In briefPredicting scientific spectra while respecting physical laws
SummaryThis paper presents DeepSciReasoner, a method for predicting scientific spectra under hard physical constraints, using electronic density of states (eDOS) as the main example. It combines structure-aware spectrum decoding with constraint-preserving physical reasoning so predictions capture spectral structure while following the underlying physics. The authors evaluate it on eDOS, phonon density of states, XANES, and Raman spectra, reporting substantially better accuracy with physical consistency maintained.
In briefBriLLM-MetaPred: A language model unifying representation and prediction through self-measurement
SummaryThis paper presents BriLLM-MetaPred, an engineering implementation of the MetaPred machine-learning framework, derived from a mathematical theory of AGI, applied to language modeling. A single continuously evolving dynamical state serves both as the context representation and as the source for predictions, and continuous self-measurement of this state reduces finite-depth inconsistency. The authors report predictive performance comparable to a GPT baseline of the same size, offered as initial evidence of the framework's engineering feasibility.
In briefResCom: Activation Residual Compensation for Compressing Large Language Models
SummaryThis paper proposes ResCom, a method that reduces accuracy loss when large language models are quantized to low-bit weights and activations. It uses the residual left by rounding activations to apply a low-rank correction that offsets layer-output errors. The correction is calculated in closed form through output-aware calibration, without gradient training or updating calibrated weights. On Llama-3-8B with 4-bit weights and activations, a rank-32 attention correction lowers WikiText-2 perplexity from 7.708 to 7.147 and raises five-shot GSM8K accuracy from 23.96% to 36.16%.
In briefThe Information Geometry of Representation Formation: Studying How Neural Representations Form During Training
SummaryThe paper studies how neural network representations form during training, rather than only analyzing trained endpoints. It uses Bures geometry and quantum Fisher information to measure movement through representation space and local sensitivity to input factors. The authors report that less direct formation paths occur when the computation needs higher-order interactions, and that path tortuosity predicts later representational drift.
In briefIs the Patient Always Right? Measuring Sycophancy in AI Medical Diagnosis
SummaryThis paper tests nine AI models to see whether they change their diagnosis when a patient claims to have a specific but incorrect condition. The models often go along with the patient's belief, sometimes even for made-up conditions. Medical fine-tuning and larger model scale make models less likely to change their diagnoses. The authors argue this sycophantic bias matters for safety as more people use AI for health support.
Viswonathan Manoranjan, Andrea Mock, Myra Cheng et al.
In briefContinual Learning Helps Muscle-Signal Gesture Models Adapt to New Users
SummaryThis paper studies how continual learning can let recurrent neural networks that decode hand gestures from surface EMG adapt to new wearers and new gesture classes after deployment. On NinaPro DB1, experience replay outperformed elastic weight consolidation and naive fine-tuning, reaching 83.3% final accuracy under class-incremental drift. The authors also report that these models should be trained on shuffled, independently windowed segments, and that hand-crafted time-domain features help, but only replay keeps that gain under class-incremental drift.
Aran Ramakrishnan, Shubhi Mishra, Tyler Ham et al.
In briefFed-Universe: a four-layer framework for negotiated alignment in federated learning
SummaryThe paper argues that conventional federated learning optimizes for global averages, which can suppress minority sites with different data. The authors propose Fed-Universe, a Semantic-Geometric-Topological-Human architecture that turns passive averaging into active negotiation among nodes. In a synthetic binational clinical network simulation, a minority node with under 3% of the data reduced its loss from 0.857 to 0.340 while dominant hubs were maintained.
In briefPartial Modal Tensor Regression: Reducing Tensor Data Dimensions Using the Conditional Mode
SummaryThis paper proposes partial modal tensor regression (PMTR), a method for modeling the relationship between tensor predictors and a response. It builds a partial modal tensor covariance that uses conditional mode information to guide dimension reduction along each tensor mode, then fits a modal regression model in the reduced space. Under suitable regularity conditions, the authors establish consistency and asymptotic properties of the estimators and use a data-driven procedure to select the effective dimension per mode.
In briefSteering the Teacher: Improved Self-Distillation for Thinking Models
SummaryThis paper examines on-policy self-distillation, where a student learns from its own reasoning under a reference-conditioned teacher, and finds the reference can cause the student to cite unavailable material and the teacher to discourage rechecking uncertain steps. The authors propose ST-OPSD, which uses activation steering with two directions taken from model reasoning trajectories to adjust the teacher's hidden states during training. Experiments on AIME and HMMT report higher accuracy than vanilla OPSD and other interventions, with reconsideration and reference-hallucination rates restored to base-model levels.
In briefResolving the Mobility-Equity-Emissions Trilemma: A Causal ML Framework for Dynamic Tolling
SummaryThis paper proposes a causal machine learning framework that integrates dynamic road tolling with mobility, equity, and climate goals. Using Double Machine Learning, the authors find commercial zones achieve nearly twice the delay reduction of residential zones. Their adaptive quota expands residential coverage 34-fold while keeping 93% of total delay savings, and under moderate traffic it estimates 743.5 metric tons of CO2 reduced, within bounds of 557.6 to 929.3 tons. The results are validated on synthetic data mirroring real-world networks.
In briefHybrid Mamba-Transformer network for reconstructing 3D shapes from a single fringe image
SummaryThis paper proposes a hybrid Mamba-Transformer network that recovers surface height from a single high-frequency fringe image in fringe projection profilometry. It uses a Swin-based U-shaped architecture with visual state space modeling and a new axial and dilated vision state space (ADVSS) block to capture both long-range consistency and local fringe detail. On a public FPP dataset, the method outperforms hNet, SwinConvUNet, UHRNet, and HiDNet on MAE, RMSE, absolute relative error, and SSIM, reaching an RMSE of 0.548 mm and SSIM of 0.9985 from one image.
In briefConstrained Gradient Equilibrium: theory on online learning under locally constrained adversaries
SummaryThis paper studies gradient equilibrium (GEQ), an online learning framework where time-averaged gradients along the learner's trajectory vanish. It shows that constrained GEQ (CGEQ) can be impossible even under simple, slowly varying constraints, so it imposes local temporal structure on the adversary. The authors give necessary and sufficient conditions in the style of Blackwell approachability and an algorithmic method that turns window-wise no-regret learners into CGEQ algorithms. Experiments on synthetic and real data show strong performance in both achieving equilibrium and respecting constraints.
In briefEmoNet: A Lightweight Attention-Based CNN for Recognizing Emotion in Speech
SummaryThis paper introduces EmoNet, a lightweight deep learning model that classifies speech into seven emotion categories. It combines a depthwise separable CNN with attention, a bidirectional LSTM, and multi-head attention pooling, using 134 acoustic features and data augmentation. On a combined TESS, RAVDESS, and SAVEE dataset, it reports 96.75% test accuracy with 4.11 million parameters. The authors state it outperforms their reproduced baseline and is competitive with much larger state-of-the-art models.
