In briefFORESIGHT: Planning future perception in streaming vision-language models without retraining
SummaryExisting streaming vision-language models keep a fixed computation path during inference, so they cannot adapt to changing scene dynamics. This paper proposes FORESIGHT, a dual-stream architecture with two Siamese LLMs sharing weights, one processing the stream and the other anticipating future context and planning computation. With a frozen Qwen3-VL-8B backbone, the method reports gains on several streaming video benchmarks without training.
Ashok Prasad Neupane, Dipan Bartaula, Ankit Belbase et al.
In briefIn-Distribution Forcing: a test-time method to reduce drift in long video generation
SummaryAutoregressive video diffusion models generate short videos well, but long videos suffer from drifting in colors and textures and from decaying motion. The paper argues that KV conditioning alone is insufficient, because cached key-value entries can become out-of-distribution beyond the training horizon. The proposed ID-Forcing framework uses self-caching, where each chunk is cached without attending to prior KV entries, keeping the rolling window in-distribution. The authors report that this extends short-horizon models to minute-scale videos while remaining competitive on standard benchmarks and reducing drift in their metrics and a user study.
In briefMethod for finding and reusing past policies to adapt continuously in lifelong reinforcement learning
SummaryThis paper studies how a lifelong reinforcement learning agent can select and combine previously learned policies to transfer knowledge to a new task. The authors propose AMSC, which estimates task similarity from online experience using non-parametric Wasserstein task embeddings and uses it to choose and weight prior policies. On CT-graph and MiniGrid, AMSC achieves higher mean performance and forward transfer than the evaluated modular composition baselines without forgetting, while Continual World results suggest layer-specific tuning may be needed. Ablations indicate that selecting relevant sources and setting how strongly to reuse them are central to these gains.
Saptarshi Nath, Inish M. D'Souza, Antonio Carta et al.
In briefStructured Pruning and Fault Tolerance in Neural Networks for Space
SummaryThis paper studies how neural network width affects robustness to single event upsets (SEUs), which are bit flips caused by radiation. The authors iteratively prune baseline models to reduce width and run fault-injection tests at each stage. They find that pruning raises per-inference sensitivity to faults, but shorter execution time offsets this effect. The results suggest pruning can save energy and latency without substantially harming overall reliability.
In briefDistilling Production LLM Signals into Interpretable Resume-Vacancy Matching Features
SummaryThe paper presents an interpretable approach to matching candidates with vacancies, showing fit as eight named dimensions that recruiters can act on. It uses an LLM-based labeler for offline labeling and a bi-encoder distilled from it that serves online requests on CPU. The model was trained on 168,772 labeled vacancy-resume pairs. The deployed student agreed with 888 of 927 recruiter-recorded values (95.79%), though this is non-blinded operational agreement rather than an independent human evaluation.
Ilya Chekin (1), Vyacheslav Malyugin (1), Vladimir Chirkov (1) et al.
In briefWhy AI-text detectors fail when machine text is made to sound human
SummaryThe paper examines a RoBERTa-based AI-text detector under semantic, structural, and tokenizer-level perturbations. When Mistral-7B-Instruct was asked to make machine text sound more human, verb diversity rose and the outputs became easier to detect. The authors report a 76.3% false-positive rate on formal human writing and find that an event-based latent space detector was unstable under paraphrasing and homoglyph changes.
In briefShared internal attention patterns across 60+ language models
SummaryThis paper examines post-softmax attention weights by marginalizing over query positions to form a token-head space, and studies how these patterns behave across many language models. Reducing along tokens yields a text-intrinsic signal consistent across models, while reducing along heads yields a model-specific signature stable across documents. The authors use this to precompute per-head budgets for key-value cache eviction offline, which performs competitively on standard benchmarks.
WhyMarginal attention reveals cross-model conserved text signals and head-specific signatures, with a Jacobian-based proof, enabling decoupled KV cache budgeting; evaluated on 60+ models but limited benc
Valentino Maiorca, Walter Nelson, Francesco Locatello
In briefA benchmark for checking whether generated pointing gestures aim at the right object
SummaryThe paper introduces a benchmark for spatially grounded gesture generation, built from about 2K pointing-annotated clips of natural VR dialogue with ground-truth 3D referents. It defines a task requiring systems to decide when, how and where to point during speech, and an evaluation protocol that separates temporal alignment, spatial grounding and perceived naturalness. The authors also provide a flow-matching baseline, MM-Conv-Flow. They find that its geometric grounding can exceed that of human pointing without improving perceived naturalness, so referential gesture quality must be measured along separate dimensions.
Anna Deichler, Rishabh Dabral, Fethiye Irmak Dogan et al.
In briefS2S-JEPA: forecasting only the predictable parts of weather two weeks to two months ahead
SummaryThis paper introduces S2S-JEPA, which applies the Joint-Embedding Predictive Architecture to subseasonal-to-seasonal weather forecasting, roughly two weeks to two months ahead. Instead of predicting unpredictable fine-scale details, it forecasts only the slowly varying components in latent space. The authors report that its skill is comparable to the ECMWF physics-based ensemble and surpasses it on multiple metrics at weeks 5 to 6.
In briefAsk, Relax, or Act? Evaluating When LLM Agents Should Intervene
SummaryThis paper studies whether an LLM agent should act directly, ask for clarification, or repair constraints when a request is uncertain. The authors build a solver-grounded benchmark covering object allocation, meeting scheduling, apartment choice, and stable matching. They find that models can recognize uncertainty yet often intervene even when a justified action already exists, and that correct decision labels do not guarantee usable responses.
In briefWhether causal effect identifiability survives variable clustering in causal graphs
SummaryThe paper studies how clustering variables in causal graphs, which simplifies the graph, affects whether causal effects remain identifiable. It notes that arbitrary clustering can change crucial causal relations and lead to wrong conclusions. The authors present a broad class of clustering operations, based on conditions tied to the c-components of the original graph, that preserve both identifiability and nonidentifiability, and demonstrate use in practical settings.
In briefComparing Across Videos: A Benchmark and Active Evidence Search for E-Commerce
SummaryThis paper introduces AdsCVR, the first e-commerce benchmark for reasoning across multiple videos, with 2,483 videos and 6,110 question-answer pairs over six reasoning dimensions. It also proposes AdSeek, an agentic framework that dynamically selects visual and audio tools during multi-turn exploration to actively gather evidence. Using offline trajectory rectification combined with reinforcement learning and supervised fine-tuning, AdSeek reaches 74.30% accuracy on the AdsCVR test split and generalizes to the open-domain CrossVid benchmark.
In briefPredictor-guided latent-space codon optimization to boost protein expression
SummaryThis paper proposes Latent-Space Codon Optimization (LSCO), which turns codon selection into a continuous problem by mapping sequences into the latent space of a pretrained mRNA language model for gradient-based search. LSCO combines an uncertainty-aware expression predictor, a structural stability regularizer, a naturalness prior, and constrained decoding to preserve protein fidelity. On a real wet-lab antibody expression dataset, it outperforms simple frequency-based and modern deep generative baselines in predicted expression while retaining suitable biophysical properties.
