In briefOptimal sample complexity for pointwise replicable agnostic learning
SummaryThe paper studies how many independent examples are needed so that two predictors trained with the same random seed agree at each fixed query with high probability. For binary classes of finite VC dimension, it gives a sample complexity bound that matches a corresponding lower bound. The result removes the remaining logarithmic overhead on countable domains.
In briefHCFM: Helmholtz-decomposed flow matching for detecting anomalies in time series
SummaryThis paper proposes HCFM, an unsupervised framework for time series anomaly detection built on flow matching generative models. It splits the learned vector field into transport, compression, and residual heads, constraining the transport head to be divergence-free so that divergence is estimated only in the compression head. The authors report that on large-scale univariate and multivariate benchmarks, HCFM improves detection accuracy over 16 baselines while keeping inference efficient.
In briefPRET: A few-shot system for recognizing many cancer types without task-specific training
SummaryThis paper introduces PRET, a few-shot system that recognizes cancers across organs, hospitals, and tasks without training for each task. Evaluated on 23 international benchmarks with 4,484 whole-slide images, it outperforms existing approaches on 20 tasks and exceeds 97% AUC on 15 benchmarks. For lymph node metastasis detection, it reaches clinical-grade performance using only eight slide examples and outperforms 11 pathologists.
WhyProposes training-free few-shot pan-cancer recognition, validated on 23 benchmarks with 4,484 slides and outperforming 11 pathologists in lymph node metastasis detection.
In briefPACT: NVFP4 data-flow LoRA fine-tuning for consumer GPUs, a method for adapting language models on affordable graphics cards
SummaryThe paper presents PACT, a LoRA fine-tuning system for language models on consumer GPUs. It speeds up frozen projections with native NVFP4 matrix operations and reduces activation memory by reusing quantized RMSNorm outputs and saving gate/up activations in NVFP4. Evaluations on a 16 GB RTX 5080 across several models and tasks report lower memory use and faster fine-tuning with competitive accuracy.
Lei Gao, Hossein Entezari Zarch, Chaoyi Jiang et al.
In briefSteadyThought: Reducing LLM “Under-Thinking” with Thought-Level Preference Optimization
SummaryLarge reasoning models often switch between reasoning paths too frequently and fail to sustain promising thoughts, a problem the authors call "under-thinking." They propose SteadyThought (ST), a thought-level preference optimization framework that treats switching points as a preference issue, preferring single-path continuations that yield higher-quality reasoning. Experiments across multiple models and datasets report token reductions of up to 63.8% while maintaining strong reasoning accuracy.
In briefA normalizing flow method for neural-enhanced track-before-detect
SummaryThis paper proposes a neural-enhanced track-before-detect framework that works directly on complex-valued raw signals and learns latent representations of unmodeled propagation effects at runtime. The signal power intensity of each object is modeled by a normalizing flow, trained online in an unsupervised way within sequential Bayesian inference, with the objective derived by maximizing the evidence lower bound. Inference combines belief propagation with sequential Monte Carlo methods for scalability.
In briefVLA-Rolling: Reusing Discarded Robot Action Tails to Denoise Faster and Control Better
SummaryFlow-matching vision-language-action models usually execute only a short prefix of each generated action chunk and discard the rest. This paper introduces VLA-Rolling, a lightweight frozen-model inference framework that reuses these action tails through inheritance, per-action confidence estimation, and confidence-guided renoising and denoising budget allocation. Across four VLA models and four benchmarks, it lowers action-chunk inference cost in all 16 pairs, with up to a 1.48x speedup. It also improves success rate in 14 pairs by up to 15.08 percentage points.
In briefScam Town: Can Scam Education Spread Like a Vaccine?
SummaryThe paper builds Scam Town, a simulated community where large language models play residents, to test whether scam warnings spread through word of mouth. Everyday conversation cut scam victims by roughly a fifth across 80 simulated towns. Official warnings that reached most residents and sparked discussion roughly halved scams in simulation. The authors present these as predictions to test with real people, not proof.
In briefBeyond the Teacher: Preserving Overshoot in Agentic On-Policy Distillation
SummaryThis paper studies on-policy distillation (OPD) for multi-turn agent tasks, where a smaller student model can temporarily outperform its stronger teacher but then regress as training continues, a phenomenon the authors call distillation overshoot. The authors attribute this to two competing effects: repairing weak student decisions and eroding successful choices the teacher disfavors. They propose Success-Preserving On-Policy Distillation (SPOD), which masks conflicting teacher supervision on successful choices and adds a soft probability floor, and report gains over ordinary OPD across two model families and three benchmarks.
In briefOne Head Away from Language Confusion: Single Attention Heads Across Model Families
SummaryLarge language models sometimes reply in an unexpected language, a problem called language confusion. The authors ablate individual attention heads in 12 checkpoints from seven model series and test them on FLORES-200 continuations and Language Confusion Benchmark chat replies. In ten models, zeroing one particular head causes far more replies to leave the expected language than zeroing other heads in its layer.
WhyShows single attention heads drive language confusion, validated by systematic ablations across seven model series and two benchmarks.
Jiwon Jeong, Anjelo Jann Laroza, Chaewoon Yoon et al.
In briefStrictly Replicable Learning: Theory of Reproducible Boolean Learning with Shared Randomness
SummaryThis paper studies strict replicability in realizable Boolean learning, where two independent training samples run with the same random seed must yield exactly the same hypothesis. It gives sample complexity bounds for finite classes under unknown and known instance marginals. The paper also examines product threshold chains, finite classes of VC dimension one, and the implementation cost of shared randomness.
In briefBuilding blocks for symplectic neural networks, with three new single-layer designs that can be more accurate than shear-based networks.
SummaryThe paper proposes three families of exactly symplectic neural network layers: learned-type, moving-plane shear, and fiber-preserving layers, and characterizes which residual symplectic maps they can represent. In experiments on sixteen systems, layers built on nonlinear integrals are far more accurate than SympNets and HénonNets when the Hamiltonian contains functions of those integrals. Reported gains include over a hundredfold accuracy improvement on nonlinear Schrödinger lattices, while fiber layers improve on learned-plane shears by at most a factor of two on generic non-separable systems.
In briefSynergy-aware residual flow matching for predicting single-cell responses to perturbations
SummaryThis paper proposes SynGeneFlow, a flow matching method for predicting single-cell expression responses to genetic and chemical perturbations. It anchors the source distribution at an additive prior and learns a residual flow toward the perturbed state, modeling non-additive gene interactions with a static gene-gene mask and a perturbation-dependent bias. The authors evaluate it on genetic and drug perturbation benchmarks under several settings, reporting a robust residual-based framework.
In briefLLM Agents Supplying Physical System Models: Calibration and Acceptance in Modelica
SummaryThe paper lets LLM coding agents build Modelica models from library components, using specifications with data, for four simulated systems and a real switching magnet. It asks what acceptance of a delivered model actually certifies. The authors propose accepting a model by membership in the set of models consistent with the handover, rather than matching one reference, and add checks such as a physical invariant.