Muhammad Moiz Khalid, Abdur Rehman Tariq, Muhammad Okasha Yamin
In briefWhich Data Quality Metrics Predict Instruction-Tuning Value Early On
SummaryThis paper examines which simple per-example metrics best predict how useful training data is for instruction tuning under a fixed finetuning budget. The authors rank examples by each metric, split them into low, middle, and high bins, finetune on each bin, and compare validation perplexity and LLM-as-judge scores. They report that response ratio, a measure of answer-dense supervision, predicts instruction-tuning utility more strongly and portably than other commonly used proxies.
In briefHierarchical Cooperative GFlowNets: several samplers divide up the search so they cover more high-reward regions
SummaryGFlowNets learn to sample objects with probability proportional to a reward, but weak exploration can make them concentrate on a few high-reward regions. The paper proposes Hierarchical Cooperative GFlowNets (HCG), which uses a shared classifier to dynamically partition the search space so higher-ranked models explore broadly while lower-ranked ones are pushed toward uncovered regions. The authors also derive a normalizer-weighted ensemble aggregation rule and prove it recovers the target distribution under stated assumptions. Experiments on multimodal distributions, sequence generation, combinatorial optimization, and biological sequence design report improved mode coverage and target approximation.
Pedro Dall'Antonia, Juan Belieni, Ana Carolina Erthal et al.
In briefRe-evaluating Neural Operator Baselines on The Well Benchmark
SummaryThis paper revisits the neural operator baselines released with The Well, a 15-terabyte benchmark of 16 physical simulation domains. The original baselines were trained with little or no hyperparameter tuning and were later treated as fully tuned benchmarks. Under the same 12-hour per-run limit, the authors ran a short hyperparameter search on a Tucker-factorized Fourier Neural Operator using only a single input timestep. Tuning alone lowered VRMSE below the published FNO and TFNO baselines on 15 of 17 tasks, sometimes by orders of magnitude, so the authors argue these baselines should not be treated as architectural benchmarks.
In briefInteractiveWorld: A Testing Environment for Whether AI Assistants Can Really Guide People
SummaryThis paper presents InteractiveWorld, a verifiable environment where an AI assistant guides a human who acts in the scene, forming a closed loop of instruction, action, and updated view. Building on it, the authors introduce InteractiveBench, a dynamic benchmark of 120 multi-turn tasks across six categories, scored by state-based rewards. Experiments with 12 frontier multimodal LLMs show they remain unreliable, with the strongest model reaching at most 20% on long-horizon Guidance and 25% on Planning tasks.
In briefBeyond Accuracy: Comparing ResNet50 and ViT Efficiency and Statistics for Explainable Gastrointestinal Disease Classification
SummaryThis paper benchmarks ResNet50 and ViT-B/16 for classifying eight gastrointestinal findings in endoscopic images from the public Kvasir dataset. In cross-validation both models reach about 92% accuracy, and a statistical equivalence test found them equivalent within a ±3 percentage-point margin. The main difference is cost, with ResNet50 having fewer parameters and operations and faster inference. The authors conclude an efficient CNN is a defensible default for resource-constrained screening, but no model is ready for deployment without local validation.
In briefQECGBERT: A Scalable ECG Foundation Model Using VQ-VAE Waveform Tokens and Self-Supervised Learning
SummaryThis paper proposes QECGBERT, a self-supervised ECG language model that tokenizes the P, QRS, T, and isoelectric waveforms using a VQ-VAE. It is pre-trained with BERT-style masked language modeling on over 800,000 unlabeled recordings from MIMIC-IV. The model is then fine-tuned and evaluated on PTB-XL, CPSC2018, and Chapman–Shaoxing–Ningbo datasets, with results reported as comparable or superior to published benchmarks.
In briefBeyond Belief: Testing Whether LLMs Actually Resist Misinformation
SummaryThe authors test five tool-using language models on fabricated notices and archived attacks to see whether stated credibility judgments predict safe behavior. Models often executed harmful tool calls even after rating the information untrustworthy, and asking for that rating itself changed their behavior. The paper argues that misinformation defenses should be evaluated on both harmful actions and legitimate task completion.
In briefHow Small Coding Agents Use Code Graphs and Embeddings: a study testing whether small open models really gain from structure-aware search tools when localizing code.
SummaryThe paper tests small open-weight coding agents on SWE-bench Verified, stratifying tasks by code-graph hop depth. Structured tools roughly double grep-only accuracy for some small models, but the gain shrinks as model size grows. Agents mostly rely on one-hop, largely lexical information and rarely chain multi-hop graph expansions.
WhyDepth-stratified ablations over ten models (79k episodes) and probes show small agents use code-graph tools mainly one-hop and lexically.
In briefA nonconvex method that jointly learns sample weights and model parameters to reduce the influence of bad data
SummaryThis paper proposes SWAPE, a framework that jointly optimizes model parameters and an individual weight for each training sample. It supports convex, weakly convex, and difference-of-convex loss functions, plus weakly convex sparsity functions that control the complexity of the weights. The bivariate problem is solved by alternating minimization, with subproblems approximated by majorize-minimize techniques, and the authors show the algorithm converges. Experiments on various learning tasks and data anomaly settings show it reduces the impact of undesirable data points compared with benchmark methods.
In briefContinual Learning Exam: A benchmark testing whether LLM agents learn from accumulated experience
SummaryThis paper introduces the Continual Learning Exam (CLE), a benchmark of 120 expert-written tasks, each with five levels, totaling 600 levels across 89 professional roles in 12 occupational domains. Agents must pass each level to unlock the next, with later levels building on earlier experience, and an evaluation agent scores submissions against human rubrics. Providing diagnostic feedback and one revision raised median completion by 5.3 percentage points, but the best system cleared only 47.7% of levels and completed all five levels in just 23.3% of tasks. The authors note the dataset can also be used to post-train LLMs for continual learning.
In briefWhat Makes a Question Curious? Measuring Model Curiosity in Human-LLM Interactions
SummaryThis paper studies curiosity-oriented questions, which are questions a language model asks to surface knowledge gaps, assumptions, and alternatives, and proposes a theory-grounded taxonomy that separates them from generic questioning. Across four conversation datasets, such questions are rare but are associated with more exploratory and epistemically rich user responses, with patterns varying by dataset. A human study finds that model curiosity can encourage exploration, especially in advice, creative, and research contexts.
Angana Borah, Rachana Ramchandar, Inderjeet Nair et al.
In briefWhy Structured Output Fails: A Study of Prompt-Enforced Pydantic Schemas Across Nine LLMs
SummaryThis paper examines why language models fail to produce structured outputs matching a Pydantic schema specified in the prompt. The authors introduce a fine-grained failure taxonomy and a deterministic classifier, applying them to 3,997 generation attempts across nine OpenAI and Anthropic configurations. They find that similar validity rates can hide different failure patterns, and that Markdown-fenced JSON is the most common failure, often concealing additional violations such as field-length errors.