Alberto Caron, Tianyu Cui, Dmytro S. Lituiev et al.
In briefPeer Influence Among Different AI Models: Who Persuades Whom
SummaryThis paper studies who persuades whom when two AI agents disagree. The authors tested seven open-weight models on three language understanding tasks and found persuasion is strong, as receivers often abandon their initial judgment after seeing a peer's answer and explanation. Neither standalone certainty nor model scale reliably predicts these dynamics, which instead depend on each specific model pairing.
Frida N{\o}hr Laustsen, Marie Haahr Petersen, Victoria Popa et al.
In briefLS-AR: Steering Language Model Generation Using Latent Goal Vectors
SummaryStandard autoregressive language models mix task instructions, state history, and transient tokens in one sequence, making it hard to keep high-level goals separate from context noise. The paper proposes LS-AR, a dual-channel architecture that uses FiLM conditioning to separate continuous goal steering from discrete token decoding, with a static goal encoder and a dynamic state tracker. The authors report results on long-horizon retrieval and Blocksworld planning, and they also discuss a latent-vector attack surface introduced by the design.
In briefULTRADISCOVERY: A test of whether AI agents can revise theories and predict cross-domain outcomes
SummaryThis arXiv paper introduces ULTRADISCOVERY, an interactive environment of five domains in which an agent revises an initially successful theory and predicts an unseen cross-domain intervention. A 2×2 design independently controls whether the representation is open or disclosed and whether evidence is distributed or aligned. With the representation open, most models retract the taught axiom but none introduces the unobserved entity or rewrites the needed variables. The authors locate the main difficulty in composing accumulated evidence into a representation that transfers.](summary-check) ---
WhyA 2×2 design separates representation openness from evidence distribution, with eleven models showing agents rarely introduce unobserved entities or rewrite variables.
In briefZephon: Elastic Determinism for Online, Stateful Data Loading in Foundation Model Training
SummaryThis paper presents Zephon, a data loader for training foundation models that keeps the sequence of training batches deterministic even when GPU topology, checkpoint-resume cycles, or execution backends change. It splits the global data stream into topology-independent lanes, serializes ordering decisions while parallelizing other work, and checkpoints only bounded in-flight state. The authors evaluate it on text and vision-language workloads and report competitive throughput with guarantees existing loaders lack for online, stateful pipelines.
Maximilian B\"other, Josh Wills, Ties Robroek et al.
In briefA lightweight method for learning entropic optimal transport couplings directly on curved spaces such as spheres and rotation groups.
SummaryEntropic optimal transport is used to learn stochastic couplings between distributions, but most solvers target Euclidean spaces and manifold extensions are often costly. This paper introduces ManifoldLightOT, which uses the kernel form of the EOT solution to build geometry-specific Gibbs kernels and potential parameterizations for spheres, tori, SO(3), and SE(3), extendable to products of manifolds. The parameters are optimized from samples using Monte Carlo estimates. Experiments on synthetic and real-world data show it often outperforms existing manifold OT methods while retaining direct sampling.
Xavier Aramayo-Carrasco, Petr Mokrov, Alexander Korotin
In briefNegT2IBench: Testing Whether Text-to-Image Models Can Honor Negated Instructions
SummaryText-to-image models are usually tested on whether requested content appears, but rarely on whether they can avoid things a prompt forbids, such as "a non-red cup." The authors introduce NegT2IBench, a benchmark of 4,800 prompts that independently varies positive and negated statements, with a reproducible detector-based scoring method that identifies which requirement failed. Across eleven models and 211,200 images, nine models score lower on a single negated statement than a single positive one, with the largest gap for color. The paper also finds that 41.5% of failed statements render exactly what the prompt forbids.
Omar Elfatairy, Maria A. Bravo, Jessica Bader et al.
In briefSmart sensing to make bridges safer, using signal processing and AI to spot anomalies in sensor data.
SummaryThis paper studies anomaly detection in real sensor data from a bridge in Norway, comparing signal processing with the machine learning method Isolation Forest. The methods are evaluated on anomaly counts, detection time, processing rate, anomaly rates, and temporal agreement, and a controlled anomaly-injection test measures each method's sensitivity. The results show the two approaches have distinct detection characteristics and computational needs, suggesting that machine learning, especially Isolation Forest, combined with signal processing can help identify anomalies in bridge measurements.
In briefRIFAR: A Replay Method That Helps Robots Learn New Tasks Without Forgetting Old Ones
SummaryThe paper proposes RIFAR, a generative replay framework for continual robot learning that rebuilds past trajectories from compact historical demonstration prefixes. A trainable inverse-dynamics model checks whether the actions match the predicted visual transitions, and training prioritizes reliable trajectories whose action predictions drifted most after adapting to new tasks. Across three LIBERO suites and real-world experiments, RIFAR outperforms the prior state of the art in world-action model-based generative replay. On LIBERO-Goal it reaches an AUC of 90.97 while retaining only 320 historical time steps per task, about 4.9% of what 50-demonstration replay keeps.
In briefAn automated pipeline that labels turns and listener backchannels in recorded spontaneous conversations.
SummaryThe paper presents an automated pipeline that extracts conversational turns and backchannels from separate-channel recordings of spontaneous dyadic dialogue. It was evaluated on 99 ten-minute Danish conversations, reaching an overall detection F1 of 0.621 with median boundary errors of roughly 0.13 to 0.18 seconds for matched segments. The authors position the pipeline as a first pass within a semi-automated annotation workflow that still requires human review.
Hanlu He, Harald Vilhelm Skat-R{\o}rdam, Ingvi \"Orn\'olfsson et al.
In briefUnmasking Propaganda: Comparing Masked and Causal Language Models
SummaryThis paper compares masked language models (XLM-RoBERTa, DeBERTa V3) and causal models from several providers for detecting propaganda techniques, using the SemEval-2020 Task 11 dataset. It tests base and chain-of-thought prompting strategies. The best masked model reaches an F1 score of 63.18 and the best causal model 63.62 in technique classification, which the authors report as improvements over state-of-the-art. The authors also note that models vary in performance across techniques and suggest fine-tuning, ensembling, and larger datasets could help further.
In briefLearn Once, Optimize Any Objective: Derivative-Free Diffusion Models for Chance-Constrained Optimization
SummaryThis paper addresses decision-making under uncertainty where constraints may be violated only with limited probability. It proposes D³Opt, which trains a risk-conditioned diffusion model on constraint-feasible decisions and freezes it as a reusable prior. At inference, a particle-based Feynman–Kac correction optimizes user-specified objectives using only function evaluations. Experiments on linear Gaussian chance-constrained programs, objective-transfer tasks, and chance-constrained economic dispatch are reported.
In briefFinding Where Compositions Fail Doesn't Show How to Fix Them
SummaryThis paper studies why text-to-image diffusion models often fail on compositions such as attribute-object pairs. The authors introduce a text-only Compositional Stress Index (CSI) that separates common from rare compositions across SD1.5, SDXL, and the SD3 text path. Experiments show defects can be diagnosed before denoising, but the representation that reveals risk is not necessarily the place or modality whose intervention best improves generation.
