In briefHow Far Can Parameter Count Drop? Fine-Tuning LLMs with Bayesian Optimization (LlMBO)
SummaryThe paper builds on TinyLoRA, a method that fine-tunes large language models with very few parameters. The authors argue that in such low-dimensional settings, global optimization becomes practical, and they propose LlMBO, a gradient-free Bayesian optimization approach. According to the paper, LlMBO is sample-efficient, memory-light, and can directly optimize greedy-decoding accuracy. On math benchmarks, it matches or exceeds GRPO's fine-tuning gains while using a fraction of the compute and memory.
Aditya Kulkarni, Davide Sartor, Carlos Soto et al.
In briefEgoOmniWorld: A Benchmark Unifying First-Person Video Understanding Through Ego-Action-World Interactions
SummaryThis paper proposes an Ego-Action-World loop describing how a wearer perceives, acts on, and receives feedback from the world, and builds a benchmark, EgoOmniWorld, around four interaction structures. The dataset of 24,166 QA groups is built from egocentric videos using visual and auditory evidence. Experiments show that current multimodal large language models have uneven limitations across these structures, while models trained on the dataset improve in-domain and generalize to existing egocentric benchmarks.
In briefContext-Aware Faithfulness Evaluation for Theory-Scale Autoformalization
SummaryThis paper frames faithfulness evaluation in autoformalization as a context-aware align-then-label task, where phrase-level correspondences between informal and formal statements are identified and labeled. The authors introduce FaithAlign, a 460-item benchmark with 4,982 annotated phrase-level links, and FaithAlign-Gap, a 146-item companion set with context and assumptions withheld. Across 8 models, the best defect F1 is 47.8%, and on FaithAlign-Gap the top model flags only 28.2% of undecidable items as such.
In briefJudges Follow the Reference: The Derivation Gap in Verifying Tool-Call Arguments
SummaryThis paper examines how judge models that evaluate argument values proposed by LLM agents for tool calls accept correct values less often when their input lacks a reference stating what the value should be. The authors call this drop in acceptance the derivation gap, which appears across all seven judges tested and is not just a truncation artifact. They introduce DeRef, where a separate model reads the full history and writes a reference for the judge, which improves argument AUROC across three benchmarks.
In briefEye movements during tasks as possible digital markers of depression severity in teens and young adults
SummaryThis single-center cross-sectional study of 155 people aged 14 to 20 grouped participants by Hamilton Depression Rating Scale scores while they completed emotional-spatial free-viewing and memory-guided recall tasks. Greater depression severity was linked to lower gaze-allocation entropy, fewer switches between areas, longer delay before first reaching targets, and lower time spent on targets, after adjustment. The authors suggest altered visual exploration and memory-guided gaze control, rather than generalized negative attentional bias, as candidate biomarkers, and call for prospective evaluation.
In briefRecalibrating many prediction questions with few labels: minimax rates and identification limits
SummaryThis paper studies how to share information and recalibrate probability predictions across many related yes/no questions that each have only a few labels. The authors derive minimax compound risk rates under fixed-effects and random-effects settings. They also analyze identification regret, showing that beyond a few labels per question, what cannot be learned is nearly unnecessary.
WhySharp minimax rates and identification limits for recalibration across questions, generalizing a known coin bound to general maps. Strong theory with complete-looking proofs, but experimental validati
In briefRethinking Membership Inference Attacks on Tabular Machine Learning Models (Systematization of Knowledge)
SummaryThis paper systematizes membership inference attacks (MIAs) on tabular data, proposing a new taxonomy and comparing four state-of-the-art attacks across diverse tabular datasets and model architectures. It finds that datasets with higher class granularity and majority class percentage are more vulnerable. Advanced attacks depend heavily on adversarial knowledge of the target model, while reference model-free attacks often perform comparably. Records with unique signatures remain highly vulnerable even under common defenses, which reduce average leakage but do not adequately protect the most sensitive individuals.
Cristina P\^era, T\^ania Carvalho, Maxime Cordy et al.
In briefPhysics-aware deepfake detection using distance-speech consistency
SummaryThis paper proposes a deepfake detector for dynamic talking videos that uses an acoustic constraint: in real recordings, speech energy changes predictably with speaker-to-camera distance. The method compares distance estimates from video with distance-dependent acoustic measures from speech. Experiments on in-the-wild benchmarks show the cue works in dynamic scenarios and complements lip-sync-based detectors, with a simple ensemble yielding consistent gains across datasets.
In briefReTerm: Recovery Training for Self-Improving Terminal Agents
SummaryThis paper presents ReTerm, a terminal agent that improves by learning to recover from its own failed terminal tasks. The authors build the ReTerm-10K dataset of recovery demonstrations for supervised cold-start training and propose ReTerm-GRPO, an online reinforcement learning method that forks from failure states to run independent self-recovery attempts. Built on Qwen3.5-9B, ReTerm-9B reaches 34.8% Pass@1 on Terminal-Bench 2.0, improving over the base model.
In briefProTok: Unified protein tokenization using modular multi-codebook quantization
SummaryThis paper presents ProTok, a tokenizer that maps protein sequence, structure, and functional annotations into one discrete vocabulary. A single checkpoint can run inference from any subset of these modalities without retraining. On StructTokenBench, ProTok ranks highest on average among structure-only tokenizers while still reconstructing structures well. The authors also report that a language model trained on its tokens can generate backbones and sequences jointly.
In briefMemoryVLN: Memory-based navigation that reuses past spatial experience for robots following language
SummaryThis paper introduces MemoryVLN, a memory-driven system for vision-language navigation that helps large vision-language models handle long navigation paths. It compresses history with forgetting-curve-inspired sampling and stores memories with 2D semantic, 3D geometric, and 6D pose tokens. A hierarchical memory bank and Top-K retrieval let the agent reuse relevant past experience. On R2R-CE and RxR-CE, the authors report success-rate gains of 2% and 6.5% over state-of-the-art baselines.
In briefSimplified On-Policy Distillation for Faster Diffusion Generation
SummaryThis paper proposes Simplified Matching Distillation (SMD), a method that reduces the sampling steps of diffusion and flow models using only on-policy velocity regression, without auxiliary score models or adversarial training. It introduces Pivot, a training-only residual proxy that separates local corrections from the student's interval-mean prediction, and repels the student from weaker teacher predictions. The authors report that at 4 steps SMD outperforms strong few-step baselines and surpasses multi-step teachers on several benchmarks across image, 3D, video, and audio-video generation, scaling to 33 billion parameters.
In briefGraph metanetworks that respect the symmetries of transformer language models
SummaryThe paper studies weight-space learning, where neural network weights are treated as data and processed by other models called metanetworks. It notes that transformer language models have richer weight symmetries than simpler architectures, including continuous linear transformations of attention matrices and permutations of whole heads, which existing methods do not fully handle. The authors propose a symmetry-aware graph-based metanetwork that works across different depths and widths, and report that it outperforms existing baselines on two new transformer-population datasets for tasks such as training-corpus prediction and fine-tuning detection.