In briefText-to-SQL without schema docs, learning database structure from query logs
SummaryThis paper proposes a text-to-SQL framework that needs no schema documentation and instead learns table relationships and column meanings from historical query logs. It builds a knowledge graph from parsed query patterns and uses entity-driven GraphRAG retrieval to link user questions to domain terminology. On the Beaver benchmark it reaches 95.7% execution validity, versus 60.8% for embedding-based retrieval. In an internal deployment, 70% of expert data engineers preferred this system and 10% preferred the alternative.
In briefTurboWAM: Learning What Matters for Faster Robot World-Action Models
SummaryThis paper presents TurboWAM, a fast world action model for robot policy learning that aims to reduce high inference latency. It adaptively fuses representations from different world model layers based on their relevance to the evolving action denoising state, and modulates action supervision gradients across layers during backpropagation. At inference, it reduces the number of executed world model layers and caches relevant representations across denoising steps. Experiments on multiple robot manipulation benchmarks show low latency with task success rates competitive with state-of-the-art methods.
In briefStable, calibrated latent force ODE filters for estimating hidden forces in physical systems
SummaryThis paper studies approximate Gaussian filters that estimate unobserved forces in mechanistic systems while inferring the system state in one forward-backward pass. The authors show that the joint posterior over state and force is a reweighted version of the conventional latent force posterior, and that stability depends on the prior variance of the system state. Based on this, they propose an improved algorithm for stable and calibrated inference, demonstrated on several simulated settings.
In briefHerHealthGPT: An evidence-based AI chatbot for women's reproductive health questions
SummaryThis paper presents HerHealthGPT, a conversational AI system for women's reproductive health that grounds its answers in retrieved evidence. It asks targeted follow-up questions when a query is incomplete and checks for premature conclusions before responding. Evaluation on 540 English messages showed that evidence retrieval raised concern-category accuracy from 54.1% to 59.3%, and 83.3% of 30 participants preferred the retrieval-enabled version. The authors note that further clinical validation is still needed.
Hassan Saeed Hassan Albattra, Mazen Mohammed Bahgat, Rahatara Ferdousi et al.
In briefEnerGNN: Learning Energy Functions That Enable Exact Constrained Combinatorial Optimization
SummaryThe paper introduces EnerGNN, which uses a graph neural network to learn energy functions over binary configurations and minimizes them exactly under hard constraints via mixed integer programming. On Quadratic Knapsack, Combinatorial Auctions, and a cell therapy supply chain problem, it reports near-optimal solutions with 100% feasibility. On the supply chain problem it reaches a 0.05% mean primal gap and a 17.4× speedup over monolithic MIP. The learned energy generalizes to larger instances and supports zero-shot constraint injection without retraining.
In briefWebViSE: a method that extracts visual design skills from the open web to guide LLM-generated websites
SummaryThe paper introduces WebViSE, a framework that automatically turns mature website implementations into reusable Visual Skills and organizes them in a multi-parent DAG library. The authors report that this structure outperforms non-hierarchical LLM retrieval baselines at 39.09% of the estimated retrieval cost. WebViSE can be injected as plug-and-play guidance or used to build a data-synthesis pipeline, and experiments and human evaluation show improved visual quality and interaction fidelity.
In briefTask Interactions in Shared Recurrent Networks Across 100 Cognitive Tasks
SummaryThe paper trains hierarchical Almost-Linear RNNs jointly on 100 cognitive tasks to study when tasks transfer or interfere with each other. The joint hierarchical model outperforms both individually trained models and a single unstructured joint model. Many tasks are solvable individually with linear dynamics, yet a shared linear model cannot solve them together.
S\'ilvia Garc\'ia-Hern\'andez, Leema Hamid, Manuel Brenner
In briefNeural Eigenfluid Bases: a learned, physically constrained basis for simplifying fluid simulation
SummaryThe paper introduces Neural Eigenfluid Bases (NEB), a neural network representation of reduced-order fluid modes that can be evaluated anywhere in space, with incompressibility and impermeable boundaries enforced directly. In a periodic vortex merger test, 1024 modes keep velocity errors below 1%. For random flows around obstacles, potential-based learning captures more initial energy and gives lower final errors than numerical Laplacian bases of the same size. Precomputed Galerkin operators allow millisecond-scale updates when the basis is reused.
In briefLearning population dynamics via variational Doob-Wasserstein flow
SummaryThis paper proposes VDWF, a simulation-free framework for learning unbalanced population dynamics on the Riemannian Wasserstein-Fisher-Rao manifold, based on the unbalanced mean-field Schrödinger bridge. It learns a data-dependent geometry, uses Doob's h-transform to incorporate boundary information for trajectory inference, and employs a growth model to construct intermediate mass samples. The authors analyze convergence using a uniform Log-Sobolev inequality and report competitive results against state-of-the-art methods on real-world datasets for balanced and unbalanced dynamics.
In briefReal-time RGB-D human trajectory prediction for mobile robots using Bi-Mamba
SummaryThe paper presents TrajMamba, a lightweight bidirectional state space model that predicts short-term human trajectories from RGB-D cameras for mobile robots. It uses one shared processing path for training and deployment to reduce the gap between idealized training data and noisy real-sensor inputs. On the ETH/UCY benchmark it reports an average ADE of 0.29 m and FDE of 0.51 m with 262K parameters, and it runs on a TIAGo robot with an onboard NVIDIA Jetson Orin.
In briefTexture-aware subregion style transfer for data augmentation
SummaryThis paper studies style transfer for image data augmentation. Existing methods pick subregions for style replacement randomly or under fixed geometric constraints, which may miss informative textures. The authors propose prioritizing regions with strong texture responses for style replacement, and a hybrid strategy that combines a texture-driven region with an auxiliary random region. Experiments report higher classification accuracy and better robustness than existing style-based augmentation methods.
In briefRank-one weight correction for trainable weight averaging in out-of-distribution detection
SummaryThis paper addresses out-of-distribution (OOD) detection, which identifies inputs outside the training distribution. It extends Trainable Weight Averaging with rank-one weight components extracted by Singular Value Decomposition, learning their contributions through the in-distribution classification objective. On ResNet50 it improves over fine-tuning and standard TWA on the evaluated datasets with MSP and RankWeight, while Mahalanobis-based evaluation gives mixed results.
In briefOS-Wizard: Scaling Computer-Using Agents to Specialized Software via Knowledge-Guided Exploration
SummaryThe paper presents OS-Wizard, a framework that automatically collects interaction trajectories for computer-using agents, combining tutorial-guided exploration with a hybrid GUI and code action space. It revisits pivotal states during tree exploration to cover alternative workflows. The authors collected 10.9K trajectories and fine-tuned open-source models, which improved performance on a specialized desktop workflow benchmark and also helped on general desktop tasks.