In briefDecentralized online optimization for exp-concave losses: agents minimize a sequence of global loss functions using only local computation and communication, with regret bounds
SummaryThe paper studies decentralized online convex optimization, where local agents minimize a sequence of global loss functions using only local computation and communication. It proposes Decentralized Online Newton Step (D-ONS), which achieves a logarithmic regret upper bound for exp-concave losses, and establishes a nearly matching lower bound. The approach is also extended to decentralized online linear regression with unbounded feasible domains.
WhyGives the first logarithmic regret bound for decentralized online exp-concave optimization, with a nearly matching lower bound; strong theory, but no experimental validation described.
In briefA Monte Carlo method extended to vector-valued quantities using cheaper models
SummaryMultifidelity Monte Carlo reduces the cost of estimating statistics from an expensive high-fidelity model by using cheaper low-fidelity models as control variates. This paper extends the method to vector-valued quantities of interest by replacing scalar coefficients with matrices and minimizing a weighted mean-squared error, with closed-form optimal full-matrix coefficients. Experiments show the best coefficient class depends on pilot size relative to the number of covariance quantities, and reusing pilot samples lowers mean-squared error.
Zhibo Dai, Hamdi A. Tchelepi, Daniel M. Tartakovsky
In briefIECLA: Interpretable Emotion Contribution Learning for Editing Image Emotions
SummaryThis paper proposes IECLA, a framework for affective image editing that aims to evoke a target emotion while keeping a source image's content and structure intact. It uses counterfactual interventions to build signed object-level supervision and trains a predictor to localize regions that support or oppose the source emotion. A target-aware editing agent then turns these regions into localized edits for existing editors. The authors report stronger target-emotion expression across six editing backbones, with preserved content and improved user-rated expression and naturalness.
In briefLatent world models should plan along the latent manifold
SummaryLatent world models such as LeWM plan by finding action sequences whose predicted final latent is closest to the goal latent in Euclidean distance. The authors find this distance is reliable only near the goal, while farther away it creates spurious low-cost regions that trap the planner. They propose Neighborhood Graph Cost (NGC), which computes shortest-path distances on a mutual nearest-neighbor graph over training latents, changing only the terminal cost without retraining. NGC improves LeWM's planning success across all four benchmark tasks, nearly solving TwoRoom and Reacher.
In briefConjugate Compilation: Self-Improving AI Agent Co-Optimization for Faster Speculative Decoding
SummaryThe paper proposes Conjugate Compilation, which treats speculative decoding's small draft model and large target model as a cross-model compilation problem and optimizes them jointly within their shared execution. An LLM-driven agentic search, guided by Monte Carlo tree search, edits optimization plans using measured performance feedback. Across six single-GPU configurations on NVIDIA B300 and RTX 5090, it achieves a geometric-mean per-round speedup of 1.37x, rising to 1.45x with interleaved kernels and reaching 1.59x at best.
Annabelle Sujun Tang, Rohan Mahapatra, Christopher Priebe et al.
In briefFineSplat: Fine Patches and Local Groups for Detailed 3D Gaussian Splatting
SummaryThis paper presents FineSplat, a feed-forward 3D Gaussian Splatting framework that builds on fine-grained image patches to reconstruct scenes quickly. It combines view-wise and 3D spatial grouping so that neighboring views and geometrically related tokens can interact efficiently. The authors also use reconstruction error as rendering-based feedback to recover finer geometry and appearance. They report state-of-the-art results under the same or smaller Gaussian budgets than existing methods.
In briefMAMath-Bench: A Benchmark for Checking Step-by-Step Math Reasoning by Multimodal Multi-Agent Systems
SummaryThis paper introduces MAMath-Bench, a benchmark of 1,003 math problems across 11 categories, with text, image, and audio inputs, for testing multi-agent AI systems. It uses milestone annotations to assess final-answer correctness, reliability of intermediate reasoning steps, and robustness to adversarial reasoning-chain perturbations. The authors report that multi-agent collaboration does not consistently beat strong single LLMs on final answers, that it can improve intermediate reasoning coverage but may add redundancy, and that current systems remain vulnerable to subtle structural perturbations.
In briefNo Escape from the Quadratic Barrier: Adaptive Attacks on Linear ℓ2 Sketches
SummaryThe paper studies whether randomized linear sketches, such as the Johnson-Lindenstrauss transform, that estimate Euclidean norms from a fixed low-dimensional projection can resist adaptive black-box queries. The authors show that a quadratic query lower bound, up to logarithmic factors, is optimal even against a computationally unbounded responder that can change its randomized rule after each query. Their attack forces large relative errors with constant probability using queries and adaptive rounds that scale with the sketch dimension. The results show inherent limits to defending a fixed compressive linear representation by adapting the prediction rule.
WhyProves the quadratic barrier is near-optimal for adaptive attacks, improving prior privacy-based guarantees via rigorous theoretical analysis.
In briefFixing Failed Proof Steps in Lean 4 Instead of Only Resampling
SummaryThe paper notes that the Lean proof assistant returns precise error messages for failed steps, which pure language-model sampling discards. The authors propose a bounded, online step-level repair module that fixes a failed tactic at most R times by feeding the compiler's error message to a pretrained repair model. On a 1,990-theorem subset of LeanDojo, the method improves the solve rate over a sampling-only baseline at all tested candidate budgets, with gains of up to 32% and 82 additional theorems solved, at the cost of extra runtime.
Haotian Deng, Chris Farber, Sonal Narasimha et al.
In briefReTrace: Visual Backtracking for Faster Multimodal AI Models
SummaryMultimodal large language models turn images into many visual tokens, which makes attention and memory costs high. Existing methods prune low-importance tokens early, but this permanently discards information that later layers may still need. The authors observe that middle layers of these models revisit early tokens, a behavior they call Visual Backtracking. They propose ReTrace, which temporarily caches inactive tokens and deletes them only after their information has been used. Tests on Qwen2.5-VL and InternVL3 report nearly 90% lower FLOPs and KV cache use, a 2.83x prefilling speedup, and strong performance on vision-language benchmarks.
In briefUpLoop: Upgrading Pretrained LLMs into Gated Looped Models After Training
SummaryThe paper introduces UpLoop, a post-training method that converts a pretrained non-looped large language model into a looped one by recurring each selected layer locally. Hidden-wise gates on attention and MLP residual branches are added so later loops act as progressive refinements. Starting from Qwen3-4B Base with SFT, the authors report a 6.17% average accuracy gain across 20 benchmarks at R=3, and state that it surpasses the saturation at two loops seen in prior looped models.
In briefUtoMe: Uncertainty-guided token merging to cut compute in weather AI models
SummaryUtoMe is a token merging method for weather transformers that uses observation uncertainty, combined with dynamics saliency, to decide which input tokens to keep and how to merge the rest. It is tested on ERA5 72-hour forecasting at two spatial resolutions. At matched token budgets it reduces overall error compared with train-time ToMe. At 1.40625° with 50% token reduction, UtoMe-E lowers overall error from 67.32 to 53.00 at 0.55x encoder FLOPs.
