Laboratoire I3S - COMRED, KAIROS) et al.
In briefEstimating 3-D Object Pose from 2-D Point and Line Matches
SummaryThis paper presents a method for optimally estimating the rotation and translation between a camera and a 3-D object using point and/or line correspondences. The authors first define an error function and then show how to minimize it, representing rotation and translation with a dual number quaternion so the function is quadratic. A trust-region optimization method is compared favourably against Newton's method, the Faugeras-Toscani linear method, and Levenberg-Marquardt. Experiments show robustness to image noise and matching errors.
Thai Quynh Phong, Radu Horaud, Adnan Yassine et al.
In briefExact cut count needed for equal-length fair cake division
SummaryThis paper studies proportional cake cutting where each of n agents must get a piece of length exactly 1/n worth at least 1/n to them. The authors prove the worst-case number of cuts needed is exactly 2n-2 for every n, improving an earlier quadratic bound. The work also covers envy-freeness, computational query protocols, and results when two-knife queries are allowed.
WhyDetermines exact worst-case cut count 2n-2, improving prior quadratic bound, via constrained necklace-splitting and Hall's theorem proof.
Yasushi Kawase, Bodhayan Roy, Mohammad Azharuddin Sanpui
In briefRL Teaches VLMs Where to Attend but Not How to Perceive: study of why RL helps vision-language agents
SummaryThis paper investigates why reinforcement learning improves vision-language model agents across four environments. The authors design grounded visual question answering tasks and find that vision-encoder probe accuracy stays relatively stable while LLM-layer probe accuracy improves, with LLM layers attending more to visual tokens during training. A lightweight fixed attention bias applied without parameter updates partially restores these VQA gains on an untrained model, suggesting better use of visual information is a key mechanism.
In briefWhat Happens to the Goal? Distinct Failures of Policy Alignment in Goal Misgeneralization
SummaryThis paper studies goal misgeneralization, where an agent acts competently in unfamiliar situations but does not pursue its intended goal. Across three Procgen environments and five robot tasks, the authors identify three forms of policy-goal misalignment: functional target substitution, a policy shortcut, and cue competition between intended and proxy goals. They find the goal is lost at different depths in the network, and the recovery method that works depends on where that loss occurs.
In briefDiagnosing World-Space Human Recovery Errors Under Fast Camera Motion
SummaryThe paper introduces the K-Pop stress test, a set of broadcast dance videos with fast, large camera motion and dense multi-person choreography. Existing methods yield plausible per-frame meshes, but their world-space trajectories drift far from performers' stage positions. The authors attribute this to camera and human estimates being recovered at mismatched scales, with drift growing as camera displacement increases. They propose a closed-form correction that estimates the mismatch and applies one scale to all performers, requiring no retraining.
In briefReducing video hallucinations with attention-guided representation perturbation
SummaryVideo large language models often give answers that the input video does not support. The paper proposes ARP-CD, a training-free contrastive decoding framework that builds negative inputs in representation space by using attention to visual tokens to guide bounded gradient updates that weaken visual reliance. Experiments on four Video LLMs and four benchmarks show it outperforms representative training-free baselines in most settings, reducing hallucinations while improving general video understanding.
In briefGenerative World Simulation: A Framework for Persistent, Interactive Shared Worlds
SummaryThis paper presents a Generative World Simulation framework that links persistent shared world state with continuous first-person audiovisual generation. DAO, a computable world engine, maintains characters, objects, spaces and their relationships, advancing the world by interaction rules and world time. JING, an Egocentric Interactive Experience Model, uses conditions derived from DAO to generate each observer's audiovisual experience. Users explore environments, manipulate objects and communicate with characters, and their actions shape both the shared state and later experiences.
XGEN Team: Guohao Cai, Haoxuan Che, Frank Fan et al.
In briefTimeRSI: A Self-Improving Agent System for Deploying Time-Series Forecasting Models
SummaryThe paper presents TimeRSI, a system that wraps frozen time-series forecasting models with a self-improving agent. A coding agent turns task descriptions into pipeline changes, tests them on historical forecasting windows, and uses observed errors to guide later revisions. At deployment, an adaptive selector picks among the resulting pipelines based on completed forecasts under similar conditions. Across 198 tasks on TIME and fev-bench, TimeRSI outperforms all submissions on six point and probabilistic metrics, and controlled comparisons favor feedback-guided revision over non-iterative search.
WhyFeedback-guided agentic pipeline revision with adaptive selection is a substantive improvement; validation on 198 tasks gives fairly strong evidence and useful relevance to time-series forecasting.
In briefRecognizing dialogue ambiguity by estimating the set of objects consistent with multi-turn context
SummaryThe paper argues that ambiguous utterances in real interactions often need multiple turns of context, and may refer to more than one object in a visual scene. The authors propose Candidate Set Visual Grounding (CSVG), a task that estimates the set of objects consistent with the dialogue context, and introduce the CandVG benchmark to track how this set changes across turns. Evaluations of vision-language models show that stronger visual grounding does not necessarily improve CSVG, as models struggle to predict the full candidate set. The authors report that instructions designed for CSVG partially mitigate this limitation in large-scale models.
In briefONCship: A Large-Scale Underwater Sound Dataset for Recognizing Ships
SummaryThis paper introduces ONCship, a large underwater acoustic dataset of 4,521 hours of recordings from Ocean Networks Canada hydrophones, paired with AIS ship data. The authors test it on self-supervised representation learning, vessel-type classification, and MMSI-level vessel classification. Joint pre-training with AudioSet reaches 62.45% accuracy on DeepShip via linear probing, and using ONCship as auxiliary training data improves vessel-type classification.