In briefCounterfactual test-time augmentation to remove shortcuts in language models
SummaryLanguage models can rely on training-data correlations (shortcuts) that fail when correlations shift at test time. This paper studies shortcut removal at test time and proposes counterfactual test-time augmentation (CTTA), which generates edits changing the shortcut value with an auxiliary model and averages the LLM's predictions over them. The authors prove the optimal shortcut-invariant predictor belongs to this family and characterize its weights, and they give conditions under which the weights can be estimated from token probabilities, along with tests for violations. The resulting Select-CTTA method outperforms test-time baselines and approaches models fine-tuned on shortcut-balanced data.
Katie Matton, Rosalind W. Picard, John Guttag et al.
In briefWhen Personas Become Harmful: Aligning LLMs with Global Citizenship Values
SummaryThe paper argues that persona-conditioned requests grounded in values are often not overtly malicious, so existing safety benchmarks may miss harms arising from persona use. The authors build a dataset of 11k controversial topics paired with value-based personas and use an agentic pipeline to collect harmful responses from well-aligned commercial LLMs, with a success rate above 47%. They use UNESCO's global citizenship principles as common ground to separate valid value expression from harmful behavior, and introduce a controllable alignment model that balances persona adherence and safety at inference time. The model is reported to be safer and more persona-adherent than baselines on most samples.
Siqi Shen, Lajanugen Logeswaran, Moontae Lee et al.
In briefWMEval: Benchmarking Interactive World Models Against Physics Simulations
SummaryInteractive world models are promising for robotics, game engines, and digital twins, but existing benchmarks mostly measure visual similarity rather than physical-causal coherence. WMEval uses Unity-based physics simulations as ground truth, running matched action sequences in both the simulator and the world model across 9 scenes. It assesses controllability, consistency, kinematics, fluid mechanics, physical causality, and lighting. Experiments on state-of-the-art models show substantial room for improvement in aspects that standard perceptual metrics do not capture.
Lanxiang Hu, Yixin Huang, Abhilash Shankarampeta et al.
In briefCurriculum Deep Reinforcement Learning for Autonomous Long-Range Rocket Landing
SummaryThis paper presents a curriculum-driven deep reinforcement learning framework that trains a reusable rocket to navigate from a far horizontal distance of 1,000 to 1,500 units and perform a soft precision landing. It uses a three-stage curriculum, critic warm-up, a rehearsal mechanism against catastrophic forgetting, and a full rotation-matrix observation, along with a redesigned reward function. Over 100,000 evaluation episodes, the proposed policy achieved a 99.98% success rate versus 68.31% for the baseline, with a mean final distance from the pad centre of 7.07 units versus 13.10 units.
In briefMin-Marginal Power Index for Choosing Which Features to Drop in Multi-Objective Models
SummaryThis paper proposes MinPI, a game-theoretic feature selection index that labels each feature as harmful, neutral, or decisive based on whether adding it drops model utility below a quota in some coalition. Harmful labels come with a concrete evaluated coalition as evidence. On six benchmark datasets over five seeds, MinPI removed 12.5% of analyzed features with almost no loss of held-out utility.
In briefIntervention-Certified Amplification: A Validity Gate for σ-Weighted Post-Training
SummaryThe paper shows that in hybrid post-training mixing SFT and RL, training-side signals can reactivate and amplify pretrained behavior that does not help the target task. It demonstrates this with DIAA, a per-sample σ-weighted hybrid that emits unexecuted Tool-Integrated Reasoning on MATH-500. The authors propose Intervention-Certified Amplification (ICA), which gates σ using paired-rollout target evidence, reducing TIR and improving accuracy, and transferring to held-out GSM8K.
In briefDIGEST: Discovery-Guided LLM Tuning for Symbolic Regression
SummarySymbolic regression aims to find mathematical expressions directly from data. The paper proposes DIGEST, which repurposes a pretrained decoder block from a lightweight LLM as a symbolic policy to seed Genetic Programming populations. High-quality expressions found by GP are fed back to update this LLM-derived policy online. On two ground-truth benchmarks, DIGEST outperforms the strongest competitor by up to 2.59% in recovery.
In briefAnchoring Explanation Regularization to Semantic Concepts in Continual Learning
SummaryThis paper addresses explanation drift in continual learning, where models keep their accuracy while shifting attention from diagnostic features to spurious correlations. The authors argue that pixel-level saliency regularization uses the model's own frozen saliency maps as targets, creating a self-referential loop that reinforces spurious attention. They propose C³L, which anchors regularization to class-consistent semantic concepts discovered offline. The authors report state-of-the-art results across five benchmarks and over 20 baselines.
In briefFlashScript: Turning Repeated Agent Workflows into Reusable Neuro-Symbolic Programs
SummaryThis paper presents FlashScript, a method that compiles a successful agent run into a reusable neuro-symbolic program (NSP). Actions and data transformations become code, while only decisions needing model judgment become LLM calls. On 120 WorkBench analytics tasks using Claude Haiku 4.5, the NSPs gave correct, identical tool calls every run, using about 70 times fewer input tokens and a third of the runtime of a skill-guided agent. On a composite version of the tasks, the skill-guided agent's accuracy dropped to 23%, while the NSPs stayed correct and consistent.
In briefStable Dimensionality Reduction with Isometric Centroid-Encoder
SummaryThis paper proposes the Isometric Centroid-Encoder (ICE), a supervised dimensionality reduction method that learns class centroids in the latent space while preserving a prescribed pairwise centroid geometry. It also introduces the Centroid Neighborhood Stability (CNS) metric, which measures how consistently class-level neighborhood structure holds across independent runs. Experiments on image and biological datasets show that ICE yields more stable embeddings that better preserve class-centroid geometry.
In briefRelateGS: Reliable evidence for 3D scene understanding using Gaussian Splatting
SummaryThis paper presents RelateGS, a structured language-aware Gaussian representation for 3D visual grounding. It suppresses unreliable observations during semantic feature uplifting and consolidates multi-view object observations into a persistent 3D graph. The method identifies under-observed object relations and guides re-querying of existing frames. It reports improvements of 7.9 and 1.7 percentage points at Acc@0.25 on Sr3D+ and Nr3D over the previous best.
In briefLate-breaking results: speech-driven robot manipulation models that skip text transcription
SummaryVision-language-action models let robots follow typed instructions, but spoken commands usually require a separate speech recognizer first, which adds delay and error. This paper presents Speech-π0.5, built on π0.5, which maps raw speech directly to motor commands by projecting voice frames into audio tokens. It uses split LoRA adapters for audio and task, trained in two stages. The authors report about 3x lower response latency than an ASR pipeline, competitive success on LIBERO, and more robust handling of unseen accents than a single adapter.