In briefBlock Recursive Transformers: Structured Parameter Sharing for Language Models
SummaryThis paper proposes Block Recursive Transformers (BRT), which split a model's layers into contiguous blocks and reuse each block recursively before moving to the next. It sits between two extremes: repeating each layer independently and cyclically reusing the whole stack as one shared function. On decoder-only models up to 1.7B parameters, BRT improves validation negative log-likelihood and downstream accuracy over static recursive sharing baselines.
Josep Marques, Hyunyoung A. Cho, Youngeun Kim et al.
In briefIndustrial Agentic AI: LLM-driven next-generation agents for smart manufacturing, explained
SummaryThis paper systematically reviews industrial agents powered by large language models for smart manufacturing. It analyzes five key challenges, proposes a unified taxonomy covering application scenarios, embodiment forms, and five capability levels, and describes a four-layer architecture for closed-loop task execution. The authors also review representative applications across the industrial lifecycle and outline future research directions.
In briefWhen Better Acceptance Helps: A Controlled Study of Speculative Decoding
SummaryThis paper examines when the extra acceptance length in speculative decoding justifies its execution cost. The authors set up a comparison protocol covering training, inference, and evaluation, and retrained six drafter recipes on shared target-generated data across ten benchmarks and two target models. Under shared conditions, gaps seen in published comparisons shrink substantially, and the payoff of improved acceptance depends on task uncertainty and serving conditions.
In briefMultiAltitudeSR: A Super-Resolution Method for Multi-Altitude Farm Drone Images Using Ground Sampling Distance
SummaryThis paper proposes MultiAltitudeSR, a framework that conditions image enhancement on the source and target ground sampling distances to improve crop monitoring images from drones. Trained jointly on real UAV images captured at 10 to 70 meters, it supports 21 altitude pairs without per-pair fine-tuning. Training on real image pairs outperformed the tested synthetic degradation pipeline, and the enhanced images showed task-dependent benefits in downstream zero-shot segmentation case studies.
In briefCertified Sparse Attention via Perturbed Retrieval: a long-context decoding method that gives verifiable error guarantees for each query
SummaryThe paper proposes a sparse attention method that adds independent Exp(1) noise to each key's score, turning threshold retrieval into sampling keys in proportion to their attention weights. It estimates the output with a Horvitz–Thompson estimator and provides confidence bounds valid across all threshold levels, so each head stops once its error certificate meets the target. Across three LLMs and ten settings at 16K–128K tokens, the authors report no violated certificates and lower 90th-percentile error than baselines at equal keys read or equal bytes. End-to-end task scores and PG-19 perplexity stay close to dense attention, and CPU decoding at 32K–128K tokens is 1.1–1.4 times faster.
WhyExp(1) perturbation turns threshold retrieval into attention-weighted sampling with simultaneous confidence bounds; experiments on three LLMs and ten settings show substantially lower error.
In briefA dataset of 118 million state media articles from 136 countries, with state-control annotations
SummaryThis paper introduces WorldStateMedia, an open multilingual dataset of 118 million state media articles from 136 countries in 114 languages, with each outlet annotated for level of state control. The authors find that the share of editorially controlled outlets in English web crawls rose from 39% in 2013 to 93% in 2025. In a retrieval-augmented generation experiment on territorial disputes, retrieving a country's state media shifted model answers toward that country's claims, most strongly for English-language media.
WhyLarge-scale audit of state media in pretraining corpora plus a RAG experiment showing answer shifts gives novel, well-supported evidence.
In briefCorrectRTL: Improving LLM-generated hardware code through decomposed self-correction
SummaryThis paper addresses using large language models to generate RTL (hardware description) code from specifications. It introduces ScaleBench, a benchmark of 200 complex RTL designs, and CorrectRTL, an agentic framework that decomposes scaffold construction from detailed implementation before correcting. On ScaleBench, Pass@1 rises from 20.5% to 49.0%, and on VerilogEval it reaches up to 96.1%.
In briefBenchmarking LLM-text detectors with item response theory as a measurement tool
SummaryThis paper proposes a benchmarking framework that uses multidimensional item response theory to treat benchmark texts as measurement items and detectors as respondents. Applied to 32K human-written and machine-generated texts across four domains and 14 detectors, it finds overlapping but distinct human and machine distributions, and detector ability patterns that vary by domain and are not fully captured by aggregate rankings. Clustering in the estimated item space yields a 2,088-text subset (6.5% of the benchmark) that largely preserves the full benchmark's detector comparisons.
In briefRubrics Set the Standard: Shapley Attribution for Checklist-Based LLM Judges
SummaryThis paper studies large language model judges that grade open-ended outputs against checklist rubrics. In controlled tests with planted flaws and on the SummEdits and RAGTruth human-labeled datasets, rubric content changed the pass rate of correct outputs far more than swapping the judge model did. The authors propose a Shapley-value method to attribute credit and blame to individual checks, and a change-control gate that keeps rubric revisions from drifting.
In briefPRECOCIAL: Automatically optimizing code-based curricula for faster pretraining
SummaryThis paper introduces PRECOCIAL, which uses an LLM to iteratively rewrite programs that generate synthetic pretraining data, scoring each rewrite by pretraining a transformer on its data and fine-tuning on the downstream task. It keeps a single network throughout the search to reduce cost. On four language and code datasets, the optimized programs achieve lower perplexity than several baselines at a fixed fine-tuning budget and improve performance on four reasoning and problem-solving tasks.
SummaryAs AI outputs grow longer and more complex, it becomes harder for humans to oversee them safely. The paper examines whether human and AI judgements, which make different mistakes, can be combined into a stronger oversight signal. The authors release the largest known suite of such tasks (3,957 instances across 8 datasets and 6 domains) and collect 35,613 judgements from 416 human raters. They find considerable complementarity potential, but the assistance methods tested led humans to correct AI errors less often, and hybrid methods added little over the better single source on average.
Sander Lanbo Shi, Miranda Anna Christ, Eric Modesitt et al.
In briefVisual Sufficiency Landscapes: How Query-Dependent Reliability Affects Visual Token Pruning
SummaryThe paper introduces visual sufficiency landscapes, which measure the fraction of sampled visual token subsets that preserve a dense model's answer, reference-answer likelihood, and output distribution at a fixed layer and budget. Experiments on Qwen2.5-VL-7B, InternVL2-8B, and LLaVA-v1.5-7B show that landscapes differ across inputs and across questions on the same image. Adding this probability improves failure prediction for pruning methods such as FastV, VisionZip, and SparseVLM on LLaVA.