In briefMeasuring how well concepts can be passed between language models
SummaryThe paper introduces two measurement tools: a prompt optimization algorithm that beats existing methods by up to 14% on benchmarks, and a transmission measure where a receiver tries to recover a sender's prompt. Using these, the authors study 1.2k concepts across eight domains as model capabilities scale. They find concepts that can be fully articulated, concepts that future models might articulate, and concepts facing articulability limits, the last tied to identity and social-group awareness.
Alexander Spangher, Daniel E. Ho, Diyi Yang et al.
In briefLearning individual research taste with hindsight distillation
SummaryThe paper models scientific taste as each researcher's mapping from criterion-level assessments to preferences over research proposals. It introduces Privileged Hindsight Distillation, which uses later-emerging evidence during training but only proposal-visible information at inference. Evaluated mainly in quantum research, the method captures distinct tastes from just 15 pairwise comparisons per researcher, and in an interactive study personalized rankings surfaced proposals with greater later user engagement.
In briefUncertainty-driven self-play for more robust multi-agent driving policies
SummaryThe paper proposes a self-play method in which half of the driving agents are rewarded for increasing their neighbors' epistemic uncertainty, while the other half learn to drive safely. Exploring agents use uncertainty as a signal to find the policy's weaknesses, creating a challenging training curriculum without manual scenario design. The authors report better zero-shot transfer to nuPlan Val14 and Argoverse 2 than standard self-play, and show that epistemic uncertainty can serve as a difficulty indicator.
Sajad Marvi, Julian Schmidt, Alessandro Canevaro et al.
In briefNew Model, Old Habits: How Swapping LLMs Carries Over Behavior in Multi-Session Agents
SummaryThis paper studies what happens when a multi-session LLM agent switches to a different model and the new model receives the earlier history. The authors find behavior transfer: the recipient tends to adopt the reasoning and action patterns of the model that generated the history, most strongly when raw history is passed and much less with summarization. They propose DASH, a direction- and authorship-aware summarization method, which outperforms raw-history transfer and naive summarization in repeated-swap experiments.
In briefBeyond State Capacity: Why Retrieval Breaks in Mamba
SummaryThis paper studies why Mamba models underperform Transformers on in-context retrieval, even when the query comes before the context. The authors show that Mamba's selection gates can change model output, but the relevance signal they depend on emerges only gradually with depth. In retrieval-critical layers the gates are not yet aligned with relevance, so retrieval often fails even when state capacity is sufficient.
In briefSceneAE: Generating and Rebuilding Scenes in a 3D-Anchored Latent Space
SummaryThe paper proposes SceneAE, which represents a scene as a sparse latent point cloud where each latent pairs an explicit 3D position with implicit content features. Unlike video-VAE latents tied to particular viewpoints, this representation describes the scene itself, so it can be projected into RGB-D views for any camera without regenerating video. Generation first produces camera-controlled 3D geometry and then content conditioned on it. The authors report better generation quality, lower computational cost, and improved camera controllability.
In briefGraph additive model separating features, topology and interactions (AMIGA)
SummaryThis paper introduces AMIGA, an interpretable graph learning model. It represents graph topology with a vocabulary of mined motifs and splits each node's contribution into three additive channels: feature, structure, and feature-structure interaction. Across ten graph classification and regression benchmarks, AMIGA outperforms GNAN on most datasets and narrows the gap with standard black-box GNNs.
Clara Alvarez-Rodriguez, Steve Azzolin, Vincenzo Marco De Luca et al.
In briefRouteWeaver: Unifying Mode Selection and Execution in LLM Routing
SummaryThis paper presents RouteWeaver, a framework that combines single-round, multi-round, and agentic routing modes within one LLM policy, learning both which mode to use and which models to select within it. It introduces COMET-GRPO, a training method that separates the trajectory-level advantage into signals for mode selection and within-mode execution, along with progressive exploration. Evaluated across mathematics, coding, knowledge, logical reasoning, and factual recall, it reports 61.4% average accuracy, an 11.03% relative gain over the strongest baseline.
In briefBANGLAPLAN: A Bangla Benchmark for Constraint-Based Planning and Decision-Making
SummaryThis paper introduces BANGLAPLAN, a benchmark of 1,000 Bangla problems that test whether language models can plan or make decisions under constraints on time, cost, capacity, and order, or recognize when no feasible solution exists. The problems span ten task families grounded in daily life and institutions in Bangladesh, across four language registers, with answers in a common structured JSON format. The authors evaluate 11 model configurations and report a large performance gap between models.
Md Rayhanul Masud, S M Maksudul Alam, Mohammad Jaminur Islam
In briefTime-JEPA: Adding Physical Time Series to Video World Models for Better Physical Understanding
SummaryThe paper proposes Time-JEPA, which fine-tunes a video world model using low-dimensional, temporally aligned physical time series as auxiliary supervision. Across autonomous driving, robot manipulation, and human motion, it improves current-state decoding, multi-horizon future-state prediction, and kinematic reconstruction compared with video-only fine-tuned controls. The authors also report aggregate gains on a second video backbone.
In briefRuntime monitoring that alerts when supermartingale safety certificates fail in deployed systems.
SummarySupermartingale certificates give formal guarantees about stochastic systems, but these rely on a mathematical model that may not match the real deployed system. This paper formulates runtime monitoring for such certificates: a monitor watches a single trajectory and raises an alarm when statistical evidence shows a violation. The authors build the monitor with betting-based sequential tests that bound the false alarm probability at a user-chosen level, and add strategies that exploit spatial regularity to improve detection. The approach is demonstrated on supermartingale certificate benchmarks from the literature.
Mirco Giacobbe, Thomas A. Henzinger, Konstantin Kueffner et al.