In briefReGE: Stronger Agents Through Experience Synthesis
SummaryThis paper presents ReGE, a framework that turns an agent's recorded experience from earlier interactive attempts into guidance for new attempts. Its main implementation, ReGE-Syn, uses one coach call to synthesize execution records and available artifacts into self-contained guidance. Across four benchmarks, ReGE improved accuracy by an average of 9.9 points for GPT and 8.8 points for Claude over unguided execution.
In briefNumina Fuse: An interactive tool for autoformalizing mathematics in Lean 4
SummaryNumina Fuse is an agent-based application for formalizing mathematics in Lean 4, where a user supervises an AI orchestrator through views of the blueprint, dependency graph, and Lean code. The paper describes its development during 2026, as it shifted from a human-paced pipeline of agents to an orchestrator that dispatches specialists, provers, and parallel children on its own. It also reports use in formalizing the three-dimensional Kakeya conjecture and in a study of prover models.
In briefTrackTok: Video tokenization that keeps the same objects consistent across frames
SummaryTrackTok is an efficient video tokenization framework that produces object-centric tokens, which keep encoding the same semantic object as it moves, deforms, or is partly occluded. Unlike patch-based or location-tied tokenizers, it learns temporally persistent representations that bind features across frames into stable units. On UCF-101 it reaches comparable class-conditional generative FVD using 8 times fewer latent tokens, and on Kinetics-600 it outperforms 1D tokenization baselines on frame prediction at low token budgets.
In briefSoftware Engineering LLMs: Using Software Engineering Knowledge to Train Coding Models
SummaryThis paper studies how software engineering knowledge can guide the training of a 92B-parameter coding LLM serving over 100,000 developers. The authors built RefactorForge, a pipeline that combines refactoring knowledge, SE tools, and verification to create executable training data from 306 Python repositories. Fine-tuning on the verified trajectories improved scores on RefactorBench and SWE-Bench ProMax, with gains extending to some SWE issue-resolution benchmarks.
Amir M. Ebrahimi, Mohammed Mehedi Hasan, Seyed Shayan Daneshvar et al.
In briefWhen early stopping helps under label noise: a phase diagram for kernel gradient flow
SummaryThis paper studies whether early stopping is needed when training labels contain substantial irreducible noise. Using kernel gradient flow on a synthetic noisy classification task, the authors derive sharp high-dimensional learning curves and a phase diagram governed by the balance between a linear kernel component and higher-order components. In one regime early stopping is necessary, in another generalization improves until interpolation but the endpoint is suboptimal, and at the boundary interpolation gives the lowest generalization error. Thus interpolation can outperform early stopping even under extensive label noise.
In briefBuild, Don't Play: Self-Improving LLM Agents Create a Hearthstone Battlegrounds AI Rivaling Top Players
SummaryThis paper asks whether LLM agents can take a complex real-world problem from raw data to a system competitive with humans, using Hearthstone Battlegrounds as the testbed. The authors present AutoTavern, in which specialized research agents share verified memory and improve from live game feedback, producing the TavernGym simulator and TavernAgent in 20 days. According to the paper, TavernAgent's last 20 ranked games averaged a placement of 1.60, better than the server's ten highest-rated players. The authors conclude that LLM agents build such systems far better than they play them.[SEP]
WhyProposes a self-improving LLM research system that builds a faithful simulator and a top-player-beating agent, backed by live ladder results and a large benchmark.
In briefApproximate universal prediction when predictions shape the data an agent sees
SummaryThe paper studies whether approximation guarantees for universal predictors, such as Bayesian mixtures and Solomonoff induction, survive when an agent plans using those predictions and thereby changes where they are evaluated. It derives a transfer coefficient for finite-horizon causal mixtures and shows that excess prediction error can scale as kappa times eta times 2 to the K(mu), a bound the authors prove is sharp. The work also gives value-error bounds for bounded returns and margins for preserving deferral decisions.
In briefSteering Vision-Language Models to the Modality You Ask For
SummaryThis paper asks whether users can reliably control which modality, image or caption, a vision-language model follows when the two conflict. It proposes Flip and Stay counterfactual tests to measure this control and learns it by training new word embeddings for visual and textual grounding. Experiments span three VLMs, three existing benchmarks, and a new multi-attribute Ground-CF benchmark.
Maxime Dassen, Alexander Martin, Andrew Yates et al.
In briefSliding Networks: Residual Designs That Keep Features Learning and Diverse
SummaryThis paper studies how to scale the width and depth of deep networks while preserving stable training and feature diversity in the residual stream. It proposes a joint scaling recipe governed by the limiting ratio of residual width, hidden width, and depth, with correlated initialization (Half Copy) partially relaxing shape constraints. Theory and residual-MLP experiments show convergence at fixed positive ratio to an SDE limit with locally maximal feature learning, but language-model experiments show performance differences when hidden computation is regrouped between width and depth at fixed parameter count.
In briefHessian-Similar Optimization with Limited Client Participation: Lower Bounds and Virtual Masters
SummaryThis paper studies how many communication rounds Hessian-similar optimization needs when each round can contact at most s of m clients. It proves a lower bound on strongly convex quadratics and proposes a virtual-master SVRS method with Catalyst whose round count does not depend on the anchor's quality. The lower bounds apply only to zero-respecting algorithms.
In briefEEGBiFormer: Real-time brain-signal classification of attention, with audio protocols chosen by EEG
SummaryThis paper introduces EEGBiFormer, a hybrid CNN-Transformer model that classifies focused versus non-focused attention states in real time from 16-channel EEG signals. The authors built a custom EEG acquisition board, and spectral and connectivity analysis confirmed differing neural patterns between the two states before training. The model reached 91.3% window-level validation accuracy and 88.2% recording-level accuracy across 17 sessions, with end-to-end latency under 100 ms. In a within-subject study with 37 participants, EEG-ranked auditory protocols improved composite attention scores by 20% over silent controls.
In briefReactive Optimizers: A Language-Model Agent Rewrites Optimizer Code During Training
SummaryThis paper proposes OPAL, a reactive optimizer in which a pretrained language-model agent inspects training dynamics and rewrites executable optimizer code within a single training run. The agent can change update rules and schedules for individual parameter groups without optimizer-specific meta-training. Evaluated on language modeling and image generation with recipes not tuned for the target setting, it matches or improves on selected fixed recipes.
In briefAdapting DNNs to Tolerate Voltage-Scaling Errors via Accumulation-Aware Quantization
SummaryThis paper proposes a hardware-software co-design method for running deep neural networks at lower voltages, based on the observation that timing-robust bit ranges in a systolic array's accumulator vary across voltage-frequency operating points. It combines accumulation-aware training, mixed-precision quantization with pruning, selective MSB masking, fine-grained clock gating, and layer-wise dynamic frequency scaling. Evaluated on a 2nm node, the authors report energy and energy-delay product gains of 39% and 23% at a low-voltage operating point versus the nominal voltage, with accuracy degradation within 1%.