In briefMeta Inverse Prompting for Video Generative Models (a method to recover text prompts from videos quickly)
SummaryVideo generative models usually map text or images to videos, with no built-in way to recover controllable text conditions from a given video. This paper proposes MVIP, the first practical framework for efficient video inverse prompting, formulating it as a meta-learning problem that learns a temporally optimized meta-prompt. The approach amortizes inversion, avoiding costly per-video iterative optimization at test time. Experiments report higher-quality inverse prompts, improved reconstruction fidelity over image-based and iterative captioning baselines, and substantial efficiency gains.
In briefDeGMatch: Selecting Training Data for Target Reasoning Skills by Removing Dominant Gradient Patterns
SummaryThis paper studies how to pick long reasoning training trajectories that match a specific target capability, using a small reference set. The authors find existing target-aware methods discriminate poorly on long trajectories, and dominant population-level structure in candidate gradients can hide target-specific differences. They propose DeGMatch, which first builds a target-relevance gate, then reranks candidates using model-specific gradients centered on the population mean with leading principal directions removed. The paper reports it outperforms heuristic, target-aware, and reasoning-specific baselines on mathematics, science, and code.
Changho Lee, Janghoon Han, Rajdeep Mukherjee et al.
In briefLabel-Space Mismatch Skews Cross-Cohort Breast Cancer Subtype Evaluation
SummaryThe paper examines why comparing breast cancer subtype classifiers across cohorts with different annotation standards is unreliable, since METABRIC uses six classes while PAM50 uses five. The authors separate the effects of restricting the label space, applying EM prior correction, and recovering target labels from unlabelled predictions. They also find that multimodal fusion gains do not transfer across cohorts, and they argue that the allowed label set should be reported and label-space compliance applied before any cross-cohort performance claim.
In briefCoworkEnv: Automatically Generating Training Environments for AI Coworkers That Handle Long Multi-Step Tasks
SummaryThe paper introduces CoworkEnv, a framework that automatically builds executable training environments for AI coworkers, including requests, service records, tool interfaces, and grading checks. Teacher executions provide demonstrations for supervised fine-tuning, while student interactions supply rollouts and rewards for reinforcement learning. Fine-tuning Qwen3-VL-8B-Instruct raised the mean Claw-Eval score from 46.6 to 64.1, with a 31.9-point gain on multi-turn tasks. The authors also report an SFT-only extension to ClawMark and preliminary evidence from RL training rewards.
In briefMeasuring LLM Breadth and Reliability Beyond Benchmark Scores via Accessibility Profiles
SummaryThis paper argues that benchmarks score each question using only one fixed wording, ignoring that the same request can be phrased in many equivalent ways. The authors define "accessibility" as the probability a model succeeds across meaning-preserving formulations, which reveals how broadly capabilities extend and how reliably they can be elicited. Across six benchmarks and over twenty models, about half of model pairs with nearly equal mean accessibility rank differently depending on whether one success across five formulations (Access@5) or all five (Reliable@5) is required.
WhyIntroduces accessibility across meaning-preserving rewordings, validated on six benchmarks and over twenty models, revealing breadth-reliability tradeoffs hidden by accuracy.
In briefMemPrice: Adaptive-pricing memory allocation for long-running AI agents
SummaryThis paper addresses how long-horizon LLM agents should allocate limited context as their interaction histories grow, deciding what information to keep, how much context to give it, and at what level of detail. The authors propose MemPrice, which treats active context as a shared resource and dynamically allocates it across memories and representation levels using an adaptive price. The framework is evaluated under matched memory budgets on LongMemEval, LoCoMo, AppWorld, and WebShop across three foundation models.
In briefCasePlay: Self-Play RL from Case Reports for Medical Reasoning
SummaryThis paper proposes CasePlay, a self-play reinforcement learning framework that turns unlabeled clinical case reports into training signals to improve medical reasoning in large language models. It pairs a knowledge-conditioned question proposer with a rubric-based judge that scores clue integration, reasoning depth, and option quality. Across seven medical benchmarks, the authors report better results than existing self-play baselines and an average gain of 3.1 points over the SFT warm-start model, including 6.3 points on LiveClin.
In briefControlJEPA: a trajectory regularizer that keeps hidden states stable using a Lyapunov tube loss
SummaryThe paper notes that next-token training leaves hidden-state trajectories largely unconstrained, and that existing geometric regularizers mainly control direction rather than magnitude. It proposes ControlJEPA, which penalizes violations of a discrete-time Lyapunov contraction condition using a normalized, scale-invariant measure of transversal deviation. The authors report that it improves accuracy and data efficiency across six benchmarks and six model families without architectural changes, and that the smoother trajectories persist during inference. Code is publicly available.
WhyApplies discrete-time Lyapunov contraction to hidden-state trajectory regularization; six benchmarks and six model families support it, but impact is confined to training regularization.
Thai-Khanh Nguyen, Nhi Y. Pham, Uyen N.B. Vo et al.
In briefWhy coding agents over-defend: learning at silent forks
SummaryThe paper finds that coding agents take defensive actions beyond what a task requires, and that these begin at "silent forks," turns where every next action leads to the same outcome. The authors propose ForkDPO, which contrasts proportionate and disproportionate actions at the same prefix at these forks. On a test set of unseen repositories, it lowers a 4B model's disproportionate behavior from 70.3% to 58.9%.
WhyForkDPO contrasts proportionate and disproportionate actions at silent forks, a novel idea; evidence is limited to one 4B model on one unseen-repository test set versus DPO.
In briefMaking Direct Preference Optimization resistant to attacks that flip human preference labels
SummaryDirect Preference Optimization is a common method for learning from human feedback offline, but it can be steered toward an attacker's chosen policy by flipping preference labels. The paper proposes Trim-K and Soft-Trim-K Robust-DPO, which remove or smoothly downweight the K most influential preference pairs. It gives gradient perturbation bounds that hold for any attack with at most K flips, and experiments on three preference datasets validate the analysis.
In briefEpi: A persistent, self-improving AI agent for autonomous research
SummaryThis paper presents Epi, an autonomous research agent that continues investigations across model sessions and refines its methods using accumulated experience. Epi reportedly produced 120 record-setting results, including 20 numerical bound records and 100 construction records. On 30 FrontierOR tasks, its lightweight configuration Epi-turbo scored 5.4% higher than Codex at 64% of its model cost.
WhyProposes a cross-session research agent with 120 record results, a computer-assisted proof, and a benchmark comparison against Codex.
In briefLowNet: A compact looped wavelet network with 24 parameters for solving Poisson equations
SummaryThe paper proposes LowNet, a wavelet-inspired neural solver for Poisson equations that has only 24 trainable scalars across four 5×5 kernels. A residual loop reuses the wavelet block to raise accuracy without adding parameters. The authors report PSNRs of 47.27, 74.14, and 100.74 dB for one, two, and three loops on 10000 images, and claim linear computational complexity.