Yonathan Wolanowsky, Guy Bar-Shalom, Fabrizio Frasca et al.
In briefHow Safe Is ChatGPT Health Advice? Real User Conversations in Four Global South Countries
SummaryThis paper evaluates the safety of health advice from ChatGPT using real, consent-based user conversations from India, Nigeria, Brazil, and Pakistan. Of 202,590 conversations from 1,252 users, 15,555 were health-related, and roughly one in five was rated high-risk. Risks clustered in symptom and condition queries, including unsafe medication guidance and missing safety warnings. Including users' own context raised high-risk ratings by 13 to 17 percentage points.
WhyFirst demographically grounded safety evaluation using real health conversations, showing user context substantially raises high-risk rates across a large corpus.
In briefHow readout scaling affects recovery of sparse features
SummaryThis paper studies how readout scaling changes what sparse autoencoders (SAEs) learn. The authors propose readout-canonicalized decision-space (RC-DS) coordinates, which remove singular-value gains while keeping retained decision information. Recovery of known factors is highest when training restores the space where the factors were generated. At condition number 100, RC-DS raised mean sparse-factor recovery from 0.506 to 0.913 compared with centered-logit training. On three language-model checkpoints, RC-DS had higher gain-free reconstruction in all 18 settings, while centered-logit SAEs reconstructed output logits better in 17.
In briefMFVSR: A Multi-Style Benchmark for Evaluating Video Subtitle Removal
SummaryThis paper introduces MFVSR, a large-scale benchmark for evaluating video subtitle removal methods. It includes 3,000 paired videos from high-quality real, low-quality real, and AI-generated sources, plus 953 in-the-wild real videos, with subtitles spanning 16 languages and varied layouts, styles, and timing. The benchmark uses a protocol that separates blind removal from mask-guided removal and measures restoration, preservation, temporal fidelity, coverage, and failures. Experiments on six methods show that conventional full-frame metrics hide differences between local restoration, outside-region preservation, and temporal consistency.
In briefTeacher-guided tool learning for small language models using two-way KL constraints
SummaryThis paper addresses how small language models learn to use external tools, a task that reinforcement learning alone struggles with and direct teacher supervision can distort. The authors propose SeTPO, which adds token-level teacher feedback derived from reverse KL to low-reward trajectories and uses forward KL supervision to keep teacher-demonstrated behavior. Experiments on BFCL v4 and API Bank report that SeTPO outperforms supervised fine-tuning, reinforcement learning, and distillation baselines across three student model scales.
In briefS³-PnP: Self-Supervised Score Priors for Plug-and-Play Denoising in Scientific Imaging
SummaryThis paper proposes S³-PnP, a plug-and-play restoration framework for scientific imaging such as hyperspectral and MRI data, where clean training targets are hard to obtain. It learns a score prior directly from observation-derived ADMM intermediate states, without clean targets or external pretrained weights, and uses it to correct the input to the analytical prior update. The authors analyze convergence and reconstruction error under an ideal score and quadratic data fidelity, and report consistent improvements over competing methods on reference-guided HSI restoration, single-image HSI denoising, and MRI denoising.
In briefLearning cell perturbation-specific representations via conditional energy priors for response prediction
SummaryThis paper proposes PerturbEnergy, a method for predicting cellular responses to perturbations such as drugs, aimed at accelerating drug discovery. It adds a conditional energy-based prior over representations, assigning lower energy to those compatible with a given perturbation, and can be integrated into autoencoders. Experiments on four benchmark datasets report improved perturbation recognition in latent space (up to 83.2%) and state-of-the-art prediction of top-50/100/200 differentially expressed genes.
WhyIntroduces conditional energy prior with a proved mutual-information bound, validated on four benchmarks, a substantive advance for perturbation prediction.
In briefEmoGenie: A Benchmark for Emotions Expressed and Evoked by Text-to-Image Generation and MLLMs
SummaryThis paper introduces EmoGenie, a benchmark for testing how well text-to-image models express intended emotions in generated images and how well multimodal large language models understand those emotions. It includes 6,400 images from eight models, with question-answer pairs and human ratings of valence, arousal, dominance, and aesthetic quality. The authors also propose EmoProbe, which combines emotion question answering with regression heads for continuous score prediction, and report that it performs better on emotion QA and correlates more closely with human ratings.
In briefFastEEP: Fast in-instance scaling for expert-parallel MoE model serving
SummaryThis paper presents FastEEP, a system for quickly resizing the expert-parallel group inside an existing Mixture-of-Experts model serving instance, rather than tearing down and rebuilding the whole instance. It is based on a Partial Reuse principle that keeps reusable runtime state and separates it from resources tied to the current expert-parallel configuration. According to the authors' evaluation, FastEEP scales 4.5x to 11.7x faster than an existing vLLM-based approach across different models.
In briefAutonomous Science Bench: Testing AI Agents on Self-Directed Scientific Discovery
SummaryThe paper introduces Autonomous Science Bench, an initial benchmark of six budgeted design-build-test-learn campaigns drawn from published biology, chemistry, and materials science experiments. Agents choose experimental batches, receive lab measurements, and deliver a predictive model, which must pass both a discovery gate and a learning gate. In a pilot of eight agent-model configurations over 144 trials, the best configuration succeeded in 67% of trials.
In briefThinker-WM: A Unified World Model Sharing Multimodal Representations for Embodied AI
SummaryThis paper presents Thinker-WM, a world model that jointly generates RGB images, metric depth, and robot actions through a shared Unified Dynamics Transformer. It supports several functions, such as video-action generation, policy inference, and inverse dynamics. The authors report that it ranks first on EWMBench, WorldOlympiad, and WorldArena. Controlled studies examine the benefits of action conditioning, geometric interaction, and multi-function training.
Thinker Team Ubtech: (Alphabetical order), Peng Gao, Jichao Jiao et al.
In briefCoTL: Learning Compact Tokens for 3D Scene Representation
SummaryThis paper presents CoTL, a framework that learns a compact set of tokens from unposed multi-view images to preserve scene geometry and appearance. An encoder compresses each view into small latent token sets that are aggregated across views, and a query-based Gaussian decoder recovers renderable 3D Gaussian primitives, trained end to end with novel-view rendering supervision. Experiments on RealEstate10K and ACID show competitive novel-view synthesis, and the token space can serve as the latent space for generative 3D Gaussian scene generation.
In briefThe Singularity Has a Distribution: Exact Timing Laws and Detection Limits for a Stochastic Intelligence Explosion
SummaryThis paper models the feedback loop from automating AI research as a stochastic process, where independent multiplicative improvements arrive at a rate proportional to a power of current capability. It shows that randomness keeps the deterministic explosion threshold unchanged but gives the blow-up time an explicit distribution, which is exactly Gamma for Lomax-distributed improvement sizes. The authors also derive minimum sample size asymptotics for detecting increasing idea-arrival rates and extensions on delays and effort allocation. The numerical illustrations are not empirical forecasts.