In briefExplainable GC-MS structure inference using diffusion models and LLM reranking
SummaryThe paper introduces DiffGCMS, a spectrum-conditioned discrete graph diffusion model that generates candidate molecular structures directly from GC-EI-MS spectra without relying on reference libraries. A large language model then validates, repairs, and reranks these candidates while interpreting fragment-ion peaks. On a 13,696-spectrum NIST 20 test set, the generative model reached Acc@1 of 6.01% and Acc@10 of 15.76%. On molecules with at most 10 heavy atoms, LLM-assisted repair and reranking raised Acc@1 from 21.28% to 21.95% and candidate validity from 91.04% to 100%.
In briefTransferable control policies learned from action-free time series via shared dynamical embeddings
SummaryThis paper proposes a hierarchical model-based reinforcement learning framework that learns system-specific control policies from action-free recordings by sharing structure across related systems. Low-dimensional embeddings capture both shared dynamics and individual variation and are used to parameterize shared policy and value networks. On Lorenz-63 and double-pendulum systems, the hierarchical policies transfer better than independently trained ones, and in neural-behavioral recordings they suppress predicted movement under constrained neural perturbations.
In briefHARPO: Reinforcement learning that balances faithful and creative language generation
SummaryThis paper proposes HARPO, a reinforcement learning framework that jointly optimizes faithfulness and creativity in large language models. It uses a Hallucination-Aware Generative Reward Model and a Selective Activation Mechanism, which applies writing rewards only to hallucination-free outputs, along with a curriculum moving from creative writing to hallucination-centric tasks. On RAGTruth, a Qwen3-4B-based reward model reaches 78.08% response-level F1, versus 66.37% for supervised fine-tuning, and on Qwen3-4B the hallucination rate on MultiHopRAG falls from 3.29% to 1.02% while the Arena-Hard-v2.0 creative-writing score rises from 16.95% to 27.54%.
In briefWhen Synthetic Relational Data Teaches Models to Use Relations: Tracing Pretraining Data to Model Behavior
SummaryThe paper compares Relational Transformer models pretrained on synthetic databases from four relational data generators to find which data property leads models to learn cross-table relational computation. The authors hypothesize that relational mechanisms emerge when cross-table information is predictively necessary for the masked-cell pretraining objective. The RelDiff-generated data shows the largest predictive gain from foreign-key-linked parents, and its model is uniquely sensitive to foreign-key interventions. Disrupting this mechanism during downstream inference removes RelDiff's advantage on relational tasks.
Shivam Dubey, Mohamed Bouadi, Nassim Bouarour et al.
In briefMOF-VERIFY: A Failure-Aware Agent Framework for Verifying Metal-Organic Framework Hypotheses
SummaryThis paper builds a diagnostic benchmark with four task families to pinpoint where large language models fail when verifying hypotheses about metal-organic frameworks (MOFs), covering knowledge access, evidence gathering, and reasoning. Based on these failure modes, the authors propose MOF-Verify, an agentic framework that addresses structural, literature, evidence-sufficiency, and computational bottlenecks before giving a final verdict. The authors report that across several backbone LLMs, MOF-Verify substantially outperforms direct inference and retrieval-based baselines.
In briefHACKTRACE: Detecting reward hacking in AI code generation by supervising behavior
SummaryThe paper releases 173,561 annotated multi-turn coding trajectories from Qwen3-8B, where agents may pass tests by shortcuts such as deleting the failing test. It introduces HACKTRACE, a monitor that reuses the model's internal states during code generation, achieving a mean per-problem AUC of 0.997 with 8 ms overhead. When used as a signal in reinforcement learning with GRPO penalties, it reduces the cheating share of passing solutions from 82-91% to 1-5%.
In briefRIPPLE: Offline Zero-Shot Client Clustering for Federated Learning via Wavelet Scattering
SummaryThis paper proposes RIPPLE, a clustered federated learning framework that assigns clients to clusters offline, before training, based on a spectral summary of each client's local data. The summary combines a wavelet scattering embedding of principal components with a Gaussian Mixture VAE trained on the server using synthetic client populations. Per-round communication matches FedAvg, and unseen clients can receive a personalized model in a single forward pass. The authors report that RIPPLE outperforms all baselines across five benchmarks, with larger gains on the most realistic heterogeneous partitions.
In briefGeometric Structure of Knowledge Accessibility in Large Language Models
SummaryThis paper studies whether large language models can reliably recall the knowledge needed to answer a query. It finds that knowledge accessibility has a simple geometric structure in the representation of the query alone, before generation: queries closer to a center are more accessible, and those farther away are less so. This geometry reveals a knowledge boundary, and the paper reports that different interventions help most in different regions relative to it.
In briefBeeWhere: Segmenting Bumble Bee Colonies to Quantify Behavior
SummaryThis paper presents BeeWhere, a workflow that combines ArUco tag detection with deep-learning instance segmentation to quantify bumble bee behavior from high-resolution colony images and videos. Using Bombus impatiens microcolonies, the authors annotated 483 frames with 8,443 bee instances and trained YOLO segmentation models. In an exploratory study of neonicotinoid pesticide exposure, BeeWhere detected more bees than tag-based tracking alone and captured spatial organization changes that tag tracking missed.
In briefExtracting 3D human motion from a frozen video diffusion model without training a separate motion generator
SummaryLarge text-to-video diffusion models produce realistic human motion in their videos without ever being trained on explicit 3D motion data. This paper asks whether that implicit knowledge can be converted into explicit 3D motion generation without training a separate motion model. The authors propose parasitic co-denoising, which decodes motion from the host model's denoising process, and implement it as the Parasitic Motion Decoder (PMD), which leaves the host unmodified. PMD outperforms dedicated motion generators on text-motion alignment with a small fraction of their trainable parameters, and produces paired video and motion in a single pass.
In briefA Score Is Not Structure: Brain Alignment and Cross-Lingual Transfer
SummaryThe paper questions whether similarity scores really show shared structure between models and the brain, or across languages. Across seventeen languages, a grammaticality probe transfers worse as languages grow more distant, but the probe's own accuracy drops along the same axis and sits at chance in four languages. Shuffled or correspondence-destroyed brain targets also achieve similar scores. The authors recommend measuring how much of a correspondence score survives without any real correspondence before interpreting it.
In briefHyperThink: A method that speeds up AI reasoning by predicting parameter updates instead of long thinking
SummaryThis paper proposes HyperThink, which uses a lightweight hypernetwork to read a question and predict updates to a small subset of a large language model's parameters, replacing long reasoning traces. A vector-quantized decoder constrains these updates to reusable patterns. Reported experiments on mathematical and general reasoning tasks show improved accuracy-latency trade-offs in the low-latency region.
