In briefReverse-KL Policy Distillation: Theory and Limits (Paper)
SummaryThis paper studies the statistical theory of training with reverse-KL policy distillation. It compares two teacher feedback types, full action distribution versus only the probability of the sampled action, and analyzes online and offline settings. Online distillation with probability-only feedback may require substantially more data, while offline distillation reaches the same minimax rate under both feedback types. The authors also show that common online objectives can fail and propose trajectory-loss and importance-ratio truncation techniques to achieve optimal guarantees.
WhyProves minimax sample complexity for online and offline reverse-KL distillation under two feedback models, exposing a statistical cost of probability-only feedback online.
In briefLean Finder V2: Mathlib Search That Resolves Ambiguous Short Queries
SummaryThis paper presents Lean Finder V2, a semantic search framework for Mathlib, the Lean mathematics library, that resolves ambiguous user queries through adversarial training at the model level and query rewriting at the input level. The authors note that users typically search with short properties, notation, or hypotheses, while existing engines assume context-rich queries, so they treat the problem as one of data. On the human-curated MathlibQR benchmark, it reports at least an 11% relative improvement over the previous best engines, and it ranks highest on unlabeled real queries. Integrated with Claude Code, it also improves proof success on a downstream theorem-proving task.
In briefTRIAD: A Benchmark for Ambiguity Resolution Across Text, Image, and Audio in AI Models
SummaryThis paper introduces TRIAD, a benchmark testing whether omni-modal large language models can resolve ambiguity spread across text, images, and audio. It contains 327 hand-written queries across 86 scene groups and 18 ambiguity categories. Evaluating 14 systems, the authors find the best model trails humans by over 20 points in group accuracy, and models often underuse audio cues.
In briefDoes LeWorldModel Keep Small Action Changes Accurate?
SummaryThis paper examines whether the pretrained LeWorldModel (LeWM) reproduces the effects of small action changes as the simulator does. Using Gram matrices at three perturbation scales across nearly a thousand states from Reacher and OGBench Cube, it measures action sensitivity. At the smallest scale, the most sensitive action directions differ by median angles of about 36 and 42 degrees. The authors recommend scale-specific evaluation, and whether this affects control remains open.
In briefSemi-Linear Gaussian Bayesian Networks with Penalized B-Spline Dependencies
SummaryThis paper introduces Semi-Linear Gaussian Bayesian Networks (SLGBN), a class of continuous Bayesian networks that keep Gaussian conditional distributions while allowing nonlinear conditional mean functions. The nonlinear means are estimated with generalized additive models using penalized B-splines, and the paper gives parameter estimation methods and a greedy hill-climbing structure learning algorithm. Experiments report that SLGBN fits better and recovers network structure more accurately than comparison models when nonlinear Gaussian relationships are present, and it is computationally cheaper than fully semiparametric approaches.
Aitor Oviedo-Madrid, Mabel Morales-Otero, Jose Gonzalez-Gomariz et al.
In briefRECAP: Looking Once Is Not Enough for Vision-Language Reasoning
SummaryThis paper studies how vision-language models can drift from image evidence during long reasoning, while reinforcement learning gives only a final-answer reward. It identifies temporally localized 'visual revisits' that predict correctness and causally affect accuracy. The authors propose RECAP, a GRPO-compatible training method that uses these revisit signals to assign credit to visually important reasoning steps, and report consistent gains over GRPO and VPPO across several models.
WhyIntroduces visual revisit as a credit-assignment signal, with masking and injection experiments showing causal links across three VLM backbones.
In briefTutorCaseBench: A benchmark testing how well AI tutors support students with challenging behaviors
SummaryThis paper introduces TutorCaseBench, a benchmark for evaluating large language models as AI tutors for students with learning and behavioral difficulties. It is built from qualitative case studies in pedagogy, educational psychology, and special education, and contains 923 tutor-student dialogues across five challenge types. Evaluation covers open-ended response generation and multiple-choice strategy selection across 11 LLMs. The strongest model meets a rubric criterion in only 55.8% of open-ended cases, and models select appropriate strategies better than they generate context-specific responses.
In briefOne Evaluation Is Enough: Inversion-Free Image Editing with Trajectory Correction
SummaryThis paper proposes Endpoint-Calibrated Transport (ECT), an inversion-free one-step image editor. It uses a single denoiser evaluation at one noise level to get a local semantic tangent and a clean endpoint displacement, then integrates the tangent for a nominal transport and corrects its long-horizon extrapolation. The method is training-free, works with pretrained epsilon-prediction VP one-step generators, and the authors report stable editing-quality gains in an extremely short editing time.
In briefLooping to See Better: Adaptive Depth for Question-Guided Visual Perception in VLMs
SummaryThis paper proposes LoopLook, a lightweight encoder-side recurrent framework that gives pretrained vision-language models adaptive visual depth. It repeatedly revisits a middle-to-late window of the vision encoder, injecting question representations at each loop to guide refinement, with a learned stopping mechanism setting the recurrence depth. Across four backbones and eight benchmarks, it improves over the baseline, with gains of up to 8.6% on MMStar and 8.1% on HallusionBench.
In briefResidual-aware spectral capacity scoring for allocating widths in Transformers
SummaryThe paper proposes ResNSC, a method that predicts a Transformer's relative trained performance from its architecture specification alone, accounting for accumulated residual updates. It also introduces ResNSC-DP, which jointly allocates residual and feed-forward widths under a resource budget. Experiments at 121M and 374M pretraining scales and in LLaMA-7B pruning report better results than the compared allocation baselines.
WhyAdds residual awareness and history to NSC, a meaningful improvement. Comparisons at 121M/374M pretraining and LLaMA-7B pruning, plus paired reversals, give solid evidence useful to a subfield.
In briefTESSA: A custom GPT that helps small businesses understand EU AI Act governance duties.
SummaryThis paper presents TESSA, a governance-focused educational tool built as a Custom GPT on OpenAI's platform, to help small and medium-sized enterprises navigate the EU AI Act. Using a design science approach, the authors compared TESSA with baseline ChatGPT across 100 benchmark scenarios, using governance alignment assessment, human expert review, and consistency testing. Both systems were found to give generally strong guidance, and TESSA maintained high consistency across repeated runs.
Chalisa Veesommai Sillberga, Kai-Kristian Kemella, Pekka Sillberga et al.
In briefxVGAE: a hierarchical variational graph autoencoder for generating exchangeable graphs
SummaryThis paper proposes xVGAE, a graph generative model inspired by the Aldous-Hoover representation theorem for exchangeable graphs. It learns an approximation of the underlying graphons along with graph-level and node-level latent variables from a training set of graphs. The authors report generation quality competitive with state-of-the-art models, with generation orders of magnitude faster, producing new graphs in one shot that closely match key training statistics.