In briefT-MUON: a tensor-based extension of the Muon optimizer using t-SVD for pretraining large language models
SummaryThe paper proposes T-MUON, which generalizes the Muon optimizer to tensors by grouping related Transformer gradients into third-order tensors and defining updates via the M-transform. The authors show the update is a steepest-descent direction under a tensor spectral norm and provide a convergence guarantee. On GPT-2-style and LLaMA-style models from 51M to 1.3B parameters trained on FineWeb and FineWeb-Edu, head-wise T-MUON with the identity transform reports lower validation perplexity than AdamW, Muon, and several transform variants.</summary>
In briefMLFont: Machine learning for splitting web fonts to load faster
SummaryThis paper presents MLFont, a framework that uses machine learning and optimization to learn how to split web fonts into downloadable slices, grouping characters by frequency, co-occurrence topics, and compact ranges. Across five Google Noto fonts, it reduces per-glyph latency on 3G compared with single-file delivery and equal-bin baselines, while keeping 100% coverage. Tests on 20 websites report cold-cache speedups and bandwidth savings, especially for CJK fonts.
WhyCombines topic models and a genetic algorithm for font slicing, validated across several fonts and 20 websites with latency and bandwidth gains.
In briefTrustAgent-RFF: An agent-based diagnostic framework for reliable radio fingerprinting under domain shift
SummaryThis paper proposes TrustAgent-RFF, a framework that judges whether each radio frequency fingerprinting prediction can be trusted when the receiver, distance, day, or environment changes. It works with a frozen I/Q classifier and needs no target domain labels, gathering diagnostic evidence and combining it into a source-calibrated trust score. A language-model planner chooses which diagnostics to run, and a decision layer outputs Accept, Reject, or Defer. Experiments on five datasets show an average source-to-target accuracy drop of 38.7%, with many wrong predictions still highly confident.
In briefLimits of Jev's one-pass decisions: what it cannot decide
SummaryThe paper tests the limits of Jev and its open replicas, which choose among supplied options in a single forward pass. Accuracy drops when an answer depends on two or more updates to one user, because similar records interfere with each other. Given a stale state summary, the model follows it, and its confidence catches its own errors but not errors inherited from stale inputs.
In briefMeans-ends analysis helps LLMs plan in continuous-state domains
SummaryMost LLM planning research has focused on discrete domains with enumerable actions, an assumption that fails in continuous-state settings. The paper applies Newell and Simon's Means-Ends Analysis to propose Medaka, a structured prompting framework that encourages recursive goal decomposition, algebraic derivation, and per-step state tracking. The authors also introduce Liquidworld, a liquid-mixing testbed for evaluating these abilities. Medaka improves performance across several model families on this benchmark.
In briefStable Signer V3: A Sign Language Avatar System Using 4D Deformable Primitive Splatting
SummaryThis paper presents Stable Signer V3, an interactive high-resolution sign language avatar system built on 4D Deformable Primitive Splatting. It targets three limitations of existing avatars: restricted facial expression, inaccurate identity-specific body deformation, and loss of fine appearance detail. Across ten identities on held-out evaluations, the authors report reductions of 71.4% in facial geometry error, 35.1% in body-boundary error, and 24.2% in high-resolution RGB reconstruction error, while retaining interactive streaming.
In briefDF-PALM: Dynamic Fusion Prompt Learning for Audio-Language Models
SummaryThis paper proposes DF-PALM, a method for few-shot audio classification that adapts class representations based on each audio input and refines audio features with asymmetric cross-attention, while keeping pretrained encoders frozen. On 11 benchmarks with 16-shot evaluation, using PENGI as the backbone, it reports 80.69% mean accuracy and 80.32% macro-averaged F1, exceeding PALM by 4.11 and 4.30 percentage points. The authors also report gains with LAION-CLAP and OpenFLAM and a 0.39% increase in inference FLOPs at 50 classes.
In briefCARTA: A label-free, reusable payment transaction embedding method
SummaryThe paper presents CARTA, a self-supervised encoder that learns a compact 16-dimensional embedding from raw payment transactions without labels. The frozen embedding can be reused for label-efficient fraud probing and unsupervised risk stratification without retraining the encoder. On IEEE-CIS and IBM TabFormer, it outperforms several baselines, though fully supervised CatBoost remains stronger.
Shreyas K. Chandrahas, Jose Munoz, Satish Babu Bellapu
In briefHow Well Can LLMs Read Math Formulas Aloud? A Multi-Domain Test
SummaryThis paper tests five leading large language models on turning LaTeX math formulas into spoken-style written text across six scientific fields, using a new corpus called Read-Aloud Math. Performance is measured with multi-reference BLEU, with GPT-5.2 scoring highest on simple transcriptions and Claude Sonnet 4.5 and Grok 4.1 Fast doing well on extended ones. The authors conclude the approach is promising but sensitive to transcription type, textual context, and formula characteristics.
In briefArctic-SWE: A Reproducible Baseline for Long-Horizon Agent Training via Co-Design
SummaryThis paper presents Arctic-SWE, a fully asynchronous reinforcement learning framework for training long-horizon software engineering agents. It combines established components, such as gated rewards, importance-weight control, aligned training and inference engines, and consistent state across updates, to create a reproducible baseline. Using only raw R2E-Gym data, Qwen3.5-4B reaches 62.0% pass@1 on SWE-bench Verified. After filtering and refining the data, the 4B and 9B models show gains across several SWE-bench benchmarks.
In briefGRaSP: Latency-Aware Retrieval Optimization for Geo-Distributed AI Question Answering
SummaryThis paper proposes GRaSP, a plug-in layer that re-ranks retrieved documents by combining relevance with network latency, cross-region cost, and data-residency limits, without retraining. Tested on a five-node testbed with the Natural Questions corpus, it cut P90 fetch latency by 41% at a moderate latency setting while keeping 78% of relevance-only recall. Selection overhead stays under one millisecond.
In briefA cross-domain comparison of prompting methods for large language models across six benchmarks
SummaryThis paper compares zero-shot prompting, few-shot prompting, retrieval-augmented generation (RAG), and a PE2-inspired automatic prompt optimization pipeline using the same LLM on six benchmarks covering arithmetic, logical, and commonsense reasoning and factual verification. No single strategy consistently outperforms the others, and pairwise statistical comparisons of top configurations showed no significant differences. RAG achieved the highest overall average accuracy of 78.33% when 300 examples were available for retrieval. The authors conclude that prompting techniques should be chosen adaptively based on task characteristics.
In briefATLAS: An Expert-Guided System for Sustainable Satellite Backhaul in Off-Grid Wildlife Monitoring
SummaryThe paper presents ATLAS, a system that uses foundation models to convert expert natural-language intent into semantic scoring logic, then combines it with environmental and energy conditions. A constrained reinforcement-learning scheduler decides which video clips to upload to satellite and when, given limited bandwidth and power. It was deployed for over six months at 18 off-grid salmon-monitoring river sites. ATLAS reached a mean clip-selection accuracy of 87.4%, 26.4 percentage points above baseline, and stayed nearly continuously operational during low-solar periods, where baselines had 16.0% downtime.