In briefWebFovea: Reliable vision-based web agent execution on live websites, where failures come from the harness rather than the model
SummaryWebFovea is a vision-based web agent that placed second in the WebRetriever Challenge 2026 with a score of 57.0 out of 100. The paper argues that many failures on real websites come from the code connecting the model to the browser, such as coordinate mismatches, silently failing dropdown or iframe actions, and chat-template token contamination. Using the same model throughout, the authors hardened these stages and added guardrails, raising the official hidden-set score from 31.0 to 57.0.
In briefBalancing multimodal learning using function-space progress signals
SummaryMultimodal learning often suffers from modality imbalance, where one modality dominates joint optimization. This paper proposes FGMO, which uses a function-space progress signal to assess each modality's optimization progress and coordinate training across modalities. It introduces Functional Progress Estimation and Functional Response Control, with theoretical analysis and experiments on multiple multimodal benchmarks.
In briefAdaptive second-order solvers that adjust step sizes from error feedback to speed up diffusion model sampling
SummaryDiffusion models generate samples by running numerical solvers over discretized time steps, and the choice of discretization affects the trade-off between sampling cost and quality. This paper adapts proportional-integral (PI) step-size control to diffusion, using both current and previous error estimates for smoother step adaptation. It also finds that per-sample trajectories share structure that can be aggregated into a fixed schedule retaining much of the benefit. Results on image and language data are mixed, and the authors report that the benefit of per-sample adaptivity is problem-dependent and marginal for image and language data.
In briefWhen Numbers Start Talking: How LLM Agents Use Numerical Signals in Strategic Games
SummaryThe paper studies AI agents built on four popular LLMs as they play four games with different cooperative equilibria. It compares natural language, numerical signals, and random sequences to see whether each message type changes cooperation levels. Structured messages alter final payoffs for most games and models, but without a predictable pattern. The authors suggest monitoring agent coordination through message-level fingerprints, which generalise across models, rather than behavioural decisions.
Alessio Buscemi, Daniele Proverbio, Alessandro Di Stefano et al.
In briefAre We Measuring Anticipation? Auditing Hidden Advantages in Procedural Video Tests
SummaryThe paper audits what an evaluation protocol actually licenses, not just the model's predictions. Using procedural action anticipation as a case study, it shows a test can stay temporally causal and free of classic leakage yet still supply a privileged intermediate representation, so the claimed capability differs from what is measured. On Breakfast, matched settings score 30.3% and 62.3%, and the authors propose a reusable four-condition audit.
In briefCrowdOcc: Single-camera scene completion for quadruped robots in crowded indoor spaces
SummaryThis paper addresses monocular semantic scene completion for quadruped robots in real crowded indoor settings, where people occluding the scene disrupt geometry and human occupancy predictions. The authors release CrowdOcc, an RGB-D dataset of 25.1K frames from 11 indoor scenes, and a framework combining normal-guided scene geometry fusion with human-centric sparse interaction modeling. On a scene-disjoint test set, the method reports 15.80 IoU, 11.40 mIoU, and 46.23 Human IoU, and it generalizes to unseen indoor scenes.
In briefSoftGene: Protein-Language-Model Soft Prompts for Explainable Gene Set Annotation
SummarySoftGene is a framework that uses large language models to annotate gene sets, encoding each set's protein amino acid sequences with a hierarchical attention encoder built on the ESM protein language model. It combines soft prompts from these embeddings with hard prompts containing LLM-generated context, then feeds the result into a local LLM. Tested on Gene Ontology and MSigDB, the integration of protein-sequence representations with text context improves overall annotation, though the benefit of protein embeddings varies by biological domain.
Drew Ross, Arya Hadizadeh Moghaddam, Dongjie Wang et al.
In briefTailoring quantization space to compress KV cache to 1 bit
SummaryLong-context LLM inference is limited by the memory and bandwidth demands of the key-value (KV) cache. This paper proposes TaSQ, which adjusts the vector quantization target space using query-guided channel weighting, cross-head normalization, and covariance-aware channel grouping. The authors report that TaSQ outperforms existing low-bit KV cache quantization baselines on general, long reasoning, and long-context retrieval benchmarks. On a single RTX 6000 Ada GPU, its SGLang implementation reportedly supports up to 14x larger batch sizes and 1.87x higher peak throughput than BF16.
In briefWhich parts of human preference can be checked or spelled out, and which stay tacit
SummaryThe paper argues that judgments such as a gripping short story, a newsworthy article, or an elegant proof are partly tacit and hard to articulate or verify. The authors introduce CreativePreferences, a dataset of 2.8M texts with 317M human preference judgments across 7 creative domains, and model the labels with executable programs, rubric banks, and trained models. They report articulability and verifiability gaps in every domain, including mathematics and software engineering, and find the gaps grow as more people take part in judgment. Articulating preferences shifts models away from the tacit dimension of human judgment.
WhyProposes a novel measurement of articulability and verifiability gaps with a 317M-judgment dataset, showing tacit preference persists even in verifiable domains; some metrics rely on modeling assumpti
Alexander Spangher, Sheldon S. Huang, Andreas Haupt et al.
In briefReSCUE: Real-time sign language translation for continuous, unsegmented long videos
SummaryThis paper presents ReSCUE, a framework for simultaneous sign language translation that works on continuous, unsegmented long-form videos rather than pre-cut sentences. It combines training designed for partial inputs, pauses, and multi-sentence context, a re-translation method that lowers latency while reducing output flicker, and a sentence commitment mechanism for online segmentation and memory management. On standard sentence-level benchmarks, ReSCUE achieves lower latency and the best translation quality under low-latency settings. On long-form unsegmented datasets, it approaches the quality of offline systems using ground-truth sentence boundaries while operating at substantially lower latency.
In briefDyadMem: A Long-Term Memory Benchmark for How AI Agents Work With Users
SummaryThis paper introduces DyadMem, a benchmark for testing whether long-term AI agents remember not only user facts and preferences but also how a specific agent should work with a specific user over time, defined as User-conditioned Relational Agent Memory (URAM). The dataset includes 3,065 episodes, 50,961 sessions, and 61,210 QA instances, evaluated in Gold-Memory and Full-Pipeline settings. Across 20 models, Full-Pipeline performance drops sharply compared with Gold-Memory, and the authors report low capture recall, incomplete recall, and unsafe deletion issues.
Yifei Tao, Xinyu Zhong, Henry Hengyuan Zhao et al.
In briefSame Output, Different Gold: How Reference Choice Shifts a Multilingual Benchmark Score
SummaryThe paper shows that benchmark scores depend on the reference answers as well as on the system output. Using the retained annotation record of a six-language PII benchmark, the authors hold outputs fixed and swap the reference, finding F1 shifts of several points and changes in comparisons between systems. They attribute this to an undocumented aggregation default that often kept one annotator's labels, so the shipped reference partly copied the output being scored. The authors argue that this reference sensitivity should be reported alongside the score.