WhyDerives exact distributions, tail behaviour and asymptotic detection sample sizes for a stochastic intelligence-explosion feedback model, with rigorous theory but no empirical validation.
In briefCopying Is Not Grounding: Rewarding Authored Sentences for Faithful Long-Form RAG
SummaryThis paper finds that faithfulness rewards for long-form RAG answers can be gamed by copying sentences from source passages, so models learn to copy rather than ground their writing. The authors propose Authored-Sentence Credit (ASC), which moves GRPO's reward and advantage to the sentence level and credits only authored, non-copied sentences that are supported by the passages. Across four models and several corpora, ASC reduces the error rate of authored sentences and produces more supported ones.
WhyProposes sentence-level authored-credit reward and advantage fixing copying hacks; consistent gains across four models and five corpora support the claim.
In briefEstimating how unexpected speech is from the audio signal alone
SummaryThis paper estimates surprisal, a measure of how unexpected a linguistic unit is, directly from speech using a speech language model, avoiding text transcriptions. To handle the many acoustic realizations of one abstract unit, the authors use an importance sampling estimator that marginalizes over speech generated by a text-to-speech model. Using pupillometry data from a listening study, they find that word-level surprisal from importance sampling predicts pupil dilation better than realization-specific surprisal. Overall, surprisal is not a strong predictor of pupil size change.
WhyEstimates surprisal directly from speech via importance sampling over realizations, validated against pupillometry at multiple granularities; novel with reasonably strong evidence.
Zachary Hopton, Samuel Kiegeland, Francesco Ignazio Re et al.
In briefSTAMP: a topology-aware molecular language for editing molecular fragments
SummaryThe paper introduces STAMP, a molecular language that separately encodes motif identity, attachment sites and tree topology so that fragments can be reused and subtrees edited directly. Compared with a matched SMILES encoder, it gave lower activity-cliff prediction error on 23 of 30 targets. Its subtree operations ranked first among six methods on 15 of 23 scaffold-preserving optimization benchmarks. A deterministic compiler also let frozen language models act on indexed fragments, improving several tasks by 9 to 56 percentage points over plain SMILES.
WhyProposes STAMP, a fragment-explicit molecular language that lowers activity-cliff error on 23 of 30 targets and ranks first on 15 of 23 optimization objectives, with solid benchmark comparisons.
In briefPrincipled Design of Expected Improvement for Simulation Optimization
SummaryThis paper studies expected improvement based on stochastic kriging surrogates for continuous simulation optimization with unknown heteroscedastic noise. It derives a high-probability upper bound on simple regret and uses it to choose the improvement incumbent, replication allocation, and final recommendation within general posterior-quantile rule classes. Under mild kernel conditions the replication order is logarithmic, and the resulting simple regret matches known-noise results up to logarithmic factors.
WhyDerives a high-probability simple-regret bound for EI under heteroscedastic noise, systematically tuning three design choices; first guarantee of this form, though limited to specific kernels.
In briefReliable Memory from Unstable Growth: Randomized Experiments for Selecting Agent-Written Playbooks
SummaryThis paper studies how language-model agents store lessons from past tasks as a textual playbook to reuse on later tasks. The authors find that the same procedure can produce playbooks whose effect ranges widely, up to 69.6 points worse than having no playbook. They propose Randomized Block Curation, which randomly includes or leaves out groups of related lessons on practice tasks and keeps the blocks whose random inclusion measurably raises success. Across two memory systems, three models and three benchmarks, it reaches an average success of 66.9, above the full playbook, no playbook, and agent-judged selection.
WhyRandomized block experiments measure each lesson group's causal effect, replacing self-judgment, validated across two memory systems, three models and three benchmarks.
In briefCALM: Decentralized Control of Cooperative Aerial Manipulation with Multi-Agent Reinforcement Learning
SummaryThe paper presents CALM, a controller that lets several drones carry a basket together while tracking a moving ground robot, despite wind, payload motion, and noisy goal signals. It uses a shared multi-agent reinforcement learning policy trained with MAPPO, combined with goal estimation, per-drone state observers, and anticipatory formation-yaw feedforward. On held-out trajectories, CALM achieves lower mean position and orientation errors than all evaluated baselines.
Jiaxi Huo, Petre Ricioppo, Fausto Francesco Lizzio
In briefReceiver Decides Outcomes in Agent-to-Agent Negotiations
SummaryResearchers ran 1,049 round-robin negotiations among fifteen language-model agents, where a requester tried to raise money from a holder given a $100 endowment. The holder's identity explained 47% of the variance in amounts transferred, while the requester's identity explained only 2%. Training the holder side, such as persuasion fine-tuning, changed how much it gave away, which the authors argue evaluations should account for.
Fardeen Sadab, Alif Hasnain, Afsana Shehrin et al.
In briefAgaram 2.5B: a Tamil-first language model trained from scratch with tool calling, plus a playbook for under-represented languages
SummaryThe paper presents Agaram 2.5B, a 2.5-billion-parameter Tamil-English model trained from scratch on about 50 billion tokens over roughly six days on one eight-GPU node. After supervised fine-tuning, it replies in Tamil script, romanized Tamil and English, and picks the right tool in 23 of 25 test prompts. It beats Sarvam-1 on translation benchmarks but scores near chance on knowledge tests, with 46% of replies rated usable on 50 everyday questions.
In briefLow-cost method to estimate 3D shape and heading of road users from roadside camera and radar
SummaryThis paper proposes a resource-efficient two-stage method for real-time 3D shape and orientation estimation of road users using a roadside RGB camera and 3D radar. The first stage uses an off-the-shelf method to extract 2D camera boxes and matching radar points. The second stage uses radar for position, predefined templates for size, and a tiny neural network to predict a site-specific yaw angle. The approach was validated at a curved road with a junction for pedestrians, bicycles, and cars, running at 25 frames per second on an NVIDIA Jetson ORIN.
Shiva Agrawal, Savankumar Bhanderi, and Gordon Elger
In briefDo Factuality Methods Really Make LLMs More Factual? A Reproducibility Check
SummaryThis paper independently reproduces five methods that aim to improve the factuality of large language models, including on-policy reinforcement learning methods (PPO, GRPO, Dr. GRPO) and offline preference optimization methods (DPO, KTO). It compares them on a common benchmark with shared evaluation metrics. The authors find that none of the methods fully reproduces its original results, and none significantly outperforms the base model. They attribute this partly to rewards that favor format, length, or brevity over factuality and relevance.
In briefRethinking how word-frequency surprisal is estimated for reading-time research
SummaryThe paper examines two predictors of reading time: contextual surprisal, usually from neural language models, and unigram surprisal, usually from independent corpus counts. The authors argue that deriving them separately breaks their probabilistic link and may add estimation error. They propose three estimation principles, including deriving both from the same language model, and introduce a Monte Carlo estimator with a lower-variance marginalized variant.