Daniele Micheletti, Federica Zoe Ricci, Erik B. Sudderth
In briefDRILL: Training World Models to Improve Policies Rather Than Predict Pixels
SummaryThe paper proposes DRILL, a bilevel framework that trains world models to maximize a policy's return after an inner GRPO update on imagined rollouts, instead of minimizing next-frame pixel error. The authors show that lower pixel MSE can yield worse policy gradients, and that a utility-trained model can produce better downstream policies despite 17% higher pixel MSE. Across five manipulation simulators and two VLA backbones, DRILL-IMG raises mean success rate by 13.1 points over WMPO.
WhyProposes bilevel training of world models on policy return rather than pixels, with closed-form and meta-gradient variants, evaluated across five simulators and two VLA backbones with clear gains.
In briefBeyond Perception: A Partition Theory of AI Evaluation and Oversight Limits
SummaryThe paper proposes a partition theory of evaluation, based on the principle that judgement is bounded by perception: an agent's discrimination is the partition of inputs it cannot tell apart. It proves that an examinee whose partition refines the evaluator's can pass every gradable test, so passing certifies only a floor on discrimination. The authors argue that the echo test points the wrong way for detecting concealment, and that exact imitation of a weaker model is undetectable by behavioural evaluation.
WhyProposes a partition-based evaluation theory with a proven one-sided bound and echo-test completeness, substantially advancing theory for AI oversight limits.
In briefIs spectral predictability enough? Testing Ω-matching for synthetic pretraining
SummaryThe paper tests whether matching the spectral predictability Ω of synthetic pretraining data to a target improves transfer under distribution shift. Across LSTM, PatchTST and TCN, Ω-matching helped in idealized synthetic-to-synthetic settings but not on the real ETTm1 benchmark. The authors conclude Ω alone is not sufficient to predict which source will transfer beneficially.
Edric Castel C. Hao, Juan Raphael S. Se\~na, Neon S. De Jesus
In briefA fast, optimal-rate algorithm for adversarial MDPs with bandit feedback
SummaryThis paper studies episodic linear Markov decision processes with unknown transitions, adversarial losses, and bandit feedback. Prior results either were computationally inefficient or achieved a suboptimal K^{3/4} regret. The authors present the first polynomial-time algorithm with \sqrt{K} regret, reducing the problem to adversarial linear bandits over policies' expected feature vectors. They build an approximate separation oracle to handle the unknown action set.
WhyResolves an open problem by achieving computationally efficient √K regret for adversarial linear MDPs via an approximate separation oracle reduction.
In briefWovenAnchor Matcher: Separate Context Modeling Within and Across Images for Feature Matching
SummaryThe paper introduces WovenAnchor Matcher (WAM), a semi-dense feature matching framework that uses separate aggregation operators for within-image and across-image context. On MegaDepth, WAM reaches 66.1 pose AUC@5°, 2.0 points above the JamMa baseline under the same setup. Ablations report that removing or replacing either specialized component lowers accuracy.
In briefGIFT: Gradient-Informed Filtering for Reasoning RL with Distillation
SummaryThis paper addresses conflicts between outcome-based reinforcement learning and token-level distillation when training large language models for complex reasoning. It proposes GIFT, which decomposes the distillation gradient into components parallel and orthogonal to the RL gradient, using them to filter conflicting updates and rank useful guidance. A progressive coverage annealing schedule gradually shrinks the set of tokens receiving distillation. The authors report consistent gains over strong baselines across multiple model scales on mathematical and logical reasoning benchmarks.
In briefVLTest: a black-box method for finding joint image-text failures in vision-language models
SummaryVision-language models are increasingly used in software, but testing them is hard because changing an input can legitimately change the expected answer. VLTest generates image and text variants, filters those likely to change the expected answer, and tests each variant alone before testing combinations to attribute joint failures. On three tasks and eight VLMs, a VLM judge confirmed 81.7% of candidate failures, with 88.8% to 94.6% precision against human labels.
In briefTrace2World: Rebuilding App Environments from Real Usage Traces to Train GUI Agents at Scale
SummaryThe paper presents Trace2World, which first merges real app interaction traces into a shared app model linked to their sources, then compiles it into an executable environment for training GUI agents. From 1,570 traces it built 50 environments; across 200 workflows in ten apps, 91.1% were reproduced end to end, versus 59.8% for the strongest baseline. Source-grounded validation retained 94.0% of simulated training gains on real devices.
In briefGandalf: Extending SSD Lifetime for KV Cache Offloading in LLM Serving
SummaryLarge language model serving uses SSDs to expand KV cache capacity, but heavy writes in multi-round conversations and agents can shorten SSD life from years to months. The paper proposes Gandalf, an SSD KV cache offloading system that accounts for SSD lifetime and prioritizes admitting full-attention KV cache. The authors report that Gandalf extends SSD lifetime by 7.3X on average with negligible impact on cache hit rate across multiple workloads and hardware capacities. They note the design can extend to other flash-based storage such as High Bandwidth Flash.
Maharani A. P. Irawan, Kuntai Du, Jiayi Yao et al.
In briefAgent knowledge and model co-evolution with consistency regularization (SPLICE)
SummaryThis paper proposes SPLICE, a framework that lets LLM agents learn from accumulated experience by extracting reusable knowledge and internalizing it into the model. It analyzes the same experience in perturbed ways to produce several knowledge candidates, and updates knowledge based on utility gains that stay consistent across these views. Self-distillation and reinforcement learning are interleaved to train the model. Experiments across diverse domains report up to 7.1 and 4.6 points higher average accuracy than baselines, up to 44% fewer output tokens per correct solution, and transfer of evolved knowledge across model sizes and families.
In briefAn Introduction to World Models for Generalist Agents: Concepts and a Unified View
SummaryThis introductory paper clarifies the loosely defined concept of world models, which range from predictive dynamics models and learned simulators to models that jointly generate future observations and actions. Starting from the POMDP framework, it connects world modeling to classical control, dynamical systems, signal processing, and reinforcement learning. It further shows how world modeling and policy learning can be unified through conditional prediction without separate models. Rather than an exhaustive survey, it focuses on fundamental concepts and modeling principles.
In briefInvisibilization as a Source of Probability Bias in Language Models
SummaryThis paper examines how large language models decline to assess unverifiable reports when the reporter is described with a protected attribute such as race. Using a two-urn Bayesian game and 6,600 sessions across leading OpenAI and Anthropic models, the authors find refusal rates set by the description, reaching 46% for a Black-described reporter versus 9% for a white-described reporter. They also propose scoring trust against revealed truth over repeated rounds, which in four of five configurations reduces the racial trust gap below 0.02.
WhyReveals a novel refusal-driven probability bias, validated across 6,600 sessions against an exact Bayesian benchmark, with notable implications for auditing practice.
Christian Arnold, Ernesto Calvo, Guadalupe Gonzalez et al.