Alexis Maras, Panagiotis Chaidos, Georgios Alexandris et al.
In briefLearning What to Share: Expert Specialization for Time-Series Forecasting
SummaryThis paper studies how multivariate forecasting models should share information across channels and specialize, examining whether mixture-of-experts gains come from extra capacity or from expert allocation. The authors propose Persistent Closed-Form Routing (PCR), a plug-and-play method that trains low-rank residual experts around a frozen pretrained forecaster, and introduce the synthetic GalaxyTS dataset to analyze when specialization helps. The paper reports that PCR reduces mean squared error relative to three frozen backbones when averaged over four forecasting horizons.
In briefEnvLunch: Using the execution environment itself as free training tasks for interactive agents
SummaryThis paper addresses the need for many human-annotated tasks to teach language-model agents to interact with executable environments such as databases for data analysis. The proposed EnvLunch is a label-free warm-up stage that samples valid action trajectories, executes them to obtain final observations, and trains the agent to recover a functionally equivalent trajectory from only that final observation. The authors report that it outperforms the baseline by over 9 points on average across downstream training budgets and reduces the annotated examples needed to reach comparable accuracy.
In briefMemory or a Clock? Auditing a Memory Advantage in XLand-MiniGrid
SummaryThis paper examines whether an LSTM agent's 25-point success advantage over a memory-free MLP on a partially observable XLand-MiniGrid task really comes from memory. The authors separate training seed, map and evaluation noise across 60 runs, and compare MLPs and LSTMs on unseen maps. They find that most of the apparent memory advantage is explained by knowing the time step, and recommend giving memory-free baselines that information.
WhySeparating seed, map and evaluation noise and adding a time-step baseline shows the claimed memory advantage mostly reflects knowing time.
In briefTask format's role in adaptation: LoRA vs. full fine-tuning for prompt-safety classification
SummaryThis paper compares LoRA low-rank adaptation with full fine-tuning for classifying prompt safety, using matched 400M encoder-decoder models across three task formats with up to 15 seeds per condition. At pre-specified learning rates, full fine-tuning outperforms LoRA by 2-3 points, and having LoRA write a reason after its verdict closes the gap. Trained models beat a prompted 8B model in-domain but trail it on two of three out-of-distribution datasets, flagging 27-66% of benign but harmful-looking prompts as unsafe.
John Alderete, Srirakshith Betageri, Jennifer Flynn et al.
In briefRectified Language Flow: A Fast, Parallel Text Generator Built on Rectified Flow
SummaryThis paper introduces Rectified Language Flow (ReLF), a language model that generates text by transporting noise to token embeddings using rectified flow. It combines fixed Hellinger-SVD token embeddings, cross-entropy-trained self-conditioning, and regularized reflow for few-step generation. According to the authors, on OpenWebText with 64 function evaluations ReLF achieves lower generative perplexity than autoregressive baselines at matched diversity across three model sizes, and after two reflow rounds it still outperforms them with 16 evaluations while generating 1,024-token sequences 15 to 19 times faster at batch size 1.
In briefGrounded exploration for open-ended image editing: first define feasible edit directions, then turn them into concrete instructions for more varied, grounded results.
SummaryThe paper addresses open-ended image-editing requests, such as "decorate this room," that allow multiple valid outcomes. It proposes a two-stage framework that first identifies diverse, feasible edit scopes and then realizes them as executable edit instructions. Across 300 tasks, it reports 24.6% higher mean semantic dispersion and 17.1% higher mean groundedness than direct diversity prompting, with more valid and distinct results.
In briefIntentDelta: A Human-Built Benchmark for Evolving User Intent
SummaryThis paper introduces IntentDelta, a human-constructed multi-turn benchmark with 638 trajectories and 3,766 turns across WebShop and TravelPlanner. Each turn pairs a user utterance with a cumulative intention state and human reference actions. Eleven agents were evaluated on intention recovery and action quality, and agents struggled to preserve earlier requirements and to separate soft preferences from must-have constraints. The authors also find that correctly recovering intent does not guarantee successful execution.
In briefLinking protein motion to function via DynaReason-Pro
SummaryThis paper introduces DynaReason-Pro, a framework that predicts GO terms and generates natural-language functional explanations from the conformational ensembles of intrinsically disordered proteins. It is evaluated on a curated benchmark from the Protein Ensemble Database, using splits that keep sequence and structure clusters disjoint to prevent data leakage. Evaluation stratified by conformational variability shows the clearest gains when variability is high.
In briefProMo3D: Finding Motion Signals in Frozen 3D Foundation Models
SummaryThis paper asks whether frozen 3D foundation models already encode motion information that lightweight readouts can extract, without retraining large decoders. The authors propose Probing-based Motion Distillation (PMD), trained on only 160 YouTube-VOS videos on a single 16 GB GPU. PMD transfers zero-shot to DAVIS, SegTrackv2, and FBMS-59, outperforming training-free heuristics and approaching heavier supervised and self-supervised methods.
Xiaoang Zhang, Mert Kiray, Yordanka Velikova et al.
In briefHas UMLS Injection into Language Models for Medical QA Been Fully Explored?
SummaryThis paper reviews existing work on injecting UMLS-derived knowledge graphs into language models for medical question answering. It critically analyzes the chosen UMLS sources and how they are used, testing each choice in isolation. The authors identify methodological issues not previously discussed, such as low UMLS coverage and weak subgraph retrieval pipelines, and offer actionable recommendations for using UMLS as an external knowledge source.
Yiran Zeng, Giacomo Frisoni, Lorenzo Molfetta et al.
In briefMTA-Zero: Self-improving multimodal deep search through online question generation and solver co-training
SummaryMultimodal language models struggle with open-world, multi-step search that needs image grounding, tool selection, and web evidence integration. MTA-Zero trains a question generator that retrieves evidence from image-based seeds and builds verified multi-hop questions online, while a separate solver selects tools and answers them. The generator is rewarded for questions near the solver's learning frontier, and both policies are optimized with GRPO. On six benchmarks, the Qwen3-VL-8B solver's mean accuracy rises from 36.91% to 51.58%.
In briefUsing neighboring captions as inference-time evidence to improve image-text retrieval
SummaryThe paper proposes using other captions of the same image as evidence at inference time, fusing them with image similarity to improve text-to-image retrieval with frozen dual encoders. Gains are largest when captions come from the same annotators as the queries and are minimal from a different annotation source. The authors analyze why the advantage arises and report that on compositional benchmarks the captions help by matching human descriptions rather than improving image-text grounding.