In briefReGRA: Residual-Gradient Adaptation for Complementary Low-Rank Fine-Tuning
SummaryThis paper proposes ReGRA, a two-stage method for fine-tuning large models with LoRA under a fixed rank budget. After the first stage, it removes calibration-set task gradients' projections onto the first update's subspace and uses the remaining residual directions to initialize a second adapter, with an orthogonality regularizer encouraging complementary updates. The authors report that ReGRA outperforms alternatives with the same total rank and matches LoRA with twice the rank. Both updates can be merged into pretrained weights, so there is no extra inference overhead.
In briefRuleTrek: A Benchmark for Rule Reasoning by AI Agents in Complex Scenarios
SummaryLarge language models often struggle to find and correctly apply domain-specific rules when the relevant scenario information is spread across multiple files. This paper introduces RuleTrek, a multi-domain benchmark built from structured rule graphs and sub-scenarios, and tests three abilities: rule identification, rule deduction, and end-to-end rule reasoning. Experiments on 11 frontier models show the best rule identification and deduction scores are only 62.47% and 60.43%, with end-to-end scores between 26.97% and 55.29%. The code and dataset are publicly available.
In briefSharing Word Indices, Not Vectors: Training-Free Fact Exchange Between Language Models
SummaryThe paper introduces SYNC (Symbol-Indexed Neural Communication), where a sender picks words from a shared list and sends their indices, and a receiver writes the described facts into its hidden states using its own word directions. It requires no fine-tuning or learned translator. On new scenes, four-answer success reaches 72.0-98.3% across nine receivers from five model families.
In briefC3: Confidence-guided credit correction for process supervision in multi-turn RL
SummaryOutcome-only reinforcement learning can reinforce undesirable local behaviors inside successful multi-turn trajectories. The authors propose C3, a plug-in method that identifies spans tied to such behaviors and selectively corrects their positive credit after advantage estimation. The correction strength is adaptively controlled by the behavior policy's confidence, while task reward and original credit estimation stay unchanged. The paper reports improved performance across several multi-turn benchmarks for both critic-based and critic-free RL methods.
In briefReFormBench: A benchmark for how reliably LLMs correct response formats across instruction changes
SummaryThe paper introduces ReFormBench, which uses each model's unconstrained answer to a question as its formatting baseline and measures whether models can correct that format when instructions vary. It includes 5,000 three-constraint test instances from 910 questions covering 66 formatting features. Across ten models, EMCR ranges from 34.04% to 64.72%, and compliance and correction scores produce different model rankings.
In briefAnnealed Lévy Diffusion: a generative model using Gaussian surrogate reverse dynamics
SummaryThe paper proposes Annealed Lévy Diffusion (ALD), which varies the stability index over time so denoising combines heavy-tailed exploration with Gaussian refinement. Stable noise is represented as a scalar Gaussian scale mixture, allowing standard Gaussian noise-prediction training with one extra scale input. Two samplers are studied, ALD-renoise and ALD-posterior, both reusing the same trained predictor. The method is evaluated on four 2D benchmarks, an imbalanced nine-mode mixture, and MNIST, CelebA64 and CIFAR-10LT experiments.
Yuling Jiao, Lican Kang, Jerry Zhijian Yang et al.
In briefPPAD-Completeness of Contiguous Envy-Free Cake Cutting with Additive Valuations
SummaryThis paper proves that computing a complete contiguous ε-envy-free cake allocation is PPAD-complete for normalized rational piecewise-constant additive valuations. The hardness holds even when ε is given in unary and each density takes only two shared values with at most six breakpoints. Exact contiguous envy-free allocation is also shown to be PPAD-complete under the same restrictions. The authors further show that computing contiguous EFc allocations of indivisible goods on a path is PPAD-complete for every fixed c≥1.
WhyProves PPAD-completeness of contiguous envy-free cake cutting even under strong restrictions like two-valued piecewise-constant densities, a rigorous and substantial complexity result.
Yasushi Kawase, Bodhayan Roy, Mohammad Azharuddin Sanpui
In briefFunction-Space Temporal Self-Distillation for Policy Optimization (FSPO): a reinforcement learning method that stabilizes training with stored past output distributions
SummaryThis paper addresses the coarse supervision of scalar outcome rewards in Reinforcement Learning with Verifiable Rewards (RLVR) for language model reasoning. It proposes Function-Space Policy Optimization (FSPO), which stores sparse historical next-token distributions as fixed temporal targets without a persistent teacher model, and adds privileged disagreement correction, reliability-adaptive grounding, and quality- and coverage-aware temporal memory. The authors report the highest average accuracy on reasoning and tool-use benchmarks, about 4 points above the strongest baseline, and peak memory reduced from about 70 GB to 56 GB on Olmo3-7B-Instruct.
In briefInterlock: Moving LLM Agents from Locally Correct Steps to Full Task Completion
SummaryThis paper addresses a problem in long-horizon LLM agents: they can satisfy individual requirements yet leave tasks incomplete overall. The authors propose Interlock, which represents each task as a typed completion graph capturing dependencies among evidence, decisions, and downstream effects, then compiles it into a resettable workspace, a public request, and an evaluation rubric. Fine-tuning two Qwen3.6 backbones on their own accepted trajectories improved pass rates or scores on AutomationBench, JobBench, and Harvey LAB.
In briefBeyond Distribution Matching: Selective Margin Preservation in Supervised Fine-Tuning
SummaryThe paper proposes SRP-SFT, a soft margin objective for supervised fine-tuning that applies a hinge loss to reference-scaled lower bounds for pairs among the reference's top-K non-target tokens. It aims to preserve selected local margin relations without freezing the full reference distribution. In one-epoch experiments on Qwen3-4B-Base with UltraFeedback, the method improves MATH-500 and code-generation scores over CE and LP-SFT across all three training seeds.