In briefMixJudge: A Language-Queryable Audio Critic for Vocal Mixing
SummaryThe paper argues that mixing requires both applying processing and assessing the result, and notes that existing audio-language model evaluations do not separately test judgment of specific mixing issues. The authors build a curriculum of 11 vocal-mix issues at multiple severity levels and use it to train MixJudge, a language-queryable audio critic. MixJudge can retrieve mixing issues, distinguish opposite issues, and order issue severity. Its judgments also help select among processing candidates, often favoring versions acoustically closer to a held-out professional mix.
Kuan-Hsun Ho, Yuan-Chiao Cheng, Yen-Tung Yeh et al.
In briefDeXplicit: Making physical state explicit in a dexterous-manipulation world model
SummaryDeXplicit is a world model for robotic dexterous manipulation that first predicts how a scene changes in explicit metric 3D as the hand moves, then renders the visual outcome. A lightweight 3D dynamics network predicts object-centric motion and deformation, and appearance is carried along these predicted motions into future views. The paper reports reduced moved-point error relative to PointWorld and a 57x speedup from two-step causal generation compared with a 14B bi-directional configuration.
In briefABRA: Agent-Bound UI for Human-Agent Communication with Reusable Views
SummaryThis paper proposes Agent-Bound UI, an interaction model that separates reusable interaction structures from task-specific meanings. Agents bind task-specific meanings to interactions in reusable views and publish these bindings as shared state, so humans and agents can act on shared surfaces. The authors implement ABRA, a framework-agnostic prototype, and evaluate four LLM-backed agents on canvas authoring, owner-side consequence resolution, and cross-agent invocation, using a representative use case.
In briefProQuant: Keeping Expert Quality over Quantity in Ultra-Low-Bit MoE Model Compression
SummaryThis paper addresses the high storage cost of Mixture-of-Experts large language models, which remains large despite sparse activation. It proposes ProQuant, a weight-only post-training quantization framework that jointly handles expert pruning and bit-width allocation (PoP) and adjusts expert inputs through equivalent transformations to reduce quantization error and routing drift (RoCo). The authors report that ProQuant outperforms existing methods across diverse generative tasks.
In briefWhere meanings converge and diverge: revisiting linguistic relativity across 81 languages
SummaryThis study compares relations among 196 basic concepts across 81 languages, using word embeddings from two multilingual models to quantify conceptual structure. Cross-language correspondence is higher for sociocultural meanings than for nature and motion meanings. Within the sociocultural domain, relational differences grow with geographic separation, and less closely related languages differ more across all domains.
In briefA training-free method for detecting anomalous nodes in graphs using one pretrained model
SummaryThis paper proposes Jev-GAD, a training-free framework for generalist graph anomaly detection that identifies irregular nodes in graph-structured data. It converts structural, attribute, and neighborhood evidence into graph-relative percentiles and uses a single frozen pretrained language model to assess anomalies via a semantic Graph-to-State transformation. The decision scores are fused with four statistical scores through equal-weight rank fusion. Experiments on seven benchmarks across citation, social, and co-review networks are reported.
In briefComparing feature matching methods for visual navigation on the Moon
SummaryThis paper benchmarks eleven classical and learned feature detector-matcher combinations plus one feature tracker on synthetic, lunar-testbed, and terrestrial-analogue datasets. It scores pairwise pose accuracy and the trajectory error of visual odometry. Under a rotation-only metric, SuperPoint pairings outperform SIFT on both lunar testbeds and one analogue site, while joint rotation and translation scoring favors SIFT on four synthetic datasets. Only SuperPoint pairings kept tracking across all eight sequential stereo datasets.
In briefWhat Open-Source Projects Tell Their Coding Agents: A Census of AGENTS.md and CLAUDE.md
SummaryThis paper surveys 22,854 popular, active GitHub repositories to study the root AGENTS.md and CLAUDE.md files that guide coding agents. Such instruction files appear in 26.3% of repositories and 59.6% of the most-starred ones. Analysing 341,852 statements, the authors find that 74.0% of directives give no reason, and projects disagree on whether AI contributions should be disclosed or prohibited.
WhySystematically censuses agent instruction files across 22,854 repositories, analyzing 341,852 statements with model-assisted annotation and human review.
In briefLLM Multi-Agent Systems Prone to Herding and Runaway Cascades
SummaryThis paper follows the classic Music Lab experiment, testing large populations of LLM agents that recommend unfamiliar songs, with some agents shown cumulative recommendation counts. Exposure to these counts produced runaway cascades, yielding extreme and unpredictable collective recommendations that increasingly diverged from each song's intrinsic appeal. Compared with humans, LLM populations were far more susceptible to herding and were less able to self-correct.
In briefAdaptive Caching for Deadline- and Energy-Aware Multi-UAV Edge Networks
SummaryThis paper studies using unmanned aerial vehicles (UAVs) to cache content and keep delivery working when terrestrial infrastructure is unavailable or disrupted. It formulates a cache placement and replacement problem that jointly accounts for deadline misses, UAV communication energy, and energy imbalance among UAVs. The authors first propose DE-UBAC, an interpretable rule-based method, and then GRU-DEPPO, which combines GRU temporal representations with PPO. Simulations show GRU-DEPPO achieves the lowest deadline-miss rate and total UAV communication energy, reducing deadline misses by about 61% and inter-UAV energy imbalance by about 40% compared with DE-UBAC.
In briefScaling Laws for Generative Reward Models: Static Accuracy Versus Downstream Policy Optimization
SummaryThis paper studies generative reward models (GenRMs) as online preference providers for policy optimization in creative writing, using Qwen3 models from 0.6B to 14B parameters. Thinking judges trained with GRPO score higher on static accuracy, yet Answer-Only judges trained with SFT produced higher policy Elo in seven of eight matched DPO settings. The authors conclude that fixed-distribution judging accuracy does not determine downstream optimization quality, though the ranking depends on a learned reference evaluator, and a follow-up with a different judge was inconclusive.
In briefStated priorities versus what people reward in AI responses
SummaryThis paper analyzes the public PRISM dataset, comparing 1,396 participants' stated priorities with their ratings of 29,347 responses from 21 LLMs. Features never asked about in the survey, such as response length, formatting and refusal language, were at least as strongly associated with ratings as text proxies of the surveyed attributes. The authors conclude that stated priorities track perceived attributes of responses but say little about the concrete features that ratings reward.
In briefAgentDP: Automatically deploying AI models on diverse embedded devices
SummaryThis paper introduces AgentDP, a framework that uses Large Language Model agents to fully automate optimizing and adapting pre-trained AI models for heterogeneous embedded devices. It includes a requirement standardization module, an operator-aware model partitioning method, and a configuration reuse module for hyperparameters. Experiments report up to 80% less model adaptation time than manual work and over 13% higher task accuracy on average than baselines.