Zachary Hopton, Francesco Ignazio Re, Samuel Kiegeland et al.
In briefNot all drift is a mechanism: how prompt adaptation changes LLM internals
SummaryThis paper compares three ways to adapt a frozen language model with prompts: discrete prompt optimization, prompt tuning, and prefix tuning. It finds that where the adaptation is applied determines how task information reaches the output, and that in mixture-of-experts models these methods also produce different expert-routing patterns. The authors report that common representation-level measurements respond strongly to generic context changes and do not reliably separate useful adaptations from matched placebos. They argue that evaluating prompt adaptation requires isolating the part of representation drift that is aligned with and causally used by the model's task computation.
Mikhail Aleksandrov, Nail Bashirov, Maksim Ivanov et al.
In briefGraph-free retrieval using graph-modified embeddings
SummaryThe paper studies how to improve vector retrieval without querying a graph at runtime, by using offline graph-based embedding transformations. It compares five such transformations and strategies for combining the modified embeddings with the originals. On a synthetic professional-network benchmark, direct replacement of the original embeddings reduced evidence recovery, while multi-stage strategies that use modified embeddings to guide search and originals for relevance improved it. The best configuration improved recovery by up to 25.9% over plain vector retrieval and matched baseline recovery with 40% fewer returned nodes.
Federica Santi (1), Vasilije Markovi\'{c} (1), Daniel Scott (1) et al.
In briefInjection-detection probes may be learning conversation features, not attacks
SummaryThis paper examines linear probes used to detect indirect prompt injection in agent tool outputs. The authors show that conversation position and text length alone reach 0.91-0.98 AUROC on standard evaluations, suggesting probes may not truly identify injections. They add matched harmless negatives to the training data, and across six open models this sharply reduces false alarms while keeping most attacks detected.
In briefMulti-stream residual computation in diffusion language models: keeping visible and masked-token information in separate channels
SummaryThis paper equips masked diffusion language models with multiple residual streams and adaptive routing between them, so visible evidence and masked-position predictions can persist through depth. With two streams, reconstruction perplexity drops 5.71% on validation data and 5.13% on average across external corpora compared with standard residual connections. The authors report that visible and masked positions learn distinct routing patterns without predefined roles, and causal interventions link this distinction to read and write routing.
Ekaterina Alimaskina, Suren Kazarian, Gleb Molodtsov et al.
In briefPreventing Forgetting When Radar Learns New Waveforms
SummaryThis paper studies how radar waveform classifiers can learn new waveform types without losing accuracy on old ones, a problem called catastrophic forgetting. It presents the first large-scale benchmark of eighteen continual learning methods across 54 waveforms in ten sequential tasks. Methods that keep data from earlier tasks performed best, and combining statistically validated mechanisms raised the leading F1 score by 2.7-5.5%.
Mohamed Rabie, Chinthana J. Panagamuwa, Konstantinos G. Kyriakopoulos
In briefBifocal: Training-free pruning of visual tokens in multimodal LLMs using mixed granularity
SummaryThis paper examines visual token pruning in multimodal large language models and finds that methods performing well on general VQA often degrade badly on fine-grained tasks such as OCR and visual grounding. The authors propose Bifocal, a training-free two-stage method that allocates separate token budgets to query-focused and image-intrinsic evidence and adds a pooled coarse view of the whole image. According to the paper, with about a quarter of the visual tokens, Bifocal is the only evaluated method that keeps over 90% of unpruned performance in every ability group across all three models.
WhyProposes training-free mixed-granularity two-stage visual token pruning, validated against ten methods on three MLLM families; impact mainly within efficient multimodal inference.
In briefLunaTP: An Efficient, Scalable Tensor Product Method for Equivariant Machine-Learning Interatomic Potentials
SummaryEquivariant machine-learning interatomic potentials aim to make atomic simulations approach quantum-mechanical accuracy, but their Clebsch–Gordan tensor product convolutions are computationally expensive. This paper introduces LunaTP, which uses node-centric execution, path-group parallelism, and multi-level reuse to speed up these convolutions and higher-order derivative computations without changing the underlying mathematics. On an NVIDIA H100 GPU, LunaTP reports kernel speedups of up to 250.0× over e3nn and 17.1× over OpenEquivariance, and end-to-end training throughput gains of up to 28.7× and 3.11×, respectively.
In briefBeyond RMSE and Granger Gain: A New Entropy Measure for Detecting Hidden Nonlinear Links Between Time Series
SummaryThis paper asks whether an auxiliary time series helps predict a target beyond the target's own past, and proposes the forecast-error distribution entropy (FEDE) framework to test this. In controlled nonlinear benchmarks, LightGBM recovers hidden source-target dependencies that Ridge regression misses, and FEDE shows higher response sensitivity than mean log-RMSE gain. FEDE stays positively aligned across all 108 tested combinations of window length, history length, and maximum forecast horizon. Code is said to be available on Zenodo.
In briefPRISM: Correcting Black-Box LLM Judges Using Counterfactual Probes
SummaryThis paper studies whether large language models used as evaluators can be corrected for systematic disagreement with human judgments. It introduces PRISM, which keeps the evaluated content fixed, measures score changes when attributes are emphasized or de-emphasized, and uses ridge regression on human labels to learn a correction to the baseline score without rubric criteria. On SummEval and HANNA, PRISM significantly improves agreement with human rankings over uncalibrated scores and remains competitive with rubric-based methods.
In briefReDiS: A Reasoning-Diagnosis Self-Evolution Framework for Multimodal Remote Sensing Reasoning
SummaryThe paper proposes ReDiS, a framework that separates reasoning exploration from candidate evaluation, with a Reasoner generating visual evidence and answer candidates and a Diagnoser assessing the full candidate set. The two experts evolve jointly, with the Diagnoser adapting to new candidate distributions. For groups where all sampled answers are wrong, the authors introduce Failure-Triggered Recovery, which gives minimal outcome-level correction. Experiments on LHRS-Bench, MME-RS, and CHOICE show gains over Qwen3.5 backbones at 0.8B, 2B, and 4B scales, with a 6.47-point category-averaged accuracy improvement at 0.8B.
In briefNoise2Match: Self-Supervised Low-Dose CT Denoising via Adaptive Inter-Slice Matching
SummaryThis paper proposes Noise2Match, a self-supervised method for denoising low-dose CT images that trains on adjacent CT slices instead of requiring paired noisy and clean images. It uses an inter-slice matching network to model label transitions between neighboring slices and corrects the denoiser's outputs to better satisfy the equal-mean condition. The authors prove that under idealized conditions this training is equivalent to supervised paired training, and they report it also works for low-dose cone-beam CT and outperforms existing image-domain self-supervised methods.
In briefUsing knowledge-guided AI agents to calibrate Earth science models
SummaryThis paper presents a general calibration harness (GCH) that lets language-model agents handle Earth science model calibration end to end, from defining parameter ranges and objectives to generating model-specific code. The harness freezes these choices before numerical search runs. Across six model families, knowledge-guided specifications in snow-model tests all executed on the first attempt, reaching validation NSE of 0.81 to 0.94, compared with 0.57 for defaults.