In briefMemory-enhanced latent world models for long-horizon planning
SummaryThis paper introduces a latent world model that adds a content-addressable memory bank to JEPA-style prediction. It learns high-level representations across multiple temporal scales in a shared latent space and uses adaptive layer normalization for conditioning. Two complementary losses are designed to prevent representational collapse and encourage effective memory use. Experiments on PushT, Tworoom, Reacher, and OGBCube report state-of-the-art results with a lightweight architecture on long-horizon planning tasks.
In briefMicroSDF: A neural method for 3D reconstruction of mixed reflective surfaces using microfacet theory
SummaryThe paper proposes MicroSDF, a neural implicit framework that models the geometry and reflectance of both reflective and non-reflective surfaces using microfacet theory. It introduces a microfacet-guided geometry model with a normal consistency loss, a dual-branch color model, and a detection-guided color blending strategy, combined with multi-stage optimization. The authors report state-of-the-art results on the DTU, Shiny Blender, Ref-NeRF, and DeepVoxels datasets.
In briefMCAIR: Catching and Fixing Agents That Misreport How They Completed Tasks
SummaryLLM-based agents can finish multi-step tasks but sometimes report results that do not match what they actually did, such as switching to other sources without disclosing it. The paper proposes MCAIR, a runtime framework that trains a risk-aware probe on the agent's internal representations to detect such mismatches during execution, and a correction loop that asks the agent to review its history. The authors also release ACUD, a dataset with step-level traces, representations, and consistency labels. Experiments reportedly show MCAIR outperforms LLM-as-a-judge baselines and reduces these unfaithful completions.
In briefRemember, Revisit, Revise: Scaffold Memory for Long-Horizon 3D Geometry
SummaryThis paper presents SMORE, a framework that gives frozen geometry foundation models a revisable external memory for estimating camera poses and 3D structure from long video sequences. It stores past observations, retrieves them, and revises them when new evidence shows inconsistencies. Across eight models and five datasets, SMORE improves pose estimation, multi-view depth, point-cloud reconstruction, and SLAM results.
In briefLearning When to Trust Medical Knowledge for Safer Clinical AI Reasoning
SummaryClinical reasoning language models rely on external medical knowledge, but retrieved evidence and reasoning chains can be outdated, poisoned, or hallucinated, sometimes turning correct answers into wrong ones. The authors build a medical knowledge-poisoning benchmark and find that no fixed knowledge-use strategy is robust across settings. They propose Med-TrustRouter, a lightweight controller that decides per instance whether to use external knowledge while keeping the solver frozen. It matches the strongest competitor on macro accuracy and substantially improves worst-case accuracy across seven medical reasoning settings.</summary>
In briefMeasuring and reducing unusable replies from LLM agents
SummaryThis paper introduces reply usability, which asks whether a requester can understand and use the reply an LLM agent returns. A reader model that takes the requester's perspective scored 480 replies from eight models, and 54% failed, with the best model still failing 30%. Failures mostly involved unseen material or coined labels. The authors report that stating explicit reply requirements, or having a secondary model rewrite the reply, lowered the failure rate.
In briefHACHIMI-II: A more faithful simulated student built by modeling the learner's behavior
SummaryThis paper presents HACHIMI-II, a method for building simulated students so that tutoring systems and teaching strategies can be tested before real learners are available. Instead of asking a language model to role-play a student, it builds a model from the learner's history, infers the learner's current state, and uses that to predict the next action. Across three item-response datasets and two dialogue datasets, it outperforms the strongest baseline in matching real student ability and reproducing real student errors.
WhyRecasts student simulation as learner behavior prediction with explicit learner state; strong gains over baselines across datasets, supported by ablations.
In briefMatching lower bounds for zeroth-order nonconvex optimization
SummaryThis paper studies theoretical lower bounds for zeroth-order nonconvex optimization, which estimates gradients from function values. It answers an open question about whether the dependence on dimension, nonsmoothness parameter, and accuracy is jointly optimal for a known upper bound. The authors prove a matching lower bound for smooth stochastic nonconvex optimization and for Goldstein stationarity under a fully adaptive, reusable scalar oracle. They also give a noiseless exact-value lower bound that applies to randomized and deterministic methods.
WhyProves a matching joint lower bound for zeroth-order nonconvex optimization, resolving an open question with rigorous proofs; impact is mainly theoretical.
In briefSPECTRA-VLA: Adding Hyperspectral Information to a Frozen Vision-Language-Action Model
SummarySPECTRA-VLA adds hyperspectral information to a pretrained vision-language-action model without retraining its visual encoder or action decoder, to help distinguish objects that look similar in RGB. Across 86 held-out acquisitions, category balanced accuracy rose from 0.812 to 0.883. Additional grouped tests report recovery of within-category spectral shape.
In briefMapping Arctic Polynyas from Sea Ice Data: Threshold and Product Sensitivity
SummaryThis paper presents a daily mapping framework that uses hierarchical partitioning, connected-component analysis, and differentiated region growing to delineate Arctic coastal polynyas and open-water areas from sea ice concentration data. Applied to ice seasons from 2002/03 to 2024/25, it reached a mean overall accuracy of 96.74% and a mean kappa of 0.93 across 63 visually assessed scenes. The authors find that absolute polynya extent depends strongly on configuration, while major temporal patterns remain comparatively stable.
In briefSelective Reasoning for Typed Decision Models: When and How to Stop Reasoning Early
SummaryThis paper proposes an interface for typed decision models that lets them use extra reasoning while keeping candidate answers. A shared answer reader is trained on a frozen Qwen3.5-4B backbone, with a benefit gate deciding when to start reasoning and an entropy rule accepting answers and cancelling generation. Across 768 questions from ProofWriter, LogiQA 2.0, and ARC-Challenge, accuracy rises from 70.96% for direct reading to 84.38%.