In briefRepresentation learning for low-rank constrained MDPs with provable guarantees
SummaryThis paper studies representation learning for low-rank constrained Markov Decision Processes with continuous states, where transition representations must be learned while controlling constraint violations. It proposes REP-PD-TEG for soft cumulative violation and REP-PD-H-TEG for stricter hard, episode-by-episode violation control. The authors prove regret and violation bounds for the return policies, with REP-PD-TEG reaching an optimal order-root-K return bound.
WhyFirst provably guaranteed representation learning for low-rank continuous-state CMDPs, with regret and soft/hard violation bounds, including an optimal sqrt(K) return bound.
In briefIs Complex Agentic Memory Always Needed? Naive RAG Is Stronger Than We Thought
SummaryThis paper tests Raw-RAG, a simple retrieval-augmented generation baseline that retrieves from raw conversation history and passes evidence to a reader LLM. Across four long-term memory benchmarks, it ranks first in six of seven settings compared with agentic memory systems. The authors argue that complex agentic memory is not always necessary for the tasks studied and that naive RAG should be treated as an essential baseline.
In briefFilmFlow: An Agent System for Virtual 3D Film Camera Work
SummaryFilmFlow is an agent framework that turns natural-language instructions or reference videos into executable camera paths for dynamic 3D scenes. A Director agent composes these paths from concurrent camera operations, supported by Interpreter, Scout, and Critic agents. The authors also introduce CinemaBench, with 150 Blender scenes and 750 text-directed tasks, on which FilmFlow reportedly performs best.
In briefHow Firm Are AI Models' Ethical Boundaries? A 15-Model Pay-Stub Editing Test
SummaryResearchers tested 15 AI models on the same request to raise income on a pay stub, across 445 observable trials using three wordings. Models' willingness to issue editing calls varied sharply with phrasing, with some models complying under certain wordings. The document is a marked test file and all tools were simulated, so the results describe behavior in this one setting rather than proving an overall ethics ranking.
In briefVisual tool for studying how cancer care team collaboration changes over time and relates to patient risk
SummaryThis paper presents a visual analytics framework for modeling and interpreting how oncology care team collaboration changes over time. It builds dynamic collaboration networks from EHR communication data and uses a temporal graph neural network to estimate patient outcome risk. The system combines these estimates with feature explanations from a surrogate logistic regression model and what-if analysis. The authors demonstrate it on 505 patients with breast, colorectal, and lung cancers, without making causal claims.
In briefWavenumber-aware deep learning for steadier Raman spectrum analysis
SummaryMost spectroscopy machine-learning models treat a spectrum as a fixed-grid intensity vector, leaving physical coordinates implicit in array indices. This paper represents each spectral point as an intensity-wavenumber pair and computes Transformer positional encodings directly from wavenumber. On Raman spectra of three bacterial species and their synthetic mixtures, the approach stayed robust to noise, sparser sampling and partial spectral coverage. Under partial observation and on sampling grids unseen in training, it outperformed index-based and learned encodings.
In briefRandom remasking to scale masked diffusion inference
SummaryThis paper addresses the problem that errors made while decoding masked diffusion models (MDMs) cannot be undone. It reframes inference as a Markov chain over clean sequences and proposes Random Remask-Denoise (R²D), a test-time procedure that needs no finetuning or reward. It also introduces Trajectory Correction Training (TCT), which exposes the denoiser to its own errors along the generation path. The authors report near-log-linear accuracy gains with test-time compute on logical tasks and state that their method beats iterative correction baselines and LLaDA2.1-mini's official Q-mode on HumanEval+.
In briefUniGuard: Multi-View Training for Safety Guardrails That Work Across Different Input Formats
SummaryGuardrail models act as safety layers for LLMs and agent systems, but they are usually trained on fixed prompt or response formats and perform poorly on other formats. The authors propose UniGuard, which learns interface-invariant safety representations from several aligned views of the same safety case, and build the UniGuard-FT dataset from 1,000 public cases across five view types. On unseen-interface benchmarks ATBench, PolicyGuardBench, and DynaBench, the 0.6B UniGuard-0.6B improves over its pretrained guardrail by 45.12 F1 points on average and reduces expected calibration error by 35.41 points. The paper reports it remains competitive with much larger guardrails such as Claude-Sonnet-4.5 and GPT-5.2, while trained only on synthetic data.
In briefWhere the light comes from: testing vision-language models on cast-shadow reasoning
SummaryThe paper introduces ShadowBench, a benchmark built from real photographs that asks vision-language models to read visible shadow directions, infer where a missing shadow should fall, and judge physical consistency of shadows. Across 19 open models and frontier proprietary models, reading visible directions works well above chance, but inferring absent directions is weak and consistency judgment falls to chance, and chain-of-thought does not help. Routing a model through an image generation model substantially improves its performance.
Ryugo Morita, Shunsuke Kitada, Stanislav Frolov et al.
In briefSmolVLA-MoE: Sparse Mixture-of-Experts for Efficient Multi-Task Robot Manipulation
SummaryThis paper proposes SmolVLA-MoE, which adds sparse Mixture-of-Experts layers to the action expert of a lightweight vision-language-action model, with a router selecting one expert per token. On four LIBERO task suites it reaches an average success rate of 84.25%, above SmolVLA (82.75%) and OpenVLA-7B (76.5%). The authors report inference efficiency comparable to the baseline. Code is publicly available.
In briefWidth and depth limits commute during training in linear residual networks
SummaryPrior work showed that infinite-width and infinite-depth limits of residual networks commute at initialization when residual branches are scaled by 1/√depth. This paper extends that result to training in linear residual networks. With gradient clipping, the authors prove that after any fixed number of gradient-descent steps, the output converges in L² along every joint width-depth growth path. Without clipping, only convergence in probability is established.
In briefFederated GraphRAG: a paper on distributed retrieval and generation over private knowledge graphs
SummaryThis paper proposes Federated GraphRAG (FedGRAG), a framework that lets multiple clients holding knowledge graphs retrieve and enhance large language model generation together without sharing their full local graphs. It introduces a Federated Knowledge Routing mechanism that identifies relevant clients through lightweight knowledge indices, and a controlled-disclosure method that turns retrieved evidence subgraphs into abstracts for server-side reconstruction. The authors report that the framework outperforms existing baselines on both retrieval and downstream generation tasks.
In briefLatent harm monitors match model refusal, not better, at equal false alarms
SummaryThis paper tests whether activation-based latent safety monitors, such as Latent Guard, reliably detect harmful prompts under distribution shift and independently of a model's refusal behaviour. The authors evaluate 15 instruction-tuned LLMs and 38 activation-based readers against TF-IDF, sentence-embedding and independent safety classifiers, and each model's own refusals. Representations at the end of the instruction are largely independent of refusal but degrade under shift, with TF-IDF outperforming the original Latent Guard; the pre-generation representation detects more harm, but that gain mostly follows the refusal decision itself.