In briefSimpler Signals Don't Mean Easier Forecasts: Why Neural Residual Models Fail in Short-Term Volatility Prediction
SummaryThe paper tests a hybrid approach that fits a statistical model first and then uses an LSTM to learn its residuals, using short-horizon volatility forecasting across five liquid U.S. assets. A HAR-style model outperforms AR, MA, and ARIMA, and its residuals show much lower variance and near-zero lag-1 autocorrelation. However, adding residual-only LSTM augmentation increases mean squared error on average, while pure LSTM achieves the lowest error. The authors call this pattern forecaster-preconditioner asymmetry.
In briefPersonalized speech recognition for a dysarthric, tracheostomized speaker using artificial conversations
SummaryThis paper presents a speech recognition system personalized for a Czech speaker with a permanent tracheal stoma and severe dysarthria. The authors release a public dataset of 33 annotated hours of the speaker's speech, collected with an artificial conversation protocol. They train Whisper Base in multiple stages, including simulated tracheostomic speech, and report a 50% relative reduction in character error rate over the baseline.
David Nadrchal, Monorama Swain, Florian Schmid et al.
In briefFrom Expression to Reaction: Role-aware Visual Transfer and Stimulus-guided Reasoning for Interlocutor Emotion Recognition
SummaryThis paper proposes RASG, a framework that predicts listener emotions using only listener videos and speaker audio. It combines Role-aware Visual Transfer, which trains a listener-focused visual model using reliable pseudo-labels, with Stimulus-guided Boundary Reasoning, which uses speaker audio and text as context when the visual model is uncertain. On the MER-Cross dataset, RASG achieves 76.25% and improves over the baseline by more than 17%. The team ranked second in Track 1 (MER-Cross) of the MER Grand Challenge at ACM MM 2026.
In briefOmniAct3D: Panoramic 3D Detection Using Foundation Geometry and Evidence-Grounded Reasoning
SummaryThis paper presents OmniAct3D, a framework that adapts 3D detectors built on Vision Foundation Models, originally trained on perspective images, to equirectangular panoramic images for full 360-degree scene perception. It includes a ray-based geometry adapter, a reasoning chain that grounds detections in panoramic evidence, and a higher-resolution module for heading estimation. The authors report gains of 2.96 NDS points over the previous best detector on Spheriverse and 24.87 mAP points over the unadapted baseline on PanoMMOcc. Source code is said to be publicly released.
In briefRYOPO: A real-time, end-to-end system that detects objects and estimates their 3D pose from RGB-D images.
SummaryThis paper presents RYOPO, an end-to-end trainable query-based RGB-D predictor that jointly detects, segments, and estimates the 9-DoF pose of objects without CAD shape priors or a separate instance segmentor. On the NOCS benchmark it substantially improves on published joint detection and pose results. It is competitive with two-stage methods on REAL275 and HouseCat6D under all-object evaluation, and runs full-frame pose estimation at 31.8 FPS on an RTX A6000.
In briefRethinking fixed time grids: frequency-separated human motion generation
SummaryMost human motion generators encode motion as tokens of identical duration on a uniform time grid, which mixes slow global trajectories with fast events such as foot contacts. The paper proposes FreqMo, which uses wavelets to split motion into frequency bands while keeping temporal localization and exact reconstruction. Its Unified Frequency Residual Quantization encodes all bands in one shared codebook, compressing the token sequence threefold. The authors report state-of-the-art fidelity with better high-frequency preservation, and note the decomposition also works with continuous diffusion backbones.
In briefDesigning Sustainable AI Agents: Comparing Agentic LLMs for Developer Tasks
SummaryThis paper empirically compares agentic large language model systems, from single-query baselines to multi-agent workflows, across five software engineering tasks. It measures accuracy, inference latency, and energy use across six open-weight models, two prompt strategies, and three hardware platforms. Multi-agent designs used on average 6.36 times the energy and ran 6.07 times longer than the baseline, while accuracy gains were limited and task-specific. Lightweight non-agentic and single-agent configurations accounted for 59 of the 66 Pareto-optimal configurations.
In briefSteering Sparse Autoencoder Features to Reduce Alignment Faking in Small LLMs
SummaryThis paper studies alignment faking in small language models, where a model behaves compliantly when it believes it is monitored, measured by a compliance gap. The authors propose a bag-of-words feature elicitation pipeline to find concept-level sparse autoencoder (SAE) steering directions tied to deontological ethics. Steering reduced the compliance gap to non-significance in Gemma-3-27B-IT but was not effective in LLaMA-3-8B-Instruct. The authors also describe a behaviour they call deceptive defanging, in which a model degrades a response's harmfulness while preserving the appearance of compliance.
In briefResponse Variability and Stability in Human Reasoning: A Geometric Study of Reasoning Patterns
SummaryThis paper examines whether people's reasoning response patterns stay stable within individuals and whether their variation can predict learning effects. The authors propose a geometry-based method to measure distances between individual reasoning patterns and their internal variability, grounded in heuristic theories. Tested on experimental data with mixed-effects models and clustering, the variation measure interacts with correctness to predict performance gains, and reasoning patterns remain stable over time within the same reasoner.
In briefAvoKV-E: Payload-Aware KV Cache Eviction for Long Reasoning
SummaryThis paper addresses the KV-cache bottleneck in long-output reasoning, where the generated trace grows large. It proposes AvoKV-E, a training-free eviction policy that delays eligibility for recent states and ranks candidates using normalized read pressure, key redundancy, and value-payload potential. Across models and datasets, it matches or exceeds baseline methods at equal active-KV budgets, with the largest gains when the cache is tightest.
In briefRecursive Self-Improvement for Models That Both Read and Draw
SummaryUnified multimodal models can understand and generate both text and images, so they can create their own training data. The paper proposes recursive cross-capability self-improvement (RSI), where the model generates images, finds its shortcomings, writes programs to address them, and verifies results through execution. On chart tasks, four rounds of RSI raised the score on differently worded requests from 45.7% to 60.2%, while continued training reached only 46.3%.
In briefNeural data needs semantic tokenization: behavior-event-bounded tokens that transfer across sessions
SummaryEach electrophysiology session records a different set of neurons, so existing neural foundation models that embed neurons and sessions fail on new sessions. This paper proposes TWS (Tokenization with States), which segments each trial at task events such as stimulus and movement onset and converts each state's population geometry into tokens, without neuron or session embeddings. On held-out International Brain Laboratory sessions, TWS decodes movement while a per-neuron foundation model performs at chance. Trained only on mice, it transfers to macaques and Utah arrays with a linear probe, reaching a Matthews correlation of 0.23 on reach direction.
WhyProposes event-segmented population-geometry states as session-transferable tokens, validated on IBL and macaque data with reasonable but limited scale.
In briefGeometric and temporal representations of neural network training to explain how networks self-organize over time
SummaryThis paper represents multilayer perceptrons and their training dynamics in hyperbolic (Poincaré ball) space. Instead of analyzing single checkpoints, the authors build temporal parameter graphs from snapshots taken during training and use a graph-based meta-learning framework to learn dynamic hyperbolic embeddings. The model is equivariant to neuron permutations within a snapshot and invariant to permutations of past snapshots. Experiments on regression and classification tasks report strong meta-network performance and hyperbolic temporal representations that show how network structure emerges over training.