In briefA method to reduce blind spots in sparse autoencoders by matching symmetries
SummaryThe paper addresses blind spots in language models, where semantically different inputs can be treated as interchangeable in their representations, a gap standard sparse autoencoder training does not target. The authors add a differentiable V_Gini regularisation term that penalises uneven responses to perturbations within a family. Across GPT-2, Gemma 2 and Qwen 2.5, the method raises weak-tail responses and reduces imbalance without material loss in general text reconstruction, while downstream discriminability is treated as a separate empirical step.
In briefAdaptive gates that let language models follow several steering directions at once across languages.
SummaryThis paper proposes contextual steering, which jointly trains gates that adjust the strength of fixed steering directions based on the current hidden state and interactions between interventions. The approach is used to compose multiple behaviour controls in language models across many languages, attributes, and models. It reportedly outperforms constant-strength steering and four strong baselines, including prompting, while better preserving general knowledge.
Daniil Gurgurov, Akansh Maurya, Josef van Genabith et al.
In briefHuman-centered explanations using mental models, illustrated by fake review detection
SummaryThis paper proposes MIND-X, a design approach that aligns AI explanations with users' mental models, commonsense concepts, and judgment heuristics rather than only model fidelity. Using fake review identification as a case study, a qualitative study with 10 participants produced a taxonomy of perceptual cues, which guided prompts for LLMs to generate adaptive, language-based explanations. In a quantitative evaluation with 166 participants, these explanations showed advantages over token-level transparency methods in understandability, satisfaction, trust, persuasiveness, and decision-making support.
Sidra Naveed (1), Md Shajalal (1), Gunnar Stevens (1 et al.
In briefHybridResearcher: A Research Agent That Reasons Across Structured and Unstructured Sources
SummaryThis paper introduces HybridResearcher, a 4B-parameter agent designed to combine evidence from structured sources, such as databases, and unstructured sources, such as the web. The authors build training data by starting from single-source tasks and synthesizing the missing source around a bridge entity, keeping only samples where neither source alone is sufficient. The model is trained with supervised fine-tuning followed by reinforcement learning. On the HybridDeepResearch benchmark, it reports gains of 30.0 pass@8 points over its base model Qwen3.5-4B and 8.3 points over the much larger Qwen3.5-397B-A17B.
In briefPredicting Lipreading Differences from Eye and Pupil Signals During Silent Speech
SummaryThe authors recorded eye-tracking and pupil data from 24 adults completing a silent four-alternative lipreading task, creating the LiPS dataset. A four-branch neural network with cross-modal attention classified participants as stronger or weaker lipreaders, reaching 0.680 accuracy. Pupil-linked and frame-level gaze signals were more informative than the time spent looking at the mouth.
In briefDistribution-based regularization for reward fine-tuning of diffusion and flow models
SummaryThis paper addresses mode collapse and distributional drift that occur when pretrained diffusion and flow-matching models are aggressively fine-tuned with rewards. The authors propose DFR, which adds a distribution-dependent, per-sample reward acting directly on the generated distribution, generalizing KL-regularized objectives to convex functionals. This yields additive per-sample rewards based on Fréchet distance, maximum mean discrepancy, and DreamSim diversity. In text-to-image fine-tuning experiments, DFR matches the peak reward of KL-tuned baselines while increasing diversity by up to 29% at equal or lower drift.
In briefLastWords: Boring Problems Revisited, a paper on old NLP problems that look simple but are hard to solve with neural methods
SummaryThe paper revisits text normalization, specifically number name reading, a seemingly trivial problem that is hard to solve with purely neural methods. The authors report that at least one recent frontier model has practically solved this problem. They argue that open issues remain for such limited-domain, resource-constrained problems, which future approaches might solve without massive model resources.
In briefField Tests of a Semi-Autonomous Vegetable Weed Sprayer at Three Speeds
SummaryThis study tests a semi-autonomous tractor sprayer that uses machine vision to target weeds in radish fields. Field trials at three operating speeds show that weed detection was strong at the lowest speed, but weed hit rate fell and crop contact rose as speed increased. The system did not meet the predefined commercial-viability thresholds, and the authors identify specific engineering improvements needed.
In briefScalable, adaptive distribution regression using random features
SummaryThis paper studies distribution regression, which learns mappings from probability distributions to response variables. It proposes RaDR, built on random feature approximations, and gives a finite-sample analysis under two-stage sampling showing its excess-risk bound keeps the order of exact kernel methods given enough features. It also introduces A-RaDR, which allocates random features adaptively according to downstream objectives. Experiments on synthetic benchmarks and 41 real-world datasets show RaDR saves substantial computation while staying competitive, and A-RaDR further improves accuracy.
In briefQuery-dependent reading for quantized inner-product estimation, with a minimax separation result
SummaryThe paper studies inner-product estimation under separate storage and read budgets, in the static bit-probe model. It shows that query-independent readers gain nothing from extra storage, with a matching minimax error bound of 1/(4^R-1). It then builds a two-batch, query-dependent protocol whose error decreases as storage grows. The protocol's error is within a factor of π/2 of the paper's lower bound in the leading constant.
In briefMeasuring Text Data Diversity with Embeddings: A Taxonomy and Benchmark
SummaryThis paper surveys 21 embedding-based measures of data diversity and groups them into four families. It introduces EmbDivBench, a benchmark that evaluates these measures on variety, balance and disparity using simulated and natural text data. The authors report that Mean Pairwise Distance, the most common measure, is outperformed by several alternatives, with Graph Entropy performing best overall, and they offer recommendations for choosing a measure.
In briefDo We Need the Largest LLMs for Every Clinical Decision?
SummaryLarge clinical language models often need multi-GPU hardware, which limits local deployment. This paper introduces CERTA, a decision-level statistical certificate that decides when a quantized model can replace its BF16 reference model and sends uncertified decisions to that reference. Across eleven instruction-tuned models and 2,500 clinical-verification decisions, 8-bit quantization served nearly all held-out decisions, and certain 4-bit configurations lowered deployment cost, with substantial variation between models.
In briefGoal Images as Control Targets: Spatially Grounded Goal-Image Synthesis for Robot Manipulation
SummaryThis paper studies the goal images that steer robot manipulation policies. The authors find that top proprietary image editors on general editing leaderboards produce worse control targets than their smaller, finetuned open models, with a larger gap on multi-step tasks. They report that downstream success depends mainly on manipulation adherence rather than photorealism. They also propose SGE-Goal, a framework that predicts edit-region masks before synthesizing goal images, and release code and datasets.
WhyArgues goal generators should be chosen by downstream control utility; spatially grounded method evaluated thoroughly on CALVIN and SIMPLER-Bridge with strong long-horizon gains.
Riddhi Chatterjee, Ritish Shrirao, Kushagra Mehrotra et al.