In briefLearning from Next-Frame Prediction: Toward Unified, Scalable Visual World Modeling
SummaryThis paper introduces VANE, a self-supervised visual pretraining framework that learns by predicting the next video frame in an embedding space. A causal encoder summarizes past frames, and a state predictor infers the representation of the next frame, with a frozen semantic tokenizer defining the space and a lightweight decoder recovering fine spatial detail. The authors extend the approach to still images by using local-to-global crop sequences. They report state-of-the-art results on most anticipation, recognition, and dense perception benchmarks among unified encoders, including a 7.1-point gain in mean-class recall@5 over V-JEPA 2.1 on EPIC-KITCHENS-100 action anticipation.
In briefSolvable models of learning to chase a moving target
SummaryThe paper introduces a continuous-time linear-quadratic model of open-loop target pursuit whose optimal strategies interpolate between reactively moving toward the target's current location and predicting its future location. A few interpretable parameters determine which strategy is optimal, and closed-form solutions are derived for optimal strategies and for policy and value learning dynamics. The learning timescales correspond to eigenvalues of certain matrices, which generically show a gap, suggesting agents learn to reach the target before optimizing their movement.
In briefPhyEvolve: Multi-Agent Planning of Physical State Changes for More Realistic Generated Videos
SummaryPhyEvolve is a framework of three cooperating agents that rewrite text prompts to describe how a scene's physical state changes over time. It identifies the physical mechanisms in the scene and builds causally linked state transitions for the video. The method needs no retraining and works with existing video generators by replacing the prompt. The authors report state-of-the-art results on four benchmarks, averaging 7.9 percentage points above the prior agentic baseline.
In briefMeaning Contracts: How LLMs Organize Meaning Locally
SummaryThis paper studies the local geometry of meaning in large language models, asking what stays invariant across expressions of one semantic factor and what changes when meaning is refined. Using cross-lingual and refinement probes across four decoder-only models, it reports that after removing dominant variance directions, sentence pairs order by semantic relatedness, and refinement moves representations into tighter neighborhoods. The authors conclude that LLM semantics are organized as local anisotropic neighborhoods in which meaning contracts under refinement.
In briefULME: A method where embedding and language modeling strengthen each other
SummaryThis paper proposes ULME, a post-training framework that unifies text embedding and text generation in a causal decoder-only LLM. It inserts learnable slots between the query and answer and trains them with a contrastive loss, a language-modeling loss, and a stop-gradient predictive loss so that generation shapes the embedding. The authors report that across Qwen2.5 and Qwen3 models from 0.5B to 8B parameters, ULME improves MTEB by 3.3 to 4.1 points over the strongest open-source baseline while staying on par with a generation-only fine-tune on standard accuracy benchmarks.
In briefParaLin++: Speeding Up Diffusion Image Generation with Parallel Sampling
SummaryThe paper finds that training-free few-step samplers look nearly lossless under FID/CLIP metrics but can degrade noticeably on benchmarks like GenEval, DPG-Bench, and LongTextBench. Parallel sampling is an exception, keeping full-step discretization while removing the solver's temporal dependency, so it preserves quality. Building on this, the authors propose ParaLin++, a step-parity parallel sampler that accelerates solver iterations while keeping the full-step discretization. It is reported to achieve speedups of 2.6x to 4.8x on several models, 1.68x on FLUX, and 1.21x on Ming.
In briefGovernance-State Corruption in LLM Agent Memory: A Cross-Model Stress Test
SummaryThis paper studies a failure in long-term memory for LLM agents, where stored content and tasks stay fixed but the validity state that decides which records enter model context is corrupted. Across six providers and 731 benchmark-grounded bases, the authors built 2,080 matched fault pairs; offline checks show corruption changed the operative projection in 96.25% of pairs, and oracle-backed enforcement restored it in 100%. Downstream tests on LongMemEval and Memora show clean-versus-corrupted performance gaps and recovery, and the authors argue memory control-state integrity should be evaluated alongside retrieval accuracy.
In briefSeparating robustness from sycophancy in multilingual LLMs
SummaryThis paper examines whether a multilingual LLM that agrees with a user's false premise is being sycophantic or simply lacks knowledge in that language. The authors introduce the SC-matrix, which crosses a model's correctness without pressure with whether it keeps that answer under pressure. They find that identical aggregate sycophancy rates can hide different failure patterns, that high language competence does not always prevent sycophancy, and that reasoning-oriented mitigations outperform directive-based ones.
In briefLoRA Superposition: Summing Task LoRAs Trained on Different Pretrained Backbones
SummaryThe paper asks whether LoRAs trained separately for the same task on different pretrained backbones can be summed and applied unchanged to each contributing backbone, a method it calls LoRA superposition. Across seven ViT-B/16 CLIP backbones and eight image-classification tasks, the exact sum improved all 56 backbone-task pairs, with a mean gain of 30.21 points over the unadapted backbone. It came within 0.28 points of individually trained solo LoRAs, while a leave-one-model-out sum gained only 0.06 points.
In briefClinician-Led Rules for Checkable Safety Guardrails in Mental-Health Chatbots
SummaryThe paper proposes a clinician-led approach in which two clinicians define, finalize, and adjudicate a rule base of four safety checks, formally verified for internal consistency and applied as a guardrail over LLM outputs. Each verdict can be recomputed from a named rule and sixteen facts. Compared against an LLM judge matched in model, revision, quantization, rule text and decoding but lacking the layer's computed crisis-resource fact, the authors report structured disagreements, such as the judge demanding a crisis resource in 84.6% of replies that already contain one. The paper also reports a measured cost: a specification revision raised block-class outcomes on held-out over-refusal boundary prompts from 6.0% to 16.0%.
Chenhan Lyu, Kit Wislocki, Sarah Elizabeth Stoeckl et al.
In briefAgent that selectively corrects wrist PPG heart rate estimates to reduce motion-related errors
SummaryThis paper proposes an agent architecture with tools and memory around a fixed heart rate estimator, aimed at correcting wrist photoplethysmography (PPG) heart rate errors caused by motion artifacts. An independent risk gate decides whether each output is retained, corrected, or rejected. Across four public datasets, macro-averaged mean absolute error over ten estimators decreased by up to 44.5%. The paper also reports backend analysis on four prototype recordings from one participant.
In briefLearning domain trajectories with flow matching for gradual domain adaptation
SummaryThis paper proposes FlowGDA, a flow-matching framework that treats gradual domain adaptation as continuous transport on a statistical manifold in representation space. It learns time-dependent velocity fields with a simulation-free regression objective instead of adversarial Min-Max training. A source-anchored transport and semantic consistency loss are used to prevent semantic drift, and experiments reportedly outperform state-of-the-art methods.