In briefModel Only What Drives Decisions: Selective World Modeling for Language Agents
SummaryThis paper proposes DGS-WM, a training method for large language model agents that does not treat all observations as world-modeling targets. It selects observation units that support subsequent policy decisions and adjusts supervision strength based on the model's prediction confidence, jointly optimized with the reinforcement learning loss. Experiments on single-hop and multi-hop question answering and interactive web tasks report more stable performance than full-observation world modeling across model scales and RL algorithms.
In briefSynthetic reaction data from transformation rules for pretraining biodegradation pathway predictors
SummaryObserved biodegradation reactions are scarce, which limits training of product and pathway predictors. The paper introduces enviGraph, which generates curated synthetic reaction data by applying environmental transformation rules to structures from PubChem, NORMAN SusDat and EPA DSSTox, yielding 776,851 parent-product pairs for pretraining a graph-to-sequence predictor before fine-tuning on observed reactions. On the EAWAG BBD database, synthetic pretraining gives the highest PR-AUC for single- and multi-generation prediction among the datasets tested, while on the Soil and Sludge datasets the strongest USPTO conditions often achieve higher mean PR-AUC.
In briefJ-tau: A Japanese Benchmark for Evaluating Tool-Using AI Agents in Customer Support
SummaryThis paper introduces J-tau, the first publicly available Japanese benchmark for testing AI language model agents that use tools during interactive customer-support conversations. The authors localized the English-centric tau-bench through machine translation, adapting locale-specific entities and applying human review. They also found and fixed two flaws in the original benchmark: invalid tasks and a score aggregation error that inflated some models' results. Experiments show that J-tau produces fewer evaluation errors than a simply machine-translated version.
In briefSparse autoencoders as a shared coordinate system for matching symbols across systems
SummaryThe paper introduces SAE-IDF, which uses a sparse autoencoder to encode symbols in their native context and re-weights dimensions by rarity to match units across different symbolic systems. Tested on 12 tasks across 6 kinds of symbolic systems, it outperforms dense baselines of the same LLM on most tasks for both Gemma-2-9B and Llama-3.1-8B. Each dimension of the sparse code has a human-readable meaning, so matches can be explained in named concepts.
In briefEye, Robot: Teaching a Humanoid Robot Expressive Gaze Using a Vision-Language-Guided Transformer-VAE
SummaryThe paper presents a humanoid robot gaze system in which a vision-language model selects a scene target, a perception module maintains fixation, and a Transformer-VAE named RGT generates coordinated neck and eye motion toward that target. RGT is trained on markerless human recordings retargeted to the robot's 11-DoF joint space. On a two-second benchmark, RGT-Base achieves the lowest reported overall seven-joint MSE, and physical demonstrations show the integrated pipeline. Perceived naturalness and open-vocabulary target-selection accuracy are left for controlled evaluation.
In briefDeciding When to Trust Machine-Learned Potentials Versus DFT, with Risk Control
SummaryAutomated materials discovery needs to screen many candidate structures, but density functional theory (DFT) calculations are slow. The paper frames the choice between trusting a pretrained machine-learned interatomic potential (MLIP) and falling back to DFT as a risk-controlled selection problem solved with conformal risk control. A lightweight predictor uses only a fixed geometric descriptor of each structure to estimate MLIP-DFT disagreement and set a DFT selection threshold. Under the same guarantee, it routes 7% and 14% fewer configurations to DFT at α=0.3 than committee baselines.
In briefSelf-conditioning lets flow matching perform deeper sequential visual reasoning
SummaryThe paper studies visual reasoning via image generation with flow matching, and shows from a circuit-complexity view that vanilla flow-matching sampling can be simulated by a constant number of parallel evaluation rounds, so extra steps do not add expressive power. It shows that self-conditioning, which conditions each step on the previous estimate of the final image, removes this bottleneck. Experiments on diverse visual-reasoning tasks report higher accuracy and reasoning depth, with performance improving as sampling steps increase, and the authors prove such models can match space-bounded Turing machines.
WhyProves vanilla flow-matching trajectories reduce to constant-depth parallel computation, and shows self-conditioning removes this bottleneck, with theory plus controlled visual-reasoning experiments.
In briefDAGFormer: Predicting Global Topology for Transformer Cross-Layer Connections
SummaryThe paper proposes DAGFormer, a Transformer language model that adds direct connections from earlier to later layers beyond standard residual connections, with a jointly trained predictor setting each connection's strength from input tokens. Unlike methods that set each layer's connections separately, it plans routing for the whole model at once. On matched OLMo-2 models from 75M to 1B parameters, it lowers WikiText-2 perplexity by 17% to 25% over the vanilla Transformer, and gains carry over to 13 evaluation tasks and instruction finetuning.
WhyProposes jointly planning whole-model cross-layer routing, with comparisons across scales, tasks, instruction tuning and pruning against routing baselines.
In briefAdaptive Observation Control for Active Visual Recognition
SummaryThis paper improves active visual recognition, where a model looks at a scene in several steps. It extends the fixed-step AdaptiveNN-DeiT-S recognizer so the model adapts both each glimpse's field of view and the number of glimpses to the input. Experiments on fine-grained benchmarks and ImageNet variants show competitive accuracy with a controllable observation budget. The same checkpoint supports accuracy-compute trade-offs via thresholds without retraining.
In briefMagicGarment: A Benchmark for Reconstructing 3D Clothing from a Single Image
SummaryThis paper introduces MagicGarment, a benchmark for rebuilding standalone 3D garments from a single image. It includes 1,518 professionally built assets across 17 composite garment types, with aligned multimodal annotations and a unified evaluation protocol. The authors also provide a garment-aware baseline that adapts a pretrained image-to-3D model using GSR and FTR-LoRA. Experiments show the benchmark is useful and that garment-aware adaptation is effective.