In briefBeyond Task Success: A Theory of Representation Reuse
SummaryThis paper studies whether a learned representation stays useful for future tasks beyond its source task, defining a task-relevant state for a given family of tasks. Experiments show that reuse improves when task-relevant distinctions are more accessible, and that reversible sample-dependent coordinate changes can sharply reduce it, with performance restored by inverting the change. Preserving a state containing the needed distinctions improves reuse on compatible held-out tasks, most clearly when target supervision is limited.
In briefSchemaPose: RGB-based category-level object pose estimation using a parametric category schema
SummaryThis paper addresses estimating the 3D translation, rotation, and size (9D pose) of objects within a category from a single RGB image. The authors propose SchemaPose, which adds an explicit parametric schema that models shared category regularities and structured instance variation. The inferred schema variables guide pose prediction in an end-to-end query-based framework. Experiments on synthetic and real data reportedly achieve state-of-the-art results among RGB-based methods.
In briefFaithRS: Hallucination-Aware Faithful Reasoning for Remote Sensing
SummaryThis paper addresses remote sensing multimodal language models that produce reasoning contradicting visual evidence. The authors build SFHD-43K, a dataset of 43K fine-grained hallucination annotations, and train a detector to identify hallucination types, causes, and supporting evidence. The FaithRS framework uses this feedback in a safety-gated corrective rollout, reporting gains of up to 11.64 points across four benchmarks.
In briefVisual servoing control using lines as features
SummaryThis paper proposes a new visual servoing approach that uses lines as features. Through a geometric study of main line representations, the authors define binormalized Plücker coordinates, a representation well suited to visual servoing that supports a proper notion of image line alignment. The control law partially decouples rotation and translation, allows analytical inversion of the motion equations, and has global asymptotic stability conditions. It was validated in simulation and experimentally for an orthogonal trihedron.
In briefUrbex: Agentic Spatial Grounding for City-Scale 3D Scenes
SummaryThis paper proposes Urbex, a vision-language agent framework that locates objects in city-scale 3D scenes by actively searching space rather than ranking precomputed candidates. The agent can query landmarks, zoom into regions, render oblique views, and output a bounding box, and it is trained with reinforcement learning on CityRefer. The authors report the best Hit@1 and IoU on CityRefer, candidate-free grounding, and faster inference than the strongest ranking baseline, with zero-shot results on STPLS3D-Refer suggesting better cross-scene robustness.
In briefWong–Zakai Convergence and FNO Learning for Stochastic Volterra Equations
SummaryThis paper studies stochastic Volterra integral equations, which model random systems with memory. It proves that pathwise solutions driven by piecewise linear approximations of Brownian motion converge to the corresponding solution. Building on this, mirror-padded Fourier neural operators are used to learn the mapping from initial conditions and noise to full solution trajectories. Experiments report competitive accuracy and generalization to finer temporal grids.
In briefABot-ER: Unified embodied reasoning via agentic data-model co-evolution
SummaryThis paper presents ABot-ER, a unified foundation model for embodied reasoning that aims to balance breadth and depth across perception, planning, and execution. It uses an agentic data-model self-evolution engine and a task-conditioned training strategy that activates lightweight low-rank residuals over a shared backbone. Across twelve benchmarks, the authors report state-of-the-art overall performance within a single parametric framework.
In briefChoosing LLMs for Strategic Agent Interactions: how to pick which language model to use in competitive game settings
SummaryThe paper argues that existing game-theoretic benchmarks test only single fixed LLMs, while model routers ignore strategic structure and opponents. The authors define strategic selection as a meta-game in which a selector picks an LLM each round to play a game with a latent, online-learned payoff matrix. Across 5 games and 10 LLMs, strategic competence varies with the game and opponent, and a lightweight multi-armed bandit performs most reliably, though it still trails the best fixed model.
Arastun Mammadli, Herbert Woisetschl\"ager, Jawad Fayaz et al.
In briefLet the Mind Conceive, Let the Hand Achieve: Robot Following of Multimodal Instructions for Long Tasks
SummaryThe paper introduces Orchestra, an agent framework that separates task planning and progress tracking from action generation, so robots can follow multi-step multimodal instructions in order. It also presents RMBench-Compose, a benchmark testing whether robots trained on short, atomic tasks can execute longer subtask sequences. The authors report a 68.22% mean success rate across nine RMBench tasks and an Average Success Length of 7.35 subtasks on RMBench-Compose.
In briefLightweight but Linguistically Fair? Quantisation Robustness in Bavarian, German and English
SummaryThis paper examines whether post-training quantisation, which reduces the memory footprint of large language models, preserves language ability in Bavarian, Standard German and English as reliably. Using a multiple-choice benchmark with 206 questions aligned across the three languages, the authors test 14 open-weight 7B–10B models under BF16, 8-bit GPTQ, and 4-bit GPTQ and AWQ. They find 8-bit GPTQ stays close to BF16, while 4-bit methods are less stable and the most affected language varies by model and method, with calibration and evaluation setting also shaping results.
Jophin John, Michael Hoffmann, Piroska Lendvai et al.
In briefCityCode: Building Agent-Operable 3D Cities from Sparse Aerial Observations
SummaryThis paper presents CityCode, a representation for persistent 3D cities that AI agents can measure, edit, and restore across tasks. The Air2Code system builds these cities from one low-altitude aerial RGB image and known camera parameters per region, recovering metrically scaled buildings and road geometry, with generative completion for occluded areas. Experiments on CitySample-Aerial and UrbanScene3D report more accurate structure recovery and better visual fidelity than the evaluated baselines.
In briefDesigning contracts to allocate resources among competing agents safely
SummaryThis paper studies strategic hypothesis testing, where a principal allocates resources among agents who can pay to join a statistical contract and receive allocations under a public mechanism. Using e-values, the authors define a continuous-closure of allocations that keeps null agents from profiting in expectation and gives post-hoc risk control for general losses. They also provide closed-form approximations, conditions under which non-null agents find participation profitable, and an extension to online agent arrival with time-uniform error control.
In briefOptimality of Random Features Under General Conditions: Misspecified Models and Robust Learning
SummaryThis paper analyzes kernel ridge regression with random features (KRR-RF) under general conditions such as misspecified models and heavy-tailed noise, where prior work was limited. The authors improve feature-map bounds to obtain minimax optimal learning rates under light-tailed noise, and extend these results to finite weak moment conditions using a Hilbert-valued Fuk-Nagaev inequality. Numerical experiments and a real data analysis validate the derived rates.
WhyImproves uniform feature-map bounds and proves minimax optimal rates under misspecification and heavy tails, backed by theory and experiments.