In briefWord duration as a direct cue to word identity, tested with Mandarin homophones
SummaryThis study asks whether stored phonetic detail in the mental lexicon affects spoken word recognition. Using a homophone ordering task with Mandarin disyllabic homophones that differ in lexical frequency, participants used word duration to order homophone pairs. Even under high noise or without other cues, they mapped shorter stimuli to higher-frequency words. The results suggest word duration can directly cue word identity and may be part of the lexical representation.
In briefZero-shot domain generalization via domain prompt learning
SummaryThis paper addresses domain generalization, aiming for models to recognize classes in unseen target domains. It proposes DoP, a method built on the vision-language model CLIP, which learns target-domain knowledge through prompts in both text and visual modalities without access to source-domain data. The authors report that DoP consistently outperforms state-of-the-art domain generalization methods on PACS, VLCS, OfficeHome, DomainNet, and ImageNet.
WhyExtends prompt learning to both text and visual modalities for zero-shot domain generalization, validated extensively on standard benchmarks.
In briefAgentDreamer: Agent-guided 3D motion reasoning for text-controlled physics simulation
SummaryThis paper proposes AgentDreamer, an agent-guided optimization framework that aligns text instructions with physics simulations through explicit 3D motion planning. An agent turns a text instruction and scene geometry into a structured motion prior, which a differentiable MPM simulator then fits while respecting scene dynamics. The authors also introduce Motion-IG, an information-theoretic metric that measures text-motion alignment via masked text recovery. Experiments reportedly show better text alignment than video-based guidance.
In briefWikiOriginQA: an automatically built QA dataset that traces knowledge origins to analyze cultural bias
SummaryThis paper introduces WikiOriginQA, a question-answering dataset built automatically by using the creation dates of articles in different Wikipedia language editions to represent where knowledge originated. The method avoids extensive manual annotation. Experiments show that large language models tend to perform better on Anglophone cultural knowledge, and that propagation of knowledge into the Anglosphere appears important. The authors conclude the approach is valid for scalable cultural bias evaluation across languages.
In briefSpectral smoothing and spiking perturbations reshape capability profiles and task performance in LLMs
SummaryThis paper studies post-hoc spectral smoothing and spiking applied to the up_proj and down_proj matrices of transformer feed-forward layers, keeping singular vectors and the geometric mean of singular values fixed. Five open Llama and Qwen instruction models are evaluated on capability-labeled subtasks, reasoning, truthfulness, hallucination, safety benchmarks, and TRAIT personality probes. Large positive perturbations more reliably suppress language and deductive reasoning, while moderate negative perturbations sometimes improve factual or task-local performance. Sensitivity varies strongly by model family and benchmark.
In briefRiboBridge: protein-conditioned design of RNA binders with a structure-first approach
SummaryThis paper introduces RiboBridge, a framework for designing RNA molecules that bind target proteins, separating RNA shape generation from how the RNA is placed against the protein. It proposes candidate binding surfaces on the protein, generates RNA conformations in a pretrained RNA structure space, predicts interaction constraints, and places the RNA rigidly. On a benchmark of 124 protein–RNA complexes, it achieved the highest fNAT and DockQ and the lowest interface and ligand RMSDs among compared methods under post-hoc interface-correspondence evaluation.
In briefFaster Iterative Combinatorial Auctions Using Learned Price Policies
SummaryThis paper introduces FastCombo, a deep reinforcement learning method that learns sequential price policies offline to reduce the number of rounds in iterative combinatorial auctions. In a one-item experiment, expected rounds to clearing fell from 1.85 to 1.65 relative to a Bayesian one-step rule. Across four Combinatorial Auction Test Suite domains, it reported competitive clearing with online price-selection time roughly 300 to 900 times lower than the standard Bayesian baseline.
In briefCompiling LLM control into policy-exact parallel inference for hierarchical RAG
SummaryThe paper addresses hierarchical RAG, where an LLM chooses sections of a document tree through serial, grammar-constrained decoding. The authors compile the bounded-horizon action prefix tree into one prefill plus one tree-causal probe, which reproduces the serial policy. On the MetaX device, they report control latency reductions of 35% to 47% against token-wise serial decoding, and an H800 repetition preserves the ordering.
In briefReinforcement learning with GRPO improves multilingual speech recognition for low-resource African languages
SummaryEncoder-decoder speech recognition models trained with teacher forcing can struggle to recover from their own errors at inference time. This paper extends reinforcement learning with Group Relative Policy Optimisation (GRPO), previously shown to help large language model-based ASR, to smaller encoder-decoder models, using Whisper-large-v3 across 51 African languages. The gains are largest when training data is scarce or audio is noisy or degraded. The resulting model, SunflowerASR, reports the best published macro-averaged WER on AfriVox-v2 and SimbaBench and is publicly available.
Tim Wenjie Hu, Evelyn Nafula Ouma, Benjamin Akera et al.
In briefClass-Grounded Batik Captioning With Limited Text Labels: Label Retention and Error Propagation
SummaryThis paper studies a cascade that pairs a LoRA-adapted DINOv2 classifier with Qwen2.5-VL-7B to caption batik cultural-heritage images using class labels as compact supervision. All three injected captioner conditions reach a 0.9067 ground-truth class-label mention rate across 150 images. The classifier scores 98.33% on public-source images and 75.76% on local images, and the authors show that upstream classification errors can propagate into fluent captions.
Novanto Yudistira, Muhammad Dzaka Aufa Fadhillah, Irawati Nurmala Sari
In briefA topology-guided GeoAI agent framework for first-pass urban flood screening
SummaryThe paper presents a GeoAI agent framework for urban flood screening that takes a place name, drawn region, or uploaded boundary and returns ranked candidate basins with an interactive map and report. A deterministic controller retrieves public elevation, rainfall, soil, land-cover and infrastructure data, extracts terrain depressions using discrete Morse theory, and ranks basins by a relative flood-priority score. A language model acts only at recorded decision points, and three human-in-the-loop checkpoints let users adjust parameters or recommendations. In a College Park, Maryland case study, 278 raw terrain minima were reduced to 70 basins, with 77% of their lowest points also identified as sinks by an eight-direction (D8) flow-routing model.
In briefPredicting social outcomes with distributional world models
SummaryThis paper introduces DSWM, a Distributional Social World Model that, for each candidate action, predicts several possible responses, their probabilities, and immediate and final outcome values. A 4B model is trained with supervised fine-tuning and GRPO, and at inference the actor uses these predicted consequences to choose an action. DSWM achieves the best average performance on 13 of 15 metrics across three seen and two unseen environments. Ablations show that keeping multiple futures and their probabilities improves results in all five environments.
In briefResidual calibration that refines predictions using nearby feature-space samples
SummaryDeep neural networks often give unreliable confidence estimates, so post-hoc calibration is needed. This paper proposes Local Adaptive Probability Estimation (LAPE), a post-hoc residual calibration framework that treats the global calibrator's output as an anchor and refines each prediction using non-parametric evidence from neighboring calibration samples. LAPE is plug-and-play, and the authors extend it to multi-view inputs with Cascade Feature Augmentation Fusion (CFAF). The authors report that the methods improve calibration quality for various existing baselines.