In briefDo AI Decision Models Show Human-Like Biases? A Jev Case Study
SummaryThis paper evaluates Jev, a model described as a fast, System One decision-maker, using 41,500 responses across cognitive reflection tasks, five classic cognitive-bias tests, and a pricing experiment on ethnic and gender bias. Jev performs well on selected answers and shows small demographic differences, but its outputs are sensitive to irrelevant numbers and option order. On an adapted probability problem, it violated the underlying rule in 54% of 15,000 responses across 30 conjunction tasks.
Ilaria Imperatrice, Taisia Kameneva, Adam Eric Greenberg et al.
In briefInstruction-Conditioned Motion Control for Robotic Arms: Language-to-Code Decomposition and Reinforcement Learning
SummaryThe paper proposes Instruction-Conditioned Motion Control (ICMC), which separates parameter specification, instruction generation, and code synthesis so robot outcomes can be traced to instructions and programs. The authors compare one-shot generation with ICMC on cube pick-and-place and letter-reshaping tasks for a UR5e arm in MuJoCo, and train an open-weight instruction policy with GRPO. A log audit found positive labels tied to geometric-failure messages, so the authors distinguish logged reward rates from independently verified task success.
In briefDivergent Paths: Learning a Representation That Predicts LLM Agent Outcomes
SummaryThe paper treats agent trajectories as paths through a shared embedding space of actions and observations, where successful and failed rollouts of the same task separate. Simple per-task discriminators predict held-out outcomes at up to 0.95 AUROC across three benchmarks and three policy models. The authors introduce Precog-RM, a transformer that meta-learns this representation for few-shot outcome prediction, which generalizes to unseen tasks and models. Used as an early stopping controller for GRPO on Tau2-Bench, it cuts training FLOPs by up to 47.3% while keeping performance comparable to baseline GRPO.
In briefAre We Really Measuring Forgetting? Pitfalls in Evaluating Multimodal LLM Unlearning
SummaryThis paper argues that existing benchmarks for machine unlearning in multimodal large language models inherit text-only evaluation protocols and have three pitfalls: no heldout evaluation, reliance on synthetic fine-tuned knowledge, and unfair comparisons from identical training settings. The authors introduce GEMMU, a benchmark built on real-world knowledge with a principled evaluation protocol. Re-evaluating representative unlearning methods under GEMMU reveals substantial generalization gaps and challenges conclusions from previous benchmarks.
In briefWhere multi-image schema tasks fail: a bottleneck in recovering ordered instances
SummaryThe paper studies vision-language models on a task requiring them to find two characters across 16 images, identify their attributes, and compose code names from a visual codebook. Using a 2-by-2 factorial design, the authors supplied either a verified textual schema or a verified ordered target-appearance representation. The ordered appearance information improved accuracy far more than the schema did. The authors conclude the main recoverable deficit lies in upstream ordered instance recovery rather than downstream schema use.
In briefCORE: A training-free method that optimizes prompts for synthesizing industrial anomalies
SummaryIndustrial anomaly detection is limited by the scarcity of anomalous samples, so realistic anomaly synthesis is important. The paper proposes CORE, a training-free framework built on large multimodal models that synthesizes anomalies in zero-shot and one-shot settings and refines its prompts through a self-supervised contrast-and-reflect mechanism. Experiments on MVTec-AD and VisA show that it generates realistic, diverse anomalies and improves downstream anomaly detection performance.
In briefTwenty Days Later: A Pre-Registered Test of Whether Past Brand Mentions Predict Future Gemini Answers
SummaryThis pre-registered study tests whether how often a brand was mentioned in past Gemini answers can forecast mentions in later answers. The same 162 requests across 54 questions in three languages were resent about 20 days later. History-based forecasts outperformed several baselines and held across all nine question families. The test covers one model, one interval and twelve brands, and does not estimate a decay rate.
In briefPN-Mix: Pair-norm guided mixed-precision key cache compression for faster long-output LLM inference
SummaryLong-output inference in large language models is limited by KV cache memory and bandwidth, and low-bit quantization of the cache can hurt accuracy. The paper proposes PN-Mix, which assigns higher precision to salient RoPE channel pairs identified once during prefill and reused during decoding. Across reasoning, code-generation, and AIME benchmarks, PN-Mix at 2.26 average bits matches or exceeds prior methods using 2.32 to 2.39 bits, and degrades more gracefully down to 1.5 bits.
In briefBeyond Static Positional Encoding: Bayesian Time-aware Attention for Long-term Forecasting
SummaryThe paper interprets self-attention as a Bayesian posterior over attended positions and derives a population-level evidence lower bound describing how positional encoding affects forecasting. It argues that static positional encoding can be suboptimal when temporal dependencies vary across input windows. The authors propose Bayesian Time-aware Attention (BTA), which uses a mixture-of-experts router to combine temporal kernels into input-conditioned positional biases. Experiments on standard benchmarks report consistent gains across several Transformer backbones.
In briefPA1-OPD: A One-Step Prefix-Augmented Method for On-Policy Distillation in Language Model Pretraining
SummaryThe paper introduces PA1-OPD, which adds teacher feedback after one student-sampled transition per selected prefix to corpus-path knowledge distillation. This supervises contexts induced by student predictions as well as corpus prefixes. In two continued-pretraining settings, it improves held-out cross-entropy over knowledge distillation across C4, PG-19, and Pile at every evaluated checkpoint.