In briefSample-efficient reinforcement learning with general utilities using neural dual critics
SummaryThis paper studies reinforcement learning with general utilities, where the objective is a nonlinear function of policy-induced occupancy measures. The authors propose a natural actor-critic method in which a neural critic learns the policy-gradient pseudo-reward directly over state-action pairs, with multilevel Monte Carlo controlling bias. Under stated conditions, the method achieves Õ(T^{-1/2}) global convergence above approximation errors and Õ(ε^{-2}) sampling complexity for a fixed critic representation.
In briefRAFT: Filtering Redundant Inputs to Speed Up Temporal Graph Neural Network Training
SummaryThis paper addresses the high cost of training temporal graph neural networks, where many events and their dependencies slow parallel processing. The authors propose RAFT, which tracks changes in node memories during training and uses them to filter out redundant prediction targets and historical events. Across three TGNN models and five datasets, RAFT reduces training inputs by 35.23% on average and cuts recorded training time by up to 52.56%, with negligible changes to model precision.
In briefBellman Language Models: a method that solves LLM agents' recurring decisions once and reuses them.
SummaryThe paper proposes "Discretize, Then Decide," which turns workflow logs or agent run data into discrete states, estimates outcome probabilities and move costs, and solves the resulting Markov decision process. The solved policy runs as a controller around the LLM, which is called only for states not yet seen. On τ²-bench telecom tasks, the controller alone solves 79% of tasks, and an open model's success rate rises from 29% to 96% with 93% fewer model calls. The approach helps little when the hard part is writing content, such as resolving SymPy issues.
WhySolving recurring agent decisions as an MDP and reusing them is novel; comparisons on τ²-bench and InventoryBench give solid evidence, though scope is limited to bounded workflows.
In briefDual Task Vector Merging: Using Emotion and Speech Recognition Knowledge to Detect Alzheimer's Disease from Limited Speech Data
SummaryThe paper proposes Dual Task Vector Merging, which adds weight differences from wav2vec 2.0 models fine-tuned on emotion and ASR tasks into each layer of a frozen speech encoder for Alzheimer's detection. Only layer-wise merging coefficients and a linear classification head are trained, updating under 0.1% of the parameters needed for full fine-tuning. On ADReSS and ADReSSo, the method outperforms frozen probes, LoRA, and full fine-tuning in the main binary settings, and performs best when combined with LoRA. It also transfers to three-class HC/MCI/dementia classification on PROCESS and gives reliable gains with 20%, 40%, and 80% of the training data.
In briefSpriteStage: Long-Horizon Animation Generation via Progressive Scene Programming and Replay
SummaryThe paper presents SpriteStage, a framework for generating long animations by building them scene by scene. Each scene is generated from an accepted prior state, executed on a shared stage, and revised using runtime and visual feedback. The results are then mapped to production assets and engine capabilities and replayed into the final video. Three complete animations of about 9.9 to 14.3 minutes are used to demonstrate the approach.
In briefFlowMaker: Automatically Choosing Cost-Efficient Generative AI Workflow Deployments From Tasks
SummaryThis paper presents FlowMaker, a system that lets users submit tasks while it automatically selects the workflow structure, models, and deployment configuration. It uses a constraint-guided, load-aware pruning algorithm to make fast, on-the-fly decisions under dynamic cluster conditions. The authors report that, compared with state-of-the-art baselines including Murakkab, FlowMaker improves goodput by up to 10.14× and reduces cost by up to 2.77×.
Dimitrios Liakopoulos, Prasoon Sinha, Neeraja J. Yadwadkar
In briefFlexible 3D MRI Super-Resolution Using Gaussian Prototypes at Any Scale
SummaryThis paper presents GSMR3D, a model that reconstructs high-resolution 3D MRI volumes at arbitrary isotropic or anisotropic resolutions, using one checkpoint per MRI contrast to help reduce scan time. It combines a learned dictionary of anisotropic 3D Gaussian prototypes, scale-conditioned frequency encoding, and a measurement-consistency projection that keeps outputs faithful to the acquired low-resolution data. Across geometries, it achieves better perceptual and distributional quality scores than geometry-specific models, while preserving information for tasks such as glioma segmentation, and reconstructs a volume in about 6 seconds.
Aditya Rastogi, Ramon Leiser, Prateek Singh et al.
In briefLLM-Based Evaluation Planner for Heterogeneous Optimization (REP-HO)
SummaryThis paper addresses heterogeneous optimization, where an optimizer must decide which candidate solutions to evaluate and which objectives or constraints to assess under varying costs and fidelities. The authors propose REP-HO, which replaces the rule-based mixed-fidelity selection step with a large language model planner, enhanced by an adaptive nearest-neighbor summary. Across five heterogeneous regimes, REP-HO is competitive with the hand-designed MFE-NSGA-III and shows adaptive allocation behavior.
In briefLearning to Write Retrieval Programs: Combining Search Tools with Reinforcement Learning
SummaryThis paper introduces Retrieval Program Learning (RPL), which lets a model combine search primitives into executable programs with local state and control flow. It also proposes program mutation, which creates variants of weaker programs sampled during reinforcement learning to improve exploration. RPL improves open-domain question answering across model sizes, with especially large gains on multi-hop tasks, and a 4B model trained this way outperforms all 8B baselines. On corpus-level reasoning it achieves the highest F1 scores while reducing context consumption and retrieval turns.
In briefMisleadQA: Testing how LLMs fail when context misleads them
SummaryThis paper introduces MisleadQA, a multiple-choice dataset with an abstention option for testing LLMs given misleading context. Out-of-scope context is true but about a different entity, time, or location, while in-scope context is false but supports a wrong answer. Six open-weight models switched to the wrong answer in 29.3% of cases with out-of-scope context alone and 75.4% with false in-scope context alone. Adding correct context removed most of the out-of-scope effect but still misled models in 21.0% of cases with false in-scope context.
In briefAgentSoV: Virtual sensing for LLM multi-agent systems using variation streams
SummaryAgentSoV is a framework that models large language model multi-agent systems as stochastic dynamic systems through a Stream of Variation. It learns how variations propagate across stages with a state-space model, without re-executing the system. The framework supports failure attribution, quantifying how intermediate inputs influence outcomes, and placing virtual sensors under a monitoring budget. Experiments across multiple architectures are reported to show a unified approach.
In briefAn impossibility result for ETF target geometry and an equinorm fix for imbalanced analytic continual learning
SummaryThis paper studies classifier heads in analytic continual learning, asking whether combining the Simplex ETF targets from Progressive Neural Collapse with exact closed-form updates helps. The authors show the fusion reduces to plain one-hot ACIL, and that ETFs are not nested across class counts, causing catastrophic forgetting. They propose equinorm post-processing of the fitted weights, which improves tail-class and average accuracy under long-tailed imbalance from 5:1 to 100:1, at a cost on near-balanced streams.