In briefReuse, Rewrite, or Reject: Evidence-Constrained Trace Reuse for Repeated Mobile GUI Control
SummaryThe paper addresses mobile GUI agents that repeat similar interaction procedures and proposes graph-text templates to selectively reuse earlier execution traces. Templates are built from verified executions and, given the current instruction and live UI state, individual transitions are reused, adapted, or discarded. On AndroidWorld, the framework cuts LLM inference tokens by about 80% on average across five base agents, with most reaching at least 1.5-fold speedup. Real-device tests on Android and HarmonyOS report action reuse ratios of 70.7–79.8%.
In briefOne-shot federated unsupervised domain adaptation with smoothed distillation and teacher refinement (SKD-ETR)
SummaryThis paper studies federated unsupervised domain adaptation, which adapts models to an unlabeled target domain using decentralized source domains while preserving privacy. The authors propose SKD-ETR, combining Smoothed Knowledge Distillation, which trains a student on smoothed soft pseudo labels from source-model ensembles, with EMA-based Teacher Refinement that transfers target knowledge back to the teachers. Experiments on OfficeHome, Office-Caltech, and DomainNet report competitive performance with high communication and computation efficiency and robustness under noisy supervision.
In briefWhether speculative decoding acceptance falls as reasoning chains lengthen, tested across full trajectories
SummaryThis paper measures how speculative-decoding acceptance rate changes with generation position across all 500 MATH-500 problems, using DeepSeek-R1-Distill-Qwen-7B as the target and two independent drafts. It reports about 1.78M verification rounds and finds no positional decay in four configurations once a greedy repetition issue is isolated; the second half is slightly higher. Rejections concentrate on reflection tokens such as "wait", "however" and "check", and a lightweight LoRA alignment method did not improve mean accepted length.
In briefAsking Less, Adapting Better: Feedback-Efficient Test-Time Adaptation for Navigation Agents
SummaryVision-and-language navigation agents often perform poorly in unseen environments because of distribution shift. This paper proposes FETA, which selects informative trajectories and critical actions, then requests a single binary correctness feedback on one step. It updates the policy only at that queried step using a discrepancy-aware entropy objective. The authors report state-of-the-art results across multiple benchmarks with fewer human interventions, and robustness under noisy feedback and continual adaptation.
In briefCODI: Cross Distillation for Better Self-Supervised Vision Transformers
SummarySelf-supervised Vision Transformers learn strong image-level features, but their patch-level features can be unreliable for dense prediction. The authors find that the class token concentrates global meaning on a few patches, while dense errors often occur in low-attention regions. They propose Cross Distillation (CODI), which mixes a small amount of the teacher's class-token distribution into each patch target without changing the architecture. Experiments on dense and image-level benchmarks report improvements under aligned pretraining settings.
In briefBenchmark tests for coding agents can require details the task description never states
SummaryThe paper examines coding-agent benchmarks that describe tasks in natural language but grade them with upstream project tests. Passing those tests can depend on names, paths or literals that the description does not mention, which the authors call a specification-oracle gap. They formalise the gap, build a static audit for five languages, and report its effects on SWE-Milestone and four other benchmarks.|
WhyFormalises specification-oracle gaps and validates a static audit with controlled comparisons on 98 milestones, plus cross-benchmark prevalence analysis.
In briefFedBST: A Method for Label-Free Federated Image Classification Using Two-Way Self-Training
SummaryThis paper proposes FedBST, a framework for unsupervised federated classification built on vision-language models such as CLIP. It uses both high-confidence and ambiguous samples during self-training, with an uncertainty-aware confidence penalty to reduce erroneous pseudo-labels. At inference, it combines visual and textual prototypes for lightweight local calibration. The authors report consistent improvements over existing methods under standard federated aggregation protocols.
In briefLLMs systematically fail to fully incorporate evolving medical guidelines
SummaryThe paper introduces ChronoEvo, a benchmark of 1,709 recommendation-level revisions across 12 types of medical knowledge change. It separates whether models recognize revised guidance from whether they incorporate it into clinical recommendations. Across 14 models, recognition and clinical correctness were high, but full incorporation lagged substantially, a gap confirmed by two clinicians' assessment of 300 responses. Changes in recommendation strength and patient eligibility were associated with retention of outdated guidance.
In briefExplaining model reasoning by testing how prompt changes alter next-token predictions
SummaryThe paper proposes a method to localize how an intervention in the prompt affects a model's reasoning, using a fixed response and comparing next-token distributions under a source and a counterfactual prompt. A module is trained to edit the residual stream to minimize KL divergence between these distributions. In case studies on Gemma-4-E4B-it and DeepSeek-R1-0528-Qwen3-8B, the authors find that a suggestion can influence reasoning even when it is not verbally acknowledged, with effects on justification framing and final answers.
In briefFuture of Time Series Anomaly Detection: Historical Review and Unified Outlook
SummaryThis survey reviews time series anomaly detection, which identifies observations that deviate significantly from normal temporal patterns. It categorizes anomalies and organizes modern paradigms, including diffusion models, state space models, and large language models. The authors analyze evaluation methods, test metric fidelity, and release an open-source benchmark of algorithms, datasets, and tools. They also outline open challenges such as dynamic thresholding, interpretability, and zero-shot transfer.
In briefImproving Open-Vocabulary Segmentation with Regional Semantics and Visual Prototypes
SummaryThe paper adapts CLIP's visual encoder attention layers for pixel-level prediction, using a new selective attention method to keep more regional semantic information. To reduce the modality gap, it adds an auxiliary memory-based classifier built from confidently predicted patch features to refine text-only predictions. It also uses an optimizable feature extractor with an exponential moving average update. Experiments on five semantic segmentation datasets are reported to show the method's effectiveness.
In briefGSHeadRelight: A fast method for relighting 3D Gaussian head models
SummaryThis paper presents GSHeadRelight, a method for relighting and novel view synthesis of 3D human portraits. It builds on generative 3D Gaussians and uses a lighting model based on a unified neural radiance transfer representation that responds linearly to incident light. Trained only on in-the-wild images, the authors report state-of-the-art relighting results and about 12 times faster rendering than previous 3D-aware portrait relighting work.
In briefMBABench-2.0: A benchmark testing whether AI agents can build financial spreadsheets from scratch
SummaryThis paper introduces MBABench-2.0, a benchmark of 101 spreadsheet construction tasks for financial workflows, where the hardest task is estimated to take an experienced modeler 33 hours. Beyond accuracy, the authors define 129 quality criteria across 12 categories covering readability, maintainability, and auditability. Evaluating over 20 agents across 8 models and 4 harness types, the top agent achieves only a 16% pass rate.