In briefBeyond Source Attribution: Author Addresses and Comparability Decisions in an LLM-Built Knowledge Graph
SummaryThe paper argues that a scholarly knowledge graph needs stable addresses for its evidence, because records gathered under one name may refer to units that differ in boundaries, rank, or counting criteria, which quotation checks alone cannot detect. It distinguishes the working address used for retrieval, the category defined by a source, and the decision that admits a passage into a given comparison, treating these decisions and record dependencies as part of comparison provenance. Using Sinitic dialectology (Yue) as an example, it presents the PaRAGraph Casebook, four re-runnable cases, and how the platform stores these decisions as objects.
Ksenia Kozha (American University of Cyprus), PaRAGraph Research Project
In briefJevCoder: When Delegating Decisions Pays Off in Code Repair
SummaryThis paper studies which decisions a coding agent makes during code repair can be delegated to Jev, a low-cost probabilistic judge, to cut token use while preserving repair quality. The authors compare interventions against unguided repair and against using the repair model itself as the judge. Based on these findings, they propose JevCoder, which combines code-snippet recommendations with action-type guidance and uses judgments to decide when to submit. Across 400 tasks from SWE-bench Verified and SWE-Bench Pro Verified, JevCoder reduces repair-model token use by 21.6% to 50.6% compared with unguided repair, with higher observed repair rates for all three models.
In briefInstructPro2: Designing New Proteins Guided by Plain-Language Instructions and Biological Context
SummaryThis paper introduces InstructPro2, a model that turns natural-language design instructions into protein sequences while adding task-specific biological context such as ligands and antigens. The authors built a pretraining set of 205.7 million function-description and protein pairs from UniProtKB, then adapted the model to ligand-binding protein and antibody design. On functional protein generation benchmarks, it reached a functional recovery F1 of 74.8 versus 62.2 for the baseline Pinal, and lowered global RMSD from 5.33Å to 3.04Å. The model achieved a DockQ score of 0.39 for antibody design and a 72.6% docking success rate for ligand-binding protein design.
In briefL3M: A Verifiable Runtime for Self-Extending, Self-Improving AI Agents
SummaryThis paper presents L3M, a runtime written in Lean 4 that lets AI agents carry out work over long periods and extend their own capabilities. All state is stored in Git, so computations can be resumed, rewound, forked, compared and merged. Agents can save a procedure as a sealed Lean program that serves as a tool, and a tool's review tier depends on how reversible its effects are. On Terminal-Bench, an untuned L3M reaches about two thirds of the pass rate of its unsafe shell-equipped setup using only tools that do not need human review.
Carson Eisenach, Robert Joseph George, Vincent Quenneville-Belair et al.
In briefJockeyBench: Testing LLM Agents on Horse-Racing Predictions
SummaryThis paper introduces JockeyBench, a benchmark that uses historical Japanese horse racing to study how AI agents make repeated forecasts from noisy evidence. The evaluation separates forecast quality, benchmark-specific utility, and retrospective betting returns and risk. Across 28 model-scaffold configurations on 324 races, every configuration had negative mean profit under capped-Kelly settlement.
In briefEstimating Hawkes process background intensity with projected Wasserstein-Fisher-Rao flows
SummaryThis paper studies estimating the background intensity of a Hawkes process when the triggering mechanism is known, treating the background as an unknown finite positive measure with a regular density. It proposes a likelihood-based variational method using Wasserstein-Fisher-Rao geometry, with dynamics projected onto statistical manifolds to obtain a finite-dimensional meaning. Under suitable conditions, the authors prove a conditional consistency theorem, and they derive an explicit finite-dimensional scheme for Gaussian-mixture manifolds.
In briefMoral Combat: Using moral reasoning to trick AI models into harmful answers
SummaryThis paper proposes a multi-agent framework based on Moral Foundations Theory that rewrites harmful requests, which are initially refused, so they appear morally justified, and reinforces them with generated precedents of similar moral reframing. The authors test the method on harmful prompts from three well-known jailbreak datasets across five open-source and proprietary models. They find that models are highly susceptible to moralization, with high single-attempt jailbreak rates that rise under repeated attempts. They also report that existing safety benchmarks are imbalanced across the moral foundations they cover, focusing on those toward which models already show bias.
WhyApplies moral foundations theory to jailbreaking via a multi-agent reframing framework, evaluated across three datasets and five models with high success rates.
In briefExplainer: Which model wrote this text? Robust model authorship attribution under distribution shift
SummaryThis paper studies Model Authorship Attribution, which identifies the large language model that generated a given text, relevant to academic misconduct, model distillation and cyberattacks. The authors propose ATLAS, which attributes text by comparing its likelihood under each source model's learned condition, and introduce the MAABench benchmark. When both domain and model version change on MAABench, ATLAS reaches 74.27% accuracy, outperforming the strongest baseline. The method is also reported to recover the teacher model of distilled models.
In briefXferAgent: Sharing Experience Across LLM Agents to Improve Performance
SummaryThis paper proposes XferAgent, a framework that lets different LLM agents share and reuse experience from past tasks. It organizes experience in local and shared graphs, merging compatible contributions while avoiding unreliable merges. Experiments across task domains show gains over representative memory-based methods and transfer between heterogeneous agent frameworks, with cumulative task success rising over successive passes.
In briefBridging the Topological Divide: A Dual-Branch Distillation Method for Robust Cold-Start Recommendation
SummaryThis paper addresses the cold-start problem in graph-based recommendation, where unobserved items become topologically isolated and their representations collapse. It proposes DualGCN, a dual-branch graph knowledge distillation framework that first trains a collaborative teacher using topology-smoothed semantic priors, then trains a semantic student via residual mapping and a magnitude-agnostic cosine distillation objective. A dynamic norm reconstruction mechanism is also introduced to calibrate item scales during retrieval. The authors report that DualGCN outperforms state-of-the-art baselines in strict cold-start settings on three real-world datasets while preserving accuracy for active items.