In briefOne Scan Is Enough: Demonstration-Free Adaptation for Language-Guided Navigation
SummaryThe paper addresses the performance drop of language-guided navigation policies when deployed in new buildings, proposing to replace human demonstrations with a single brief iPhone scan. Scan2Instruction turns the scan into a full training environment and automatically synthesizes executable navigation episodes with grounded instructions. PRPT-KS adapts the policy using a closed-form Pareto utility computed directly from the reconstructed scene, without a learned critic or human labels. Across five real environments, the method reaches an 85% real-robot success rate versus 49% for a baseline post-trained on human demonstrations, and 82% versus 41% in unscanned regions of the same buildings.
In briefSegmenting moving objects while the camera moves, using the OWL function
SummaryThis paper presents an analytical framework for separating multiple independently moving rigid objects seen by a single moving camera, based on the OWL perception function. It clusters per-point estimates of relative translation direction, which are recovered from perceived looming and rotation, to partition the scene into rigid bodies. The three-dimensional framework is derived using quaternions, and the authors state it is parallel, invariant to camera rotation, and requires no training data, depth, or semantic priors.
In briefStructured Pruning and Adaptive Fine-Tuning to Improve PPG Foundation Models for Wearable Signals
SummaryThis paper addresses the poor adaptation of PPG foundation models, trained mainly on hospital oximeter data, to sparse and low-quality wearable device data. The authors find that pretrained models contain inherent sparsity and redundancy, and propose SPAFT, a framework that prunes and then fine-tunes within a compact, more relevant parameter space. Across four models, five datasets and ten downstream tasks, SPAFT reports win rates of 82.8% on regression and 87.5% on classification.
In briefLipschitz-controlled greedy orthogonal polynomial manifolds for dimensionality reduction and reduced modeling
SummaryThis paper represents the polynomial terms of a polynomial-manifold decoder with features that are orthonormal on the training data. Based on this, it introduces Greedy-OPM, a greedy algorithm that jointly selects encoder modes and polynomial features to minimize reconstruction error under a prescribed bound on the decoder's Lipschitz constant. Numerical experiments show that this constraint improves generalization away from the training data while keeping reconstruction accuracy high.
Tomoki Koike, Benjamin Peherstorfer, Elizabeth Qian
In briefThe Geometry of Limits in ReLU Function Spaces: a study of how the set of functions a fixed-architecture ReLU network can produce behaves under limits.
SummaryThe paper studies the geometry and closure of function spaces realized by ReLU networks with fixed architecture. It proves a general result on limits of continuous piecewise-linear functions and characterizes functions realizable by shallow networks with vector outputs. The authors show that fixed-width shallow scalar ReLU realization spaces are closed in the weighted supremum norm, but construct an architecture (2,3,2|1) where closedness fails, disproving a conjecture by Petersen, Raslan, and Voigtlaender.
WhyProves lower semicontinuity of polyhedral complexity for CPWL limits and disproves a conjecture via an explicit counterexample. Rigorous, but impact is mostly theoretical.
In briefRoboReason: Scalable Data Generation to Improve Vision-Language Models' Reasoning for Robotics
SummaryThe paper notes that evaluating vision-language models (VLMs) on robotics tasks in the real world is costly and hard to scale. It proposes spatial and temporal ordering tasks using stereo and wrist camera images from real robots, covering frame ordering, depthwise ordering, and left-to-right ordering. Experiments show several state-of-the-art VLMs struggle on these tasks. The authors then introduce an autonomous training data generation pipeline and report that both supervised finetuning and reinforcement learning on this data improve performance on ordering and downstream robotics and visual reasoning tasks.
In briefInfiniteWorld: Testing whether video models trust visual evidence over familiar physics
SummaryThis paper introduces InfiniteWorld, a benchmark that pairs physics questions with simulator videos whose physical parameters are systematically altered. Five tasks test law identification, parameter estimation, temporal reasoning, counterfactual comparison, and law induction under matched text-only and video input settings. The evaluation finds that video-language models persistently prefer familiar Earth-normal physical rules, often giving those answers even when the observed dynamics have changed.
In briefTriageRAG: Failure-aware self-training to adapt retrieval-augmented models to specialized fields
SummaryThis paper proposes TriageRAG, a self-training method for adapting retrieval-augmented generation systems to specialized domains such as medicine when only an unlabeled corpus is available. The model probes each self-generated question with and without retrieved context and sorts it into four categories, then applies training signals matched to each category. Experiments on 11 datasets and CS-Bench report that it outperforms a budget-matched SimRAG and recovers closed-book performance where retrieval hurts the base model.
In briefA Practical Guide to System 1 Models: Fast Decisions Beyond LLMs
SummaryThe paper surveys System 1 decision models, which map an input directly to a structured decision in a single forward pass, for routine tasks such as routing, gating, scoring and classification that need sub-second latency. It proposes a five-family taxonomy covering architecture, training and deployment trade-offs. Calibration is a central theme, and the guide also covers how these models combine with slower LLMs in agentic workflows and cascades.
In briefA neural model that learns atmospheric chemistry quickly and picks its own training data
SummaryEarth-system models usually simplify atmospheric chemistry because near-explicit mechanisms like the MCM are too costly to solve in every grid cell. The authors present FNO-MoE, a Fourier neural operator with a mixture of experts that predicts the full near-explicit chemical state over 30 minutes in one call. The network uses disagreement among its experts to select the most informative box-model simulations for training, which yields better accuracy than random or evenly spread sampling at equal cost. Fine-tuning on longer trajectories keeps it stable over multi-hour runs.
WhyIntroduces FNO-MoE with disagreement-driven active sampling for long-window near-explicit chemistry; novelty is notable, but evidence rests on box-model experiments with limited comparisons.
In briefCFL-aware edge design for graph neural solvers of hyperbolic transport
SummaryThis paper studies whether explicitly giving graph neural networks an edge-local CFL signal helps them model hyperbolic PDEs with shocks. The architecture is held fixed while only the edge input form changes. The authors report that CFL conditioning helps most when edge features already separate interaction direction from spatial scale, and the benefit grows with larger time steps.
Jinyoung Jeong, Joseph B. Choi, Xinlun Cheng et al.
In briefEntropy-Fluctuation Detection of Black-Box LLM-Generated Text
SummaryThis paper studies zero-shot detection of machine-generated text when the source model is unknown. The authors find that existing dual-model and resampling detectors perform poorly on instruction-conditioned text. They propose the Relative Peak Prominence Rate (RelPPR), a metric that captures structural differences along the entropy trajectory, and report state-of-the-art black-box detection across model families and datasets.
WhyProposes RelPPR, a new entropy-fluctuation metric, validated across many model families and datasets, offering useful black-box detection for this subfield.
In briefStateWise: Teaching Language Models to Keep Track of Changing Rules in Multi-Turn Chats
SummaryThis paper studies how chatbots lose track of user constraints across multi-turn conversations, either forgetting active rules or re-applying ones the user withdrew. The authors propose StateWise, which trains a single model to maintain its own revisable constraint state and then uses a preference stage to discourage responses that enact revoked rules. On two backbones and unseen constraint families, it reduces relapse by 8 to 22 points compared with supervised self-training and raises conversation success.