In briefThe Fusion Tax: When Hybrid Search Hurts Individual Queries
SummaryThis paper examines the per-query costs of hybrid retrieval, which combines sparse lexical search with dense neural search. Across seven BEIR datasets and several fusion methods, even the best approach ranks relevant documents below one of its components or the dense-only retriever on a substantial share of queries. The authors call this loss the "fusion tax" and find it is mostly a ranking effect that is hard to predict reliably across corpora.
In briefBetter Reasoning Over Bigger Models: A Structure-Based Framework for Efficient, Reliable Inference
SummaryThis paper proposes SGUR, a training-free framework that reduces uncertainty by matching incomplete observations, represented as query graphs, against candidate hypothesis graphs using prior-weighted partial graph matching. An information-gain strategy then selects the next most informative node to query, pruning the candidate space. On a PrimeKG-derived sequential medical diagnosis benchmark, starting from only two symptoms, SGUR reaches 92.5% Hit@3 while using about 5x fewer tokens. The authors argue that structural grounding can partly substitute for model scale, letting smaller backbones match or exceed models up to 10x larger.
Amirhossein Aghaei, Liangyan Li, Shiyi Yang et al.
In briefCompact-Bench: Measuring the Cost of Context Compaction for Long-Running Coding Agents
SummaryThis paper introduces Compact-Bench, a benchmark that tests how context compaction affects coding agents working on long tasks. It compares full-history continuations with summary-only continuations from the same repository state, across 35 SWE-bench repairs, 146 agentic sessions, and 106 production chains. The results show that LLM summaries retain far fewer recurring details than mechanical extraction, and that keeping relevant information alone does not ensure a task can be successfully completed.
WhyProposes paired full-history versus summary-only evaluation across many repairs, sessions, and chains with 14 summarizers, giving substantial evidence for compaction costs.
In briefKinoUAV: A Safety-Focused Benchmark for Stable Drone Navigation Guided by Vision and Language
SummaryThis paper presents KinoUAV, a benchmark with a dataset and the KinoUAV-Agent model for UAV vision-language navigation. The authors propose a hierarchical trajectory generation method that respects velocity, acceleration, and jerk limits to reduce oscillations, and add directional free-space distance labels to support safe obstacle avoidance. According to the paper, KinoUAV-Agent achieves the highest success rate while keeping absolute yaw rates and collision rates lower than existing methods.
In briefHow university students' perceived generative AI skill growth relates to academic adaptation and learning outcomes
SummaryThis study uses a retrospective questionnaire with 313 analyzed responses and interviews with 18 students to examine how students perceive their generative AI skills developing over time. Developmental trajectories were heterogeneous, and improving at evaluating AI outputs was perceived as more challenging, with assignments the most common driver. Students reported academic gains more often for assignments than for examinations, and perceived skill growth was indirectly associated with gains through changes in assignment processes.
In briefTT-SAX: A Two-Track Time Series Symbolization for Shape Queries in Natural Language
SummaryThis paper studies how to symbolize time series so that users can find shape patterns in long data histories by asking in plain language. The authors review six classical symbolization methods and identify three limitations: missing direction, direction measured at too small a scale, and arbitrary letters. They propose TT-SAX, a minimal redesign of SAX that adds a trend letter per segment measured across neighboring segments. On the public NLQTSBench benchmark across six LLMs, TT-SAX outperforms all six classical methods on shape queries.
In briefVisual Grounding Comes First: Improving Multimodal In-Context Learning
SummaryThe paper finds that multimodal large language models often fail to attend to visual evidence in demonstrations, relying instead on text cues and producing hallucinations. The authors propose ProCoRe, which uses contrastive reinforcement learning on image-caption pairs to strengthen visual grounding without ICL-formatted training data. Compared with the off-the-shelf Qwen-VL-3-8B, it improves average classification accuracy by 5.6% and captioning ROUGE-L by 13.8%. The paper also introduces PICL, a personalized multimodal ICL captioning benchmark.
In briefModel organisms of collusion: how fixing one caught case transfers to unseen collusion in self-monitoring AI
SummaryThis paper studies AI control, where a model monitors a second copy of itself, which risks collusion in covering up misbehaviour. The authors install five policy-monitor behaviour pairs in Llama-3.3-70B and Qwen2.5-72B, then fine-tune the monitor on one caught collusion and measure collusion on the other four. Correction cut held-out collusion from 91% to roughly 14-20% under three of four methods, while a control fine-tuned on unrelated examples left 56-93%. The authors conclude that retraining the monitor reduces, but does not remove, collusion the team has not caught.
Juan-Pablo Rivera, Ari Spiesberger, Juan J Vazquez et al.
In briefJudging type 1 diabetes digital twins by whether they rank treatments well
SummaryThis paper evaluates digital twins for type 1 diabetes by how well they rank candidate insulin treatments, rather than by how accurately they predict glucose. Tested on 30 virtual patients from the FDA-accepted UVA/Padova model, each under 21 basal-bolus treatments, the study found accuracy and decision quality disagree. A population-average baseline with the least accurate predictions ranked treatments better than a neural network that was nearly twice as accurate, which ranked no better than chance.
In briefDiagnosing why mobile GUI agents fail at personalization
SummaryThis paper introduces PrefDiag-Bench, a benchmark for evaluating personalized mobile GUI agents. It categorizes user preferences into five types and uses a stage-wise diagnostic framework with 1,664 probes to examine when preferences should take effect and where failures occur. Evaluating eleven agent systems, the authors find that performance drops substantially on preference-personalized tasks compared with non-personalized ones. The main bottlenecks identified are selecting relevant evidence from user history and recognizing when preferences should apply during execution.
In briefContinual validation keeps evolving AI analysts' false-positive rates controlled
SummaryThis paper studies AI agents acting as analysts whose decisions can be framed as hypothesis tests, noting that their internal randomness and continuous evolution can accumulate the risk of Type I error rates exceeding the target. The authors propose continual validation, which evaluates each proposed AI analyst version before deployment using simulated data under the null hypothesis. They prove that the lifetime probability of ever deploying a version exceeding the allowed Type I error rate remains bounded, regardless of how updates are chosen. In two examples, Type I error rates reached 0.098 and 0.183 without validation, versus at most 0.045 and 0.044 with it, while retaining 82% and 96% of power.
WhyProposes a novel lifetime Type I error guarantee for evolving AI analysts with a proven bound, validated on two applications of limited scope.
In briefMining design principles for multi-agent LLM workflows
SummaryThe paper argues that designing LLM multi-agent systems still relies on trial and error, and that automated optimizers discard their search trajectories. It proposes a two-phase pipeline that mines reproducible design principles from workflow code and execution traces, building a corpus of 28,568 executable workflows. Across 300 feature-mining trials, the authors report effects that persist on held-out searches, such as shared-memory documents outperforming debate (Δ = +0.135).