In briefActive learning that targets treatment effects for multi-outcome preference comparisons
SummaryThis paper proposes an active learning method for estimating treatment effects across multiple clinical outcomes when the preference rule weighting those outcomes must be learned from costly expert comparisons. Instead of maximizing information about the preference rule itself, the method selects queries that most reduce uncertainty in the treatment effect and the optimal policy. Under a Gaussian process model, the authors derive a closed-form approximation of this criterion. On semi-synthetic data from a Parkinson's disease cohort with 13 clinical outcomes, it yields lower estimation error and policy regret than existing criteria at equal query budget.
Lola Giordani, Mathieu Even, Chlo\'e Geoffroy et al.
In briefOmniConfess: Token-level self-reports to reduce hallucinations in omni-modal AI
SummaryThis paper presents OmniConfess, a training-free inference-time method that reduces hallucinations in omni-modal large language models handling text, images, audio, and video. It re-scores a fixed candidate response at token level under controlled channel-wise evidence interventions, producing a structured confession that shows which evidence supports each commitment. The authors also build OmniHalluBench, a 3,540-example benchmark from six datasets, and report that the method mitigates hallucinations across modality and task settings. Code and benchmark are publicly released.
In briefSentry: Learning to recover from LLM agent failures at test time
SummaryThis paper addresses LLM agents failing mid-task due to invalid tool calls, repeated actions, or poorly grounded reasoning. It proposes Sentry, a failure-management layer that runs alongside the agent, retrieving matching lessons from an external playbook when a failure is detected. Sentry verifies recovery without task rewards and stores a new lesson only on success, keeping the full playbook out of the agent's context. The authors report outperforming the strongest baselines across several benchmarks and that learned lessons transfer to held-out tasks.
In briefDifferentiable Koopman Operator for Contrastive Learning on Dynamic Graphs
SummaryThe paper proposes KAIROS, a self-supervised framework that embeds a differentiable Koopman operator in dynamic graph contrastive learning to linearize how node representations evolve over time. It uses a dual-view encoder combining raw node features with a graph-diffused structural view. For anomaly detection, it combines Koopman prediction residuals with temporal inconsistency and local neighborhood deviation, and the authors report state-of-the-art results on nine dynamic graph benchmarks.
In briefOLMo-Detect: A Controlled Benchmark for Detecting Whether Text Was in an LLM's Training Data
SummaryThis paper introduces OLMo-Detect, a benchmark for membership inference, which asks whether a given text was part of a large language model's training data. Built on the fully open OLMo 2 pipeline, it covers pre-training, mid-training, and post-training, aligns members and non-members on three key factors, and filters non-members with infini-gram. Evaluating 15 unsupervised and 3 supervised attacks, the authors find overall performance is limited, with the best methods reaching an AUC of only 0.68.
In briefPLCWorld: A Closed-Loop Benchmark for Testing LLM-Written PLC Programs in Simulated Factories
SummaryThe paper introduces PLCWorld, a closed-loop environment and benchmark that runs Structured Text PLC programs against simulated plant responses and sensor feedback. It contains 100 synthetic tasks across Motion Control and Material Handling, and reports Task Success and Safety Violation separately. Evaluations of six LLMs and four generation-and-verification workflows reveal differences between completion, safety, and generation cost.
WhyBuilds a closed-loop PLC benchmark with validated evaluators separating task success and safety violations, a substantive advance mainly relevant to industrial control.
In briefProtecting Mutual Correction in Source-Free Adaptation via Cut Statistics
SummarySource-Free Domain Adaptation adapts a pretrained model to an unlabeled target domain without access to the original source data. Prior methods that use Vision-Language models mostly supervise in one direction, overlooking that the two models make different errors and could correct each other. The paper proposes SafeCut, which uses the cut statistic as a label-free reliability measure to control, per sample, the direction and strength of cross-model supervision. The authors provide theory claiming this gating guarantees a net-positive correction signal and report state-of-the-art results across SFDA benchmarks.
In briefRASPER: Reward-aligned summarization of clinical notes to improve EHR outcome prediction
SummaryThis paper proposes RASPER, a method that summarizes unstructured discharge notes from electronic health records to extract evidence relevant to a prediction task. It trains the summarizer with reinforcement learning, using a reward derived from the downstream predictor's loss, and adds patient context from structured codes. The authors report that RASPER consistently outperforms strong baselines on readmission prediction and medication recommendation using MIMIC-III and MIMIC-IV.
Arya Hadizadeh Moghaddam, Mohsen Nayebi Kerdabadi, Chen Chen et al.
In briefRelevant Evidence Decoding (RED): a training-free method reducing cross-modal hallucinations in audio-visual language models
SummaryAudio-visual large language models can wrongly let one modality distort predictions about another. This paper proposes Relevant Evidence Decoding (RED), a training-free method that first determines which evidence type a question needs, then uses pointwise mutual information to strengthen the audio or video contribution. Across three benchmarks and three models, RED improves accuracy over standard decoding by up to 7% on CMM, 6.3% on AVHBench, and 3.8% on SVHalluc. Its average time to first token is 1.5 times that of standard decoding.
In briefReliable self-evolving search using imperfect proxy rewards
SummaryThis paper addresses a problem in LLM-based self-evolving search: cheap proxy rewards used in place of costly evaluation can give high scores to infeasible candidates. The proposed method, CISE, builds candidate-specific reward intervals using conditional conformal inference and online density-ratio estimation, and returns only candidates whose required property intervals lie fully within feasible regions. On three materials science search tasks, all candidates returned by CISE were true positives under high-fidelity evaluation, while baselines returned more candidates that included false positives.
In briefFrom Language Priors to Field Adaptation: Preference Learning for Traversability Estimation
SummaryThis paper addresses image-based traversability estimation, which depends on the robot platform, deployment domain, and mission preferences. The method represents traversability with von Mises-Fisher mixture prototypes in a frozen vision-language feature space, using natural-language rules as a prior and sparse relative image annotations for efficient adaptation. Experiments on WayFAST report accuracy competitive with end-to-end trained estimators while requiring fewer target-domain annotations.
Simon Schwaiger, David Seyser, Alessandro Scherl et al.
In briefCreateScore: Cutting Costs of LLM CV Screening with Bayesian Routing
SummaryThis paper presents CreateScore, a Bayesian network that routes each CV screening criterion either to a local 8B language model or, when uncertain, to a 120B reference model. On 200 synthetic Data Science CVs, 77.7% of criterion decisions were resolved locally, cutting token use by 65.2% compared with using the 120B model for everything. The authors found that the uncertainty signal did not reliably identify the 8B model's errors, so the system is presented as an auditable cost-reduction tool rather than an autonomous hiring system.
In briefGAMA: a guideline-augmented multi-agent system for schema-based biomedical named entity recognition
SummaryThe paper presents GAMA, a multi-agent framework for biomedical named entity recognition that uses schema-as-code. It induces annotation rules from labeled training data, verifies them against annotations, and stores them as dataset-specific guideline memory. A planning component proposes ranked span-type hypotheses, a coding component converts them into schema-constrained entities, and a verification module checks grounding, type validity, and structure with dual-loop refinement. Experiments on three BioNER datasets with two LLMs show the best results in four of six combinations, and ablations support the value of its components.