In briefAnalyzing how multilingual models represent color and whether it matches human perception
SummaryThis paper examines whether multilingual language models learn a shared internal geometry for colors and whether that geometry reflects human color perception or mainly language. The authors measure alignment with the CIE-LAB perceptual color space and cross-lingual symmetry across layers. They find perceptual alignment peaks in intermediate layers and weakens toward the output, and that alignment correlates with color classification accuracy of 96.7%.
Pablo Loyola, Edison Marrese-Taylor, Andr\'es Hoyos-Idrobo
In briefAssetForge: Asset-Augmented Task Synthesis for Training Business Automation Agents
SummaryThis paper presents AssetForge, a pipeline for building training tasks for business automation agents by pairing a reusable asset library with task-specific Rubrics. Candidate tasks are checked by executable tests and a model-based review, with bounded repair before acceptance. Fine-tuning Qwen3.6-35B-A3B on 6,000 AssetForge examples reaches 41.17% strict pass@1 on AutomationBench public-600, compared with 9.33% for the base model and 30.33% for traditional Rubrics.
In briefShoeFit: A New Dataset and Dual-Stream DiT Model for Virtual Shoe Try-On
SummaryThis paper addresses virtual footwear try-on, which must handle scarce matched data, viewpoint misalignment, and color distortion from background lighting. The authors introduce MVShoes, a dataset of 7,305 annotated image triplets, and ShoeFit, a dual-stream DiT architecture using 3D rotary position embedding and LayeredRefAttention. Experiments reported in the paper show improved rendering fidelity and robustness on real-world product shoes.
In briefOmni-RS: A unified framework for remote sensing understanding and generation
SummaryThis paper presents Omni-RS, a unified framework covering four remote sensing understanding tasks and five generation tasks. It uses a two-stage training strategy: a multimodal large language model is first adapted to remote sensing understanding, then its hidden states guide a diffusion transformer for image generation. The authors also build UniGen, a unified multi-task generation dataset, and introduce detailed visual conditioning with VAE-encoded visual conditions to support tasks needing reference images. Experiments report strong performance on understanding tasks and outstanding results across various generation tasks.
In briefOver-parameterization in Matrix Sensing: A Geometric Theory of Solution Robustness
SummaryThis paper studies how over-parameterization affects the robustness of solutions in matrix sensing, a problem that can also be viewed as training a two-layer quadratic neural network. The authors develop a rank-aware robustness theory for the Burer-Monteiro factorization, showing that robustness depends on both the degree of over-parameterization and the local geometry of the solution. They propose a trajectory-aware algorithm using spectral balancing and randomized sketching, which reduces sensitivity in experiments without degrading reconstruction accuracy.
In briefComplex Materials Engineering with Provenance Action Graphs: Potential, Structure and Demonstration
SummaryThis paper examines and tests reaction graphs for materials synthesis, where events are reaction steps (actions) and their outputs (observations). It notes that synthesizing useful materials, especially kinetically stabilized ones, relies on many unpublished failures as well as published recipes. The authors propose learning algorithms to extract rules from multi-dimensional synthesis pathways, aiming to speed up discovery through a shareable global knowledge base. Several scenarios in solid-state synthesis are discussed, since these reactions are kinetic and require close monitoring of reaction history.
Yevgeny Rakita, Paul Todd, James R. Neilson et al.
In briefOptimizing how doctor recommendations are presented to patients using simulated patient choices
SummaryThe paper addresses patients' difficulty comparing recommended physicians by generating short, case-specific summaries grounded in doctor profiles. It uses persona-conditioned LLMs to simulate patient choices offline and compares prompt-level optimization (GEPA), parameter-level optimization (GRPO), and their combination. GRPO substantially outperforms GEPA on satisfying all constraints, and combining the two yields the highest simulated conversion.
In briefSimInter: Interview-based profile enrichment for realistic user simulation in recommender systems
SummarySimInter is a pipeline that uses synthetic interviews between two LLM agents, grounded in user statistics and conversation history, to infer implicit traits such as decision-making style and domain expertise for conversational recommender simulators. An audit agent flags possible fabrication and stereotyping. On movie recommendation with ReDial users, SimInter significantly increased the diversity of expressed preference aspects compared with a statistics-only baseline. Human evaluation found 94.3% of enriched-attribute claims supported or partially supported.
Erfan Shahabi, Amirhossein Roshandel, Mohammad Sadegh Sadeghi et al.
In briefBayesian optimization of CVaR when the model's hyperparameters are unknown
SummaryThis paper studies Bayesian optimization of Conditional Value-at-Risk (CVaR) for an expensive black-box function that depends on a decision and an environmental variable. Existing methods assume the Gaussian process kernel and hyperparameters are known, but in practice they must be estimated from limited data, which can misspecify the model and distort tail-risk estimates. The authors propose HE-CV-UCB, which eliminates candidate models inconsistent with the data, queries where surviving candidates disagree, and returns a design robust across them. They prove its cumulative regret matches the oracle CV-UCB rate up to a factor depending only on the candidate set size, and report competitive results on synthetic and real-world tasks.
Valentin Margraf, Jonas Hanselle, Viktor Bengs et al.
In briefAffectWorld: A Benchmark for Emotion-Driven Future World Prediction
SummaryThis paper introduces AffectWorld, a benchmark testing how emotions shape actions and world states across 1,498 everyday videos annotated with emotions, action tendencies, and world-state constraints. A 233-sample counterfactual subset checks whether changing an agent's emotion produces matching changes in actions and consequences. Evaluating five state-of-the-art world models, the authors report world-state change scores of 0.100–0.280 and counterfactual emotion sensitivity scores of 0.027–0.070, indicating substantial gaps in emotion-driven prediction.
In briefLR-Guard: fast, explainable AI safety filtering via latent reasoning that can reconstruct readable explanations
SummaryThis paper introduces LR-Guard, a guardrail that classifies prompts for safety without generating explicit reasoning text for every decision. It distills a frozen teacher model's analysis of safe and unsafe interpretations into compact latent representations, trained jointly for safety prediction and for reconstructing the teacher's reasoning traces. The authors report that LR-Guard outperforms its teacher in most in-domain and out-of-domain settings and gives a 19.4x inference speedup at batch size one.
In briefA dataset of formal theorem statements from Annals of Mathematics papers
SummaryThis paper releases AnnalsChallenge, a dataset of 50 theorem statements hand-written in Lean 4 using Mathlib. Each is the main result of a paper published in the Annals of Mathematics since 2020. Every formalisation was manually reviewed and then audited by AI. Only the formal statements are released, without formalised proofs.
Thomas Browning, Katerina Hristova, David Ledvinka et al.
In briefLLM evolutionary search should be compared by token cost, not proposal count
SummaryLLM evolutionary search is usually evaluated by counting candidate program proposals, which hides the real cost of search. The authors replayed 600 logged runs across 5 tasks and 3 methods and found per-proposal token costs vary widely. When token budgets are matched, the evolutionary advantage over independent sampling mostly disappears.