In briefRoleHear: Pseudo-Labeling for Semi-Supervised Speaker-Role Streaming Speech Recognition
SummaryStreaming speech recognition transcribes speech with low latency but does not label which speaker role said each word, which real-time uses such as customer-service analytics require. The paper presents RoleHear, a semi-supervised pipeline in which a single streaming transducer emits role tokens inline, marking turn boundaries and then each speaker role a few words later. An offline teacher model relabels unlabeled audio, and a multimodal large language model provides silver role labels where gold annotations are unavailable. Tested on customer-service calls with gold roles and earnings calls with silver roles, RoleHear attributes roles accurately, keeps transcription quality, and narrows the gap to the offline teacher.
In briefAll the Facts Right, Still the Wrong Answer: Vision-Language Models Struggle to Combine Evidence
SummaryThis paper studies why multimodal models fail when they must combine several pieces of evidence into a decision that satisfies all constraints. The authors evaluate controlled subset-selection and path-finding tasks, linking accuracy on individual facts to validity of the overall decision. They find that models recover local facts accurately yet often fail globally, with Gemma-3-27B reaching only 19.75% global validity on eight-item visual instances.
In briefC-DFF: Concept-level diagnostics for checking whether simulated and real driving perception rely on the same evidence
SummaryThis paper presents C-DFF, a diagnostic method that checks whether a frozen perception system relies on the same concept-level visual evidence for real and simulated images. It extracts decisive evidence, maps it to a shared semantic concept space, and compares concept-importance vectors, also indicating which scene concepts become more or less decisive in simulation. It was evaluated on 2126 matched KITTI and Virtual KITTI 2 pairs across three perception tasks. The authors report that matching outputs alone does not establish mechanism parity.
Danial Safaei, Siddartha Khastgir, Matthew Higgins et al.
In briefRoad Infrastructure Perception for Autonomous Driving: A Survey of Taxonomy, Representations, Tasks and Challenges
SummaryThis survey reviews research on road infrastructure perception for autonomous driving, which recovers road elements such as lane markings, drivable areas, and traffic signs. It groups existing work into geometric-structural, functional-spatial, and regulatory-semantic categories and reviews sensing modalities, scene representations, methods, benchmarks, and evaluation protocols. The authors note that current benchmarks have limitations in relation reasoning, temporal consistency, robustness, and planning-oriented assessment, and they discuss future directions.
In briefVision-based autonomous vehicle navigation in simulation using PPO reinforcement learning
SummaryThis paper evaluates whether Proximal Policy Optimization (PPO), a reinforcement learning algorithm, can learn a policy that drives a simulated vehicle along road lines. The camera image is the input, and the outputs are rear-wheel speed and front-wheel steering angle. The learned policy successfully navigated the vehicle in a simplified environment, and the authors suggest improvements for more complex scenarios along with a framework for other developers to test algorithms.
Dinis Fernandes, Tiago Ribeiro, Ant\'onio Ribeiro et al.
In briefRATE: A New Protocol for Evaluating Machine Translation at Human Level
SummaryAs machine translation nears human quality, existing gold-standard methods (MQM and ESA) struggle to catch subtle errors. The authors built a high-quality English-Russian dataset from WMT24 using professional linguists and introduced RATE, a protocol with a streamlined error taxonomy, expanded severity ratings, and multidimensional scoring of accuracy and fluency. They find that some state-of-the-art systems may surpass human translations in accuracy but still lag in fluency.
Dmitry Popov, Vladislav Negodin, Ekaterina Enikeeva et al.
In briefProtAgent: Training Compact Protein Tool Agents with Step-Level Credit Assignment
SummaryThis paper presents ProtAgent, a two-stage post-training pipeline for a compact model that selects bioinformatics tools for protein annotation. It first uses supervised fine-tuning to learn the tool interface, then applies on-policy REINFORCE with rewards for intermediate tool calls and a final answer reward. On a fixed closed-loop protocol with real tools, the 8B ProtAgent-8B reached 61.00% success under the annotation-coverage metric, versus 41.60% for a 671B model. The authors note the metric measures recovery of reference annotations and does not by itself establish biological correctness.
In briefPOD-RoPE: Rotary position encoding over physical observation distributions for robot AI
SummaryThis paper proposes POD-RoPE, which represents each token's observation as a distribution over camera centers, viewing directions, and sampling times, and encodes relative relationships through expected rotations. Integrated into the Fast-WAM robot policy, it raises mean success on all 50 RoboTwin 2.0 tasks from 91.83% to 92.88%. The authors also report faster forward-plus-backward speeds than the compared RoPE variants.
In briefBeyond Prompts: A Systematic Study of LLM-Based Machine Translation Evaluators
SummaryThis paper systematically studies how instruction design, few-shot demonstrations, backbone model choice, and multi-step workflows affect large language model evaluators of machine translation. The authors collected new expert annotations on WMT24 and WMT25 English-Russian translations and validated transfer on benchmarks for five other language pairs. They present ORBIT, a family of LLM-based evaluators whose strongest configurations reach expert-level performance at roughly 20 times lower cost than expert human annotation.
Dmitry Popov, Roman Bokhyan, Ekaterina Enikeeva et al.
In briefA distribution-matching view of how large language models are trained
SummaryThis paper proposes a unified framework that explains several large language model training algorithms as minimizing a divergence between the current model and a target distribution. The authors find that on-policy distillation outperforms classical knowledge distillation when the model's distribution is close to the target, but does worse when they are far apart. They also show that when the target depends on the model being optimized, estimation variance can degrade the target, and they propose Conditional KL regularization to mitigate this while preserving mode diversity.
In briefA quantum k-nearest neighbors method that compares vector magnitudes, using fewer qubits for Euclidean distance
SummaryThis paper proposes a quantum encoding that lets Euclidean distance be computed between vectors with different L2 norms by appending one component to each vector, and applies it to quantum k-nearest neighbors (QKNN). The authors say it is simpler to understand and implement than comparable methods and uses fewer qubits. They benchmark their implementation against a recent Euclidean-distance QKNN method on standard labeled datasets. They report better results under simulated realistic noise.
In briefProof that polynomial residual networks are generically identifiable
SummaryThis paper proves identifiability results for constant-width residual networks with polynomial activations and bias terms. For activations of degree at least five with generic coefficients, the generic fiber is a singleton. For degree at least three with no linear term after translating the input, the generic fiber is given by translating biases and scaling by real roots of unity. The authors also derive necessary conditions for critical points of the ResNet parametrization, which are central in singular learning theory.
Isabel Dahlgren, Giovanni Luca Marchetti, Kathl\'en Kohn