In briefPCHMR: A benchmark for privacy-constrained human mesh recovery in real-world settings
SummaryThis paper studies how to evaluate and improve human mesh recovery (HMR) models when raw user images cannot be centrally collected. The authors benchmark state-of-the-art HMR models trained with federated learning and secure aggregation across varied client scales and data heterogeneity. They also propose DePoser, a local annotation and fine-tuning pipeline using depth foundation models, which improves personalization accuracy on in-the-wild data measured against DePoser-generated pseudo-ground truth.
In briefA statistical view of Adam, Shampoo and Muon optimizers
SummaryThis paper models stochastic gradients as normally distributed and whitens them using their covariance to design optimization methods. It proposes an online estimator for gradient mean and variance, and uses diagonal and Kronecker-factored covariance approximations to derive a method close to Adam and a new MN method, a Shampoo variant. The framework is also applied to gradient directions to obtain MN-Muon, a preconditioned version of Muon. In GPT-2-type experiments, the authors report that MND outperforms Distributed Shampoo and KL-Shampoo in stability, validation loss and time, and MND-Muon improves on Muon in most paired runs, while NorMuon performs comparably and slightly better in some settings.
Tetiana Parshakova, Robert M. Gower, Paulo Orenstein
In briefPattern-driven evaluation of LLM code optimization, separating optimizations compilers cannot already do
SummaryExisting evaluations mostly measure end-to-end speedup, which does not show whether a model actually performed the intended optimization. The paper introduces PatternBench, a benchmark covering seven transformation classes that remain beneficial after aggressive compiler optimization, with an AST-based check of whether the target optimization appears in the generated code. Experiments with frontier, open-weight, and fine-tuned models find that runtime speedup alone is incomplete, that faithfulness is strongly associated with realized speedup, and that fine-tuning improves both.
In briefMeasuring Agreement Among AI Agents: An Evaluation Framework
SummaryAs AI agents connect through networks and handle tasks needing negotiation, this paper proposes a framework that treats consensus as public positions converging toward a shared region, using accumulated communication history as context. It tracks position trajectories, group dispersion, and information in public messages. Two case studies cover staffing decisions from distributed facts and reconciling duplicate information with a cancelled contract. Across 1,872 completed runs, sharing explanations was associated with more accurate group decisions.
In briefVTF900: A 900-Subject Benchmark for Thermal-to-Visible Face Recognition
SummaryThis paper introduces VTF900, a large benchmark with 900 subjects and paired visible and thermal face images, including variations in expression, pose, and occlusion. It supports verification and closed-set identification at large gallery scale. After training on ARL-VTF, the authors compare convolutional, transformer, and domain-adaptive models. At N=900, ViT-B reaches 97.68% AUC and 51.10% TAR@1%FAR on frontal images, while DPIT-ResNet50 is more robust off-pose (94.68% AUC, 36.08% TAR@1%FAR). The results show that high AUC at small scale does not ensure reliable large-scale performance, and pose variation markedly worsens degradation.
Jacob Abaare, Bianca Ianetta, Cedric Nimpa Fondje et al.
In briefRecurrent RBM as an unsupervised encoder for biomechanical movement sequences
SummaryThe paper adapts a Recurrent Temporal Restricted Boltzmann Machine into an unsupervised encoder that maps windows of movement data to fixed-length embeddings, and integrates it into the SIT-FUSE segmentation pipeline. It evaluates the embeddings on simulated musculoskeletal reaching and two real gait datasets, finding that clustering recovers movement structure tied to target progress without labels. On gait data the embeddings beat PCA and autoencoder baselines under a nonlinear probe, but removing the recurrent connection showed no measurable loss on reaching, and simple summary statistics classified walking speed at least as well as the embeddings.
In briefBlock-wise long-range forecasting with shared parameters versus sequential compute (Compose the Horizon, paper)
SummaryThis paper proposes Recursive Functional Decomposition (RFD), which generates the future horizon by applying one parameter-shared operator three times, each time producing a contiguous block fed back into the context for the next. Experiments suggest the evolving forecast state itself carries the useful signal, and widening the operator helps more than reusing it. In a frozen long-context evaluation RFD has lower test error in 23 of 24 cells, but the comparison uses different context lengths, and its sequential channel-wise execution is slower and more memory-intensive at high channel counts.
In briefTraining Code Agents to Ask Questions Without Leaking Private Code
SummaryLocal coding agents can consult a stronger cloud model, but these consultations often expose private identifiers, file paths, and source code. The paper proposes utility-verified abstraction (UVA), which trains the agent to describe private entities by role and type while keeping the technical details an expert needs. The trained agent consults more often and solves 37.6% of repository-level issues, compared with 25.3% for the same model working alone. Its questions disclose 3.7-fold fewer private entities and 1.4-fold less verbatim code than the untrained agent's questions.
In briefPair-Aware Optimization: Combining Exploration and Refinement to Jailbreak LLM Agents
SummaryThis paper studies how adversarial inputs can make large language model agents carry out attacker-specified malicious behaviors, a practice called jailbreaking. The authors compare pure zeroth-order, pure first-order, and two orderings of combining them, then propose a ZO-FO framework that explores prefixes with zeroth-order search and refines suffixes with first-order optimization. Experiments show that hybrid optimization outperforms single-optimizer methods and that the optimization order substantially affects results.
In briefEvent rate as a feature: curriculum FiLM training for event-based GNNs
SummaryThis paper argues that event-camera event rates, which vary with scene motion and other settings, carry useful information that standard pipelines discard through normalization or rate-invariance training. The authors condition a PointNet++ backbone on each recording's log event rate via FiLM layers, using a two-stage curriculum of natural rates followed by rate-augmented fine-tuning. The method reaches 67.2% on N-Caltech101 and 82.5% on DVS128 Gesture with about 0.2M extra parameters. A matched-capacity control indicates the gain comes from rate conditioning rather than added model capacity.
In briefExpert-guided rubrics as dense rewards for reinforcement learning in hard-to-verify domains
SummaryThis paper studies training LLM retrieval agents with reinforcement learning in non-verifiable domains such as financial services, where query quality needs graded expert judgment rather than binary rewards. The authors convert a small, fixed amount of subject-matter expert judgment into a dense reward and combine it with retrieval metrics into a hybrid signal compatible with GRPO. According to the paper, this yields up to 6% higher Recall@3 and 6.7% higher MRR@3 than sparse-reward baselines.
Ashish Bansal, Hossam Amer, Keerat Kaur Guliani et al.
In briefEfficiency measured in English: tokeniser vocabulary limits small language models on Indic scripts
SummaryThe paper argues that small language models are usually judged by a single memory figure measured in English, which does not carry over to Indic scripts. Testing TinyLlama-1.1B on a 131k-sentence corpus across eleven Indic scripts, the authors find its tokeniser covers almost none of several scripts, falls back to raw bytes for some, and reports a 0.0% UNK rate anyway. They conclude the binding constraint is vocabulary rather than parameter count, leaving a Kannada user with about 91 words of context versus 1,528 for English in a 2048-token window.