In briefLegalRewardBench: Benchmark for Legal Preference Reward Models
SummaryThis paper introduces LegalRewardBench, the first human-verified benchmark for evaluating reward models in the legal domain. It organizes legal preference evaluation into five capability levels (knowledge, understanding, analysis, decision, and writing), with examples built through rubric-guided automatic judging plus human verification. Experiments on general and fine-tuned, generative and discriminative judges find that strong general-purpose generative models are currently the most robust, while existing fine-tuned reward models transfer less reliably to legal preference evaluation.
In briefBeyond Equation Search: SCOPE Framework and ClosureBench for Conjecture-Guided Scientific Discovery
SummaryMost symbolic regression methods only improve how well candidate equations fit data, without judging whether the underlying scientific conjecture should be kept, revised, or rejected. The paper introduces SCOPE, which runs multiple scientist agents that pursue different conjectures through equation modeling, questioning, and experiments, then exchange evidence and consolidate findings into a scientific report. It also presents ClosureBench, an interactive benchmark of 60 closure-discovery tasks across five scientific domains, evaluated with metrics for evidence faithfulness, mechanism insight, and experimental effectiveness. The authors report that SCOPE achieves higher equation accuracy than diverse baselines on ClosureBench and a real-world case.
In briefOutput Discipline Sets the Floor: A Model-Agnostic Agent Protocol for Auditable Decisions
SummaryThis paper examines why large language models perform poorly on auditable, hypothesis-driven tasks, using the TCMEval-SDT benchmark for Traditional Chinese Medicine syndrome differentiation (200 cases). The authors argue that output discipline and reasoning structure, rather than model capability alone, are the main constraint. They propose a model-agnostic agent framework that enforces a five-stage closed loop through a candidate-by-evidence falsification matrix. The protocol helps models whose output discipline is weak, slightly hurts already-disciplined models, and also transfers to misinformation-robust mathematical reasoning.
In briefDetecting logical anomalies robustly under changing capture conditions
SummaryThis paper studies detecting logical anomalies such as missing, extra, or wrongly combined parts in images. It shows that existing methods comparing test images with normal references lose substantial accuracy when brightness is reduced or images are zoomed out. The authors propose training a classifier on normal images and synthetic anomalies, generated with a language model and an image editor, without comparing to normal data at test time. The method achieves high AUROC on clean images and under perturbations, even without perturbation augmentation during training.</br>
In briefWhich optimizer leaks most? How optimizer implicit bias affects training-data privacy
SummaryThis paper studies whether the choice of optimizer affects how easily training data can be reconstructed from a released neural network checkpoint. The authors argue the key factor is the weight-specificity of each optimizer's implicit bias. Models trained with SGD and Muon retain enough weight-specific information to enable reconstruction, while Adam's ℓ∞ implicit bias leaves only a discrete pattern of maximal coordinates and signs. Experiments on MNIST and CIFAR-10 show Adam-trained models were consistently more resistant to the evaluated attacks.
In briefReliable Risk Estimation with Kernel Mean Matching under Dependent Sampling
SummaryThis paper studies Kernel Mean Matching (KMM), a method that estimates risk under covariate shift by reweighting labelled source samples to match the target distribution's kernel mean embedding. The authors give a finite-sample analysis for dependent sampling, replacing nominal sample sizes with effective sample sizes. The bound separates the effects of overlap, dependence, and smoothness. Applied to prediction, the approach yields confidence intervals with near-nominal coverage in synthetic experiments and reduced undercoverage on real monitoring data.
In briefInfrared small target detection: deep learning methods, challenges and outlook (Part 2)
SummaryThis paper is the second part of a survey on infrared small target detection, focusing on deep learning models such as convolutional neural networks, GANs, Transformers, Mixture of Experts, and graph and hypergraph neural networks. The authors categorize existing approaches by their modeling strategies and discuss the critical situations each claims to handle. They also present specific performance evaluation protocols and conclude with several promising future research directions.
Thierry Bouwmans, Anastasia Zakharova, Belmar Garcia et al.
In briefLongForcing: Extending Temporal Manifold Matching for Long-Horizon World Rollouts
SummaryAutoregressive video diffusion models drift over long rollouts beyond their training windows, with degrading quality, inconsistent motion, and collapsing object identity. The paper proposes LongForcing, which replaces the short-horizon teacher in the final distribution-matching stage with an extended-horizon bidirectional teacher that supervises long student self-rollouts, without changing the student architecture or inference cost. It reports a VBench Quality score of 85.52 and a 23.9-point gain on WorldRoamBench 60s action score in first-person ablations. Qualitative tests sample interactive rollouts up to 24 hours.
In briefFixing text in generated images by regenerating only the target regions
SummaryGenerated images can have small text and complex characters that are locally wrong, which hurts readability. The paper introduces PixText, which treats local text refinement as grouped image generation in pixel space, conditioned on the original image, generating only target regions and preserving unedited content with MeanFlow in four steps. The authors also introduce Text Refinement Bench to evaluate text correction and image preservation.
In briefPING: Physics-informed graph neural network for predicting cloud-driven solar power
SummaryThis paper proposes PING, a physics-informed graph neural network that forecasts advection-diffusion particle concentrations on irregular sensor grids. It adds a discretized advection-diffusion equation as a soft constraint in the loss function. The model performs on par with state-of-the-art methods on synthetic data and real sea surface temperature and cloud index datasets. It is also coupled with a photovoltaic power forecasting model to fuse ground sensor and satellite cloud data for intra-day solar output prediction.
Jelena Simeunovi\'c, Baptiste Schubnel, Pierre-Jean Alet et al.
In briefConventional and low-rank/sparse model-based methods for infrared small target detection: a survey (Part 1)
SummaryThis paper is a survey of infrared small target detection, which is used in ground, air, and sea surveillance. The authors group existing approaches into mathematical, machine learning, and signal processing models, and discuss the critical situations each claims to handle. They also present available resources, datasets, and libraries. This first part covers conventional and low-rank and sparse decomposition methods, while deep learning approaches are covered in Part 2.
Thierry Bouwmans, Anastasia Zakharova, Belmar Garcia et al.
In briefJevGPT: Fast and Slow Thinking for Autonomous Research Agents
SummaryThis paper presents JevGPT, a dual-system architecture for autonomous research agents that combines fast evidence processing with deliberative research planning. System Two sets evidence objectives and integrates findings, while System One locates supporting material, assesses relevance, and guides further inspection. Experiments on fixed-corpus question answering and document ranking report leading task performance, with faster completion and lower cost than comparable agents, and broader coverage of supporting documents.
In briefPAINT: A Multilingual Framework for Representing Pain in Medical Exams and Clinical Texts
SummaryThis paper introduces PAINT, a framework for studying how pain is expressed in medical licensing exam questions and real-world clinical texts across English, Spanish, Portuguese, and Japanese. It defines a structured schema covering pain relevance, presence, location, descriptors, and demographics, and uses an ensemble of large language models to label the data. Ten clinicians validated the labels, and the authors release the ensemble labels, extraction pipeline, validation scripts, and annotation sheets.