In briefReasoning With Evidence, Not Just Rationales: Verifiable Preference Proofs for LLM Recommendation
SummaryThis paper observes that rationales generated by LLM recommenders may not reflect the information actually used, leaving a gap between grounding and influence. The authors propose PROVE-REC, which compresses interaction history into a preference proof linked to selected evidence, then predicts the next item using only that proof. Verification is done by masking evidence and removing preference claims to measure grounding and influence. Experiments on real-world datasets show gains over strong baselines of up to 7.45%.
In briefPost-training text-to-image models by combining preference and rubric rewards
SummaryThis paper presents a post-training recipe for text-to-image models that combines a preference reward, trained on large-scale human preference data with a Bradley-Terry objective, with rubric-based rewards that evaluate prompt faithfulness and guard against reward hacking. The authors find that a naive weighted average of these signals performs poorly and propose a reward composition strategy that better balances preference optimization with rubric satisfaction. On the Arena text-to-image leaderboard, their RL-trained Flux2dev scores 69 Elo points above its base model, and post-trained Ideogram-4 reaches 1223.5 Elo, surpassing all open-source models listed. They also release Arena-T2I-Training, a 1K training data subset.
Yuanhao Ban, I-Hung Hsu, Anastasios Angelopoulos et al.
In briefNeural model that finds how physical components connect, without simplifying them
SummaryThis paper proposes Dirac-interconnected neural elements (DINEs), a neural network model that represents a physical system as a differential-algebraic equation. Its algebraic constraints are given by a Dirac structure, which captures how components are interconnected. DINEs learn both this interconnection and the behavior of each component from data, keeping subsystems unreduced so they can be isolated or recombined without retraining. The authors state the model can also handle partially observable systems.
Reiho Li, Razmik Arman Khosrovian, Takaharu Yaguchi et al.
In briefTerraVis: Evaluating whether AI-generated images follow real-world logic
SummaryTerraVis is a framework for checking whether text-to-image generated pictures are consistent with the real world, such as having plausible object structures, anatomy, physical interactions, and spatial relations. It defines a taxonomy of violations at the object, interaction, and scene levels, and uses a multimodal large language model workflow to detect them and label each as minor or major. On two widely used benchmarks, TerraVis correlates most strongly with human judgments of world consistency among existing metrics. The authors also show that models scoring well on conventional metrics can still fail these consistency checks. Code is publicly available.
In briefTrajectory heterogeneity in federated world models for clinical prediction
SummaryThis paper benchmarks federated learning for world models using hourly action-conditioned clinical prediction on eight MIMIC-IV disease cohorts, comprising 40.87 million transition memberships. It compares ten federated algorithms across 32 disease-partition configurations. The findings indicate that client data ownership and participation jointly limit long-window coverage, and finer severity partitions generally accompany higher FedAvg error. Algorithm gains are small and depend on prediction horizon.
In briefChoosing hyperparameters for learning equations with biologically-informed neural networks
SummaryThis paper addresses hyperparameter selection for biologically-informed neural networks (BINNs), a subclass of physics-informed neural networks used to learn partial differential equation terms from data. The authors present a diagnostic workflow, usable without known ground-truth equations, guided by questions about network capacity, training epochs, and whether learned terms stabilize. Tested on synthetic systems with known ground truth, the validation loss generally tracks the true error in the learned terms. The work offers practical rules of thumb and suggested starting values.
William Lavery, Jodie A. Cochrane, John T. Nardini et al.
In briefSlimKV: a method that compresses the KV cache along both token and feature dimensions, with beacon attention that avoids reconstruction.
SummaryThe paper addresses the memory bottleneck of KV caches in long-context LLM serving. It proposes SlimKV, which jointly compresses tokens and features using low-rank-aware training that turns long contexts into beacon memory states with latent KV representations, plus layer-adaptive rank allocation. The authors also observe that removing key-side RoPE affects beacon tokens much less than raw tokens, and exploit this to run attention in latent space during decoding.
In briefGTDD: Generative test-driven development with adversarial testing for AI coding agents
SummaryThe paper proposes Generative Test-Driven Development (GTDD), in which a separate testing agent creates new test inputs after each candidate implementation is fixed, based on a human-specified behavioral contract and earlier feedback. A trusted evaluator checks these inputs, returns reduced counterexamples to the coding agent, and saves them for regression testing. In a paired experiment on a stateful key-value store, both policies that regenerated tests during development ended with lower mean failure rates than the policy whose tests were generated once.
In briefDynamic Expert Pruning (DEP) for Multi-Agent Systems: Generating Per-Request Expert Masks From Prompts
SummaryThis paper addresses the problem that static expert pruning of Mixture-of-Experts language models fits poorly in multi-agent systems, where one model serves many tasks and roles. The proposed method, Dynamic Expert Pruning (DEP), relies on the finding that an agent's system and task prompts alone identify the experts it needs. A lightweight predictor, trained once on workflow transcripts, produces a specialized mask for each request in a single forward pass. The authors report better overall accuracy than static pruning and merging baselines across tasks, roles, model scales, and architectures, with generalization to unseen workflows without retraining.
In briefImproving Biomedical Named Entity Recognition with Multi-Language Code Instruction Tuning and Ensembling
SummaryThis paper proposes MITE, a method for biomedical named entity recognition (BioNER) that expresses instructions and entity outputs in code format. Each training example is converted into Python, C++, and Java versions to provide structurally diverse supervision without external knowledge. At inference, predictions from the different code formats are combined through entity-level voting. Experiments on three widely used BioNER datasets show competitive performance against BERT-based and LLM-based baselines, with the best results on two benchmarks.
In briefWhen Empty Predictions Beat SAM 3: Rethinking Video Object Segmentation Evaluation
SummaryThis paper introduces FaVOS, a benchmark for video object segmentation where target objects appear only intermittently in long videos. The authors show that the standard J&F metric rewards empty predictions on frames where the target is absent, so a trivial empty-mask predictor can outperform SAM 3. They propose Volumetric J&F, which evaluates mask sequences as spatio-temporal volumes to reduce this bias while remaining sensitive to segmentation quality and temporal structure.
In briefContinual Graph Memory for Mathematical Research Agents: A Memory System That Organizes and Reuses Proof-Search Results
SummaryThe paper presents Ansatz, a mathematical research agent built on Continual Graph Memory, a graph-based system that stores intermediate results such as facts, plans, and counterexamples along with their relationships. The memory is reused across problems through dependency-aware retrieval and scoped recall of earlier findings. The authors report closure on all ten First Proof Second Batch problems, and state that Ansatz produced solutions to the Jamison caterpillar conjecture and Erdős Problems 289, 348, and 488 without human intervention, with partial progress on several open problems.