In briefTransferring Learned Patterns Across Temporal Graph Representations
SummaryThis paper studies whether predictive structure learned from relational histories can be reused between continuous-time dynamic graphs (CTDGs) and temporal knowledge graphs (TKGs). It finds that reuse works in both directions, and that sources sharing the same representation are not always more useful. Independently pretrained TKG and CTDG models can be fused at the score level to improve prediction on a third dataset without entity or relation alignment.
In briefInherent Privacy of Posterior Sampling: A Unified Rényi Divergence Framework
SummaryThis paper develops a unified Rényi divergence framework to characterize the differential privacy guarantees that arise inherently from sampling a Bayesian posterior, without adding calibrated noise. The authors show that the privacy loss of One-Posterior-Sampling (OPS) depends on posterior exponential moments of the single-record log-likelihood ratio, and give explicit guarantees under bounded, sub-Gaussian, or sub-exponential tails. The framework is applied to categorical, Gaussian-conjugate, exponential-family GLM, and Gaussian linear regression models, recovering existing OPS guarantees as special cases, often with tighter bounds.}
In briefDivergence-Routed On-Policy Distillation (DrOPD) for Language Models
SummaryThis paper studies on-policy distillation, where a student language model learns from a teacher along its own generated trajectories, usually by minimizing reverse KL divergence. The authors find that routing by divergence identifies positions behind the teacher-student gap better than entropy-based routing, and that high-JS positions suffer from teacher-mode under-coverage. They propose DrOPD, which keeps reverse KL everywhere and adds forward KL only at high-JS positions. Across six math benchmarks, DrOPD outperforms standard OPD and entropy-based routing, and reduces severe under-coverage cases by 37.9 percent relative to OPD.
In briefLED-BENCH: A Benchmark for Understanding Large-Format Engineering Drawings
SummaryThe paper introduces LED-BENCH, a multidisciplinary benchmark for testing how well AI models understand large-format engineering drawings in mechanical, architectural, and transportation fields. It evaluates 11 multimodal large language models, including 8 proprietary and 3 open-source ones. The best model reaches 52.65% average accuracy, performance varies by domain, and all models do worse on questions requiring broader evidence integration.
In briefGeometric Kernel Diversity: an axiom-satisfying, efficiently computable diversity measure in Euclidean space
SummaryThe paper addresses how to measure the diversity of a set of objects with a reliable and efficiently computable metric. It introduces Geometric Kernel Diversity (GKD), which attaches a radial kernel solid to each point and measures the volume of their union, and shows it satisfies monotonicity, uniqueness, and continuity on Euclidean-realizable inputs. The authors give exact polynomial-time algorithms for dimensions up to three and a randomized approximation for arbitrary dimensions, while showing that two previously known axiomatic measures remain NP-hard on this restricted domain.
WhyProposes a Euclidean diversity measure satisfying all three axioms, with exact low-dimensional algorithms, a randomized approximation, and hardness proofs for prior measures.
In briefINTACT: Inducing Latent Reasoning on a Frozen Backbone
SummaryThis paper presents INTACT, a method that uses a small external encoder to induce latent reasoning in a frozen backbone model. The encoder maps the backbone's hidden states into latent blocks that form a reasoning prefix, after which the frozen model writes the rest of its reasoning and answer as text. The encoder is first trained by distillation and then reinforced on answer correctness. Across four frozen backbones from 1B to 8B, it improves GSM8K and average out-of-domain accuracy and reduces per-problem latency by 19% on Llama-3.2-3B.
In briefFrame Evidence: A Benchmark Testing Reference Frames and Visual Evidence in Multi-Image Spatial Reasoning
SummaryThis paper introduces Frame Evidence, a benchmark of 320 matched triplets and 960 Chinese multiple-choice questions based on captured and rendered scenes. It tests whether answers stay correct when the language-specified reference frame changes or when supporting images are deleted or added. Results show that high single-item accuracy does not translate into consistent joint correctness across all frames and evidence variants, and few-shot prompting and task-specific adaptation do not reliably fix this.
In briefSaTPO: Aligning Text-to-Image Models for Safety via Step-wise Transition Preference Optimization
SummaryThis paper proposes SaTPO, a method that enforces safety alignment step by step during the denoising process of text-to-image diffusion models. It optimizes preference pairs formed by an unsafe prompt and its safe rewrite under a shared parent state and noise, and adds multi-scale drift constraints and benign replay to preserve ordinary generation. Experiments on Stable Diffusion v1.5 and SDXL report improved safety-quality trade-offs, and the authors also release the CoProPV dataset.
In briefREWIND: Uncertainty-guided repair of reasoning paths at test time
SummaryThe paper notes that early mistakes in generated reasoning can constrain all later steps, and that repeated-sampling methods are limited when they randomly choose where to regenerate. On MATH500 with Qwen2.5-Math-7B, incorrect responses show substantially higher mean token entropy than correct ones, suggesting entropy can signal reasoning errors. The authors introduce REWIND, a training-free controller that uses entropy dynamics to target suffix regenerations, and report that it outperforms all baselines across three model families and five benchmarks.
In briefτ-Emotion-Bench: Testing How User Emotions Break LLM Agents During Tool Use
SummaryThis paper introduces τ-Emotion-Bench, a ten-domain benchmark that tests tool-calling AI agents when users express different emotions while the task goal stays the same. It also presents Tracer, a pipeline that generates training data and filters out interactions where users must depend on the agent excessively. Experiments show that frontier agents remain unreliable, and Tracer-generated data improves Qwen3 models and transfers to other benchmarks.
Rishitosh Kumar Singh, Amir Saeidi, Shamanthak Hegde et al.