In briefFALLNEST: Condensing Imbalanced IMU Data into Small Balanced Sets for Wearable Fall Detection
SummaryThis paper addresses wearable fall detection, where falls are rare, labeled IMU data is scarce, and training deep models on imbalanced data is costly. It proposes FALLNEST, a framework that distills large, imbalanced IMU datasets into tiny class-balanced synthetic training sets. Evaluated on SisFall, KFall, and MobiFall with subject-wise splits, a TinyCNN trained on only about 1% of the training windows reaches 93.5-95.8% mean accuracy and 0.89-0.98 PR-AUC.
In briefNear-optimal online regret for bandits with lumpable contexts
SummaryThis paper studies stochastic contextual bandits where contexts are partitioned into r blocks and each arm has the same mean reward within a block. The authors propose an algorithm with expected regret Õ(√(r(S+K)T)), which matches the lower bound of Lee et al. (2023) up to logarithmic factors. This improves their previous upper bound by a factor of r.
In briefLookahead Self-Distillation (LEAD): Teacher Learns From Other Samples First
SummaryThis paper proposes LEAD, a method that improves on-policy self-distillation for training large language models. Unlike prior methods that use only privileged information from the same instance, LEAD lets the teacher first learn from other rollouts via virtual RL updates, then supplies token-level supervision across groups. It also adds harmonious modulation, which adjusts the GRPO advantage based on agreement between the two signals. Experiments on scientific reasoning and interactive-agent benchmarks report the best in-domain average across model scales and out-of-domain generalization comparable to GRPO.
In briefSurvey on shrinking LLM benchmarks to cut evaluation cost
SummaryThis survey reviews 119 papers on benchmark compression, which estimates full-benchmark scores and rankings for large language models from partial evaluations. It proposes a taxonomy covering compression within single benchmarks, across benchmark suites, and during adaptive evaluation. The authors compare methods by information needs, total cost, reliability checks, and limitations.
In briefPACT: A Prompting Method for Speech Translation That Keeps Tone and Voice
SummaryThis paper proposes PACT, a framework for speech-to-speech translation that aims to keep not only the content but also the speaker's voice and expressive features such as emotion, tone, rhythm, and pauses. It extracts pragmatic cues from the source speech, maps them into the target language's expression space, calibrates them, and turns them into continuous prompts for a frozen speech generator. On bidirectional Spanish-English TOPI evaluation, the authors report that PACT achieves the best overall performance among evaluated systems on content fidelity, voice preservation, and expression preservation.
In briefLiFUSE: A self-supervised pretraining method for liver disease analysis on non-contrast CT
SummaryThis paper proposes LiFUSE, a self-supervised pretraining framework for analyzing liver disease on non-contrast CT (NCCT), where low tissue contrast makes lesions hard to identify. A frequency branch uses random band-stop masking and spectrum reconstruction to capture subtle differences between lesions and surrounding tissue, while a spatial branch performs hybrid-granularity masked volumetric reconstruction. A cross-domain consistency scheme links the two branches. The authors report that LiFUSE consistently outperforms strong baselines on three clinical tasks and a public target-domain transfer setting.
In briefBeyond Answer Sharing: A Plug-in for One-to-One Peer Feedback in Multi-Agent Debate
SummaryMulti-agent debate lets language models see each other's answers and revise iteratively. This paper proposes TrueMAD, a test-time plug-in that adds one-to-one, free-form advice between answer broadcasting and self-revision. Source agents support, rebut, refine, or ask questions, while the target agent decides how to revise. It requires no parameter updates and can augment standard MAD and its variants.
In briefRebuilding Heading Hierarchies in Long, Complex Documents Using Level-Wise Multimodal Memory
SummaryVision-language models can extract page elements accurately, but predicting headings page by page often lacks document-level consistency. This paper proposes a VLM-based framework that incrementally rebuilds heading hierarchies using a level-wise multimodal memory, which keeps a compact set of earlier pages selected by anchors. The authors also introduce CLDoc-HHRBench, a benchmark of 250 documents and 11,212 pages with expert-verified annotations. Experiments on this benchmark and three existing ones show the method's effectiveness.
In briefConsider, then choose: order-independent answers from System One models
SummaryThe paper shows that System One models give different answers to choice questions depending on option order, with the effect growing as options increase, and that averaging over cyclic shifts only reduces it. It proposes Runoff, a training-free method that first scores each option alone to keep a few candidates, then compares them in listings where each candidate takes every position. The answer does not depend on option order by construction.
In briefWhen Does Opponent Modeling Help? Spatial Inference for Stable RL Under Partial Observability
SummaryThis paper studies how reinforcement learning agents in partially observable competitive games can act on beliefs about an unseen opponent's intent. The authors propose a transformer-based spatial opponent model that outputs a subgoal heatmap from the opponent's observed trajectory, fused into a dueling double DQN. The module stabilises training and transfers zero-shot to an unseen opponent, but its advantage disappears against a full-information adversary and on resource-dense maps.
In briefA Controlled Test of a Simple Curvature-Inspired Feature Scaling Method for Domain Generalization
SummaryThe paper evaluates a deliberately narrow method called HyperDG, which rescales Euclidean encoder features with learned positive scalars during training and uses an ordinary linear classifier at inference. It does not implement hyperbolic geometry, geodesic prediction, or Sharpness-Aware Minimization. Across PACS, VLCS, Office Home, and TerraIncognita, the full implementation scores 1.61 percentage points below native DomainBed ERM, but the difference is not significant after multiple-comparison adjustment. The authors conclude this is a bounded negative result for this scalar scheme and do not test genuine hyperbolic representation learning.
In briefTextualRL: Reinforcement Learning for Managing Agent Context
SummaryThis paper presents TextualRL, a gradient-free reinforcement learning framework that lets LLM agents improve by revising the textual context guiding their behavior, without changing model parameters. It treats a group of rollouts from the same task under the same context as the unit of critique, contrasting successful and failed attempts to propose reusable context updates. Averaged over six benchmarks, TextualRL improves test accuracy over the base model by 24.4 points for GPT-5.5 and 15.3 points for Qwen3.8-27B.
In briefOperator Surrogates: Finite-parameter maps that approximate complex operators between function spaces
SummaryThis chapter gives a mathematical introduction to operator surrogates, finite-parametric maps that approximate complex, typically nonlinear, operators between infinite-dimensional function spaces. A common approach is the encoder-approximator-decoder paradigm, which maps inputs to a finite-dimensional latent space, processes them, and maps back. It surveys representation methods, compares spectral methods based on polynomial chaos with neural network surrogates, and discusses architectures such as Fourier Neural Operators and transformers. It also covers approximation and generalization error theory, explaining how holomorphy and other low-dimensional structure help overcome the curse of dimensionality, with convergence bounds depending on trainable parameters.
Christoph Schwab, Josephine Westermann, Jakob Zech