In briefMultimodal 3D Human Pose Estimation with Privacy Protection for Individuals
SummaryThis paper presents a framework for multimodal 3D human pose estimation that fuses RGB, LiDAR, and mmWave radar data, with skeletal structure injected to improve accuracy. It also formulates a black-box subject membership inference attack and analyzes the resulting privacy leakage. To train with user-level differential privacy, the authors propose Action Temporal Stratification and evaluate on the MM-Fi dataset across three protocols.
In briefFASTDIAR: A Frame-Level Speaker Encoder for Real-Time Streaming Speaker Diarization
SummaryThis paper presents a streaming speaker diarization system that identifies who is speaking when, designed to run in real time on a CPU. It converts a speaker recognition architecture into a causal frame-level encoder that emits an embedding every 80 ms from a two-second window of past audio, paired with online clustering gated by stream self-similarity. Trained only through distillation from an utterance-level teacher, it is reported to be the most accurate streaming diarizer on low-overlap benchmarks and to run five times faster than real time on a single CPU thread.
In briefRAIN: Region-aware networks helping robots generalize tasks to new objects and layouts
SummaryExisting vision-language-action models tend to memorize training trajectories and struggle to adapt learned interactions to new tasks and scenes. The paper proposes Region-Aware Interaction Networks (RAIN), which represents a task as goal-conditioned subtasks defined by an action type and a target-region mask, with components for cross-view encoding, trajectory prediction, and subtask transition. It also introduces a Reference-based Retargeting Strategy that synthesizes approaches to alternative target regions from recorded trajectories without extra data. Experiments on the new LIBERO-Analogy benchmark and real-world tabletop tasks show RAIN generalizes more robustly than existing VLA baselines.
In briefWhen Should a Computer-Use Agent Be Compiled? Measuring Payback for Token Savings
SummaryThis paper introduces PACE, a system that decides when to compile GUI procedures an agent repeatedly runs into programs to reduce token costs. It includes a measurement protocol that records compilation costs and compares agent and program execution costs, plus an online algorithm that uses past task arrivals and compilation outcomes to decide. In simulations on recorded task arrivals, PACE reduced token costs by 17.3% versus ReAct, 24.9% versus an AutoRPA adaptation, and 17.3% versus a ToolPro adaptation on average.
In briefTesting whether verification suites for AI agent infrastructure can actually catch bugs
SummaryThe paper examines whether invariant suites used to gate AI agent deployment are truly effective, since they are usually validated with only one passing and one failing case. The authors applied mutation analysis to a suite for a multi-session agent state-projection layer and found that a mutant removing event-identity deduplication survived all checks. After repairing and freezing a twelve-check suite, an independently designed set of ten mutants killed only five. The surviving mutants failed either because no fixture exercised them or because their corrupted internal state was unobservable, and adding fixtures for the missing dimensions killed all five.
In briefOutput Language Confusion from Foreign-Language Distractors in LLMs
SummaryThe paper introduces Multilingual Distractor Interference (MDI), an evaluation protocol that precedes factual questions with an irrelevant foreign-language sentence, tested on five instruction-tuned LLMs across TruthfulQA and TriviaQA. For Llama-3.1-8B with a Hindi distractor, 58% of responses switched to Devanagari, giving a raw hallucination proxy of 0.710, but manual review lowered the adjusted semantic hallucination rate to 0.470. The authors conclude that exact-match hallucination rates in mixed-language settings should be split into script-switching and semantic error components before judging model reliability.
Riju Marwah, Ritvik Garimella, Khusham Bansal et al.
In briefPositive-Unlabeled Learning for Auditing False Alarms in Agent Safety Monitors
SummarySafety monitors for language-model agents often raise many false alarms, which costs review effort and reduces trust. This paper frames false-alarm auditing as a positive-unlabeled ranking problem and proposes a two-stage framework that needs no alarm safety labels and leaves the monitor unchanged. Across mainstream safety monitors it reaches a macro AUPRC of 0.6444, beating eight PU baselines by 5.27–16.98 points. Compared with PULDA, the strongest baseline, it recovers 33.3% more false alarms at a 5% review budget.
In briefELMS: LLM agents with Monte Carlo tree search for protein motif scaffolding
SummaryThis paper introduces ELMS, a search framework for protein motif scaffolding that feeds structural evaluator results back into the design process instead of using them only as a final filter. A Critic Agent diagnoses failures in each scaffold state, a Policy Agent selects targeted operators, and Monte Carlo Tree Search decides which historical states to refine further. On the GeomMotif benchmark, ELMS reports success rates of 86.41% for single-motif and 84.57% for paired-motif tasks. On MotifBench, it solves 26.7 of 30 tasks on average (88.89%) under a matched 100-candidate budget.
In briefHASTE: Automatically Evolving Agent Safety Harnesses from Sparse Threat Evidence
SummaryThis paper introduces HASTE, a multi-agent framework that automatically updates agent harnesses, the safety constraint layers that block unsafe actions, using only sparse evidence such as brief threat report descriptions or a few attack examples. It works through an adversarial loop in which safety specifications guide harness updates and generated attack cases test for remaining weaknesses, with evaluation results fed back into both steps. Experiments across multiple backbone models, attack types, and evidence forms show that HASTE consistently lowers attack success rates while preserving performance on benign tasks.
In briefLearning Jazz Pianist Style via Cross-Attention Conditioning
SummaryThe authors use a pretrained symbolic music transformer to show that its representations can identify jazz pianists with high accuracy. They add cross-attention over learned pianist embeddings so the model can generate music in a specific artist's style. Evaluations report that generated continuations are classified as the correct artist and that the classifier can locate the most characteristic moments in a performance.
In briefCustom Forcing: Training-free subject customization for autoregressive video generation
SummaryThis paper presents Custom Forcing, a training-free method that lets a frozen autoregressive video model generate specific subjects from user-provided images instead of generic ones. It stores reference anchor frames in the model's persistent KV cache and adds drift-adaptive value amplification and anchor contrast guidance to keep identity stable. The authors report that identity holds steady over two-minute rollouts and that the method generates frames much faster than the long-video baselines it is compared against.
In briefChecking the Test Setup: Why Stale-Data RL Comparisons in Language Models Can Mislead
SummaryThis paper argues that details of the experimental setup can reverse which training method appears better when language models learn from outdated samples. The authors introduce PTH (Probe The Harness), a checklist for exposing such setup details, and compare SAN with truncated importance sampling (TIS) on verl and a single-GPU trainer. After the setup was checked, TIS matched SAN on verl, while on the trainer TIS learned steadily but SAN kept a margin.
In briefConstraint-Aware Training: Moving Generation-Time Program Analysis Out of the Training Objective
SummaryThis paper addresses language models that generate programs with constrained decoding, which uses program analyses to exclude tokens violating syntax, scope, or typing rules. Standard training duplicates this work by teaching the model to suppress the same tokens. The authors propose a general constraint-aware objective that externalizes these analyses from training and formalize its benefits in three theorems on model size and data efficiency. In a controlled synthetic experiment, the approach achieves lower prediction loss than ordinary cross-entropy training at matched parameter count and data.
