In briefCan AI agents handle customer complaints? A simulation study of multi-turn complaint resolution with unequal information
SummaryThe paper introduces ServiceBench, a benchmark of 2,500 scenarios covering 184 complaint categories across five service industries, along with an Asymmetric Progressive Complaint Simulation method. It also proposes a justice-based evaluation framework. Experiments show that even capable LLM agents handle complaints poorly, often offering compensation prematurely. Targeted post-training with SFT and DPO improves these abilities, letting a smaller model match or surpass much larger ones on several outcomes.
In briefGLOSS-MT: controllable machine translation for low-resource languages using an editable dictionary
SummaryThis paper proposes GLOSS-MT, which replaces source noun phrases with Twi translations from an offline English-Twi dictionary and fine-tunes a translation model once to handle these mixed inputs. The method is applied to NLLB-200, OPUS-MT and TranslateGemma, and on an 84-document English-to-Asante Twi benchmark every fine-tuned model beats its pretrained base in chrF2. The authors also report that the tested fine-tunes reproduce most dictionary terms and adapt to swapped terms without retraining, and they release the models, data and code.
Mich-Seth Owusu, Stephen E. Moore, Akwasi Asare et al.
In brieffxBench: A benchmark for evaluating formula suggestions in spreadsheets
SummaryThis paper introduces fxBench, a manually curated benchmark of 503 formula suggestion tasks based on Sheetpedia. It breaks the problem into four capabilities: gathering context, understanding context, understanding intent, and writing formulas. The authors evaluate small and large models across different context representations, including fine-tuned smaller models, and find the best configuration produces correct suggestions 58.8% of the time.
In briefShlokaDhara: A Multi-Speaker Sanskrit Corpus for Meter-Aware Verse Recitation Speech
SummaryThis paper presents ShlokaDhara, a 43-hour multi-speaker Sanskrit speech corpus covering diverse meters, recitation styles, and melodic variations. It targets Text-to-Recite synthesis, which requires both prosodic accuracy and metrical correctness. Using the automated UTMOS predictor, XTTS-v2 was rated highest in perceptual quality among the benchmarked TTS models. The XTTS-v2 outputs were further validated through subjective ratings from 16 Sanskrit experts and syllable-level Laghu-Guru duration analysis.
Harshal Kausadikar, Onkar Kishor Susladkar, Prathmesh Santosh Choudhari et al.
In briefWhere Natural Language Ends in Agents: Executable Skill Composition via MetaSkill
SummaryThe paper proposes MetaSkill, which formalizes coordination between agent Skills so that invocation dependencies and predecessor-output bindings run through a directed acyclic graph runtime, while task judgments stay in natural language. Across 20 multi-Skill artifact tasks and two LLMs, blinded judgments show favorable preference point estimates against several baselines. The authors also report that audited execution traces satisfy all planned nodes and declared precedence constraints.
In briefPV-Decoding: Progressive Verification Decoding for Faster Diffusion Language Models
SummaryThis paper proposes PV-Decoding, a speculative decoding method for diffusion language models that uses confidence thresholds to sort tokens and verifies speculative tokens in progressively stronger stages. Compared with confidence-based decoding, it reduces sequential model forward passes by 8% to 37% across GSM8K, MBPP, and HumanEval, with the largest accuracy drop at 1.3%. On Dream-7B-Instruct with GSM8K, the single- and dual-batch variants improve throughput by 7.0% to 13.6%, while the multi-batch variant is slower under the current runtime.
In briefTensor network models for tracing how training data shapes model behavior
SummaryThe paper introduces Tensor Network Datamodels (TNDMs), which use low-rank tensor networks to capture interactions among training data units that linear datamodels cannot represent. On a controlled reasoning task, TNDMs outperform all baselines while observing only half as many training subsets. In factual-association fine-tuning, a single TNDM-CP raises held-out R² from 0.579 to 0.668 over the linear datamodel.
Farzaneh Heidari, Md Rifat Arefin, Louis Jaburi et al.
In briefInhibitory message passing that keeps rare evidence visible and cuts token use in multi-agent reasoning
SummaryThe paper proposes inhibitory message passing, a training-free communication method for multi-agent systems. It cancels redundant evidence separately for each evidence sentence and receiver, so rare but decisive evidence stays visible, and it excludes senders' reasoning and answers. Compared with standard message passing, accuracy in peer groups rose by 39 to 77 percentage points across three benchmarks, while tokens per run fell by 10% to 46%.
In briefDiffuPlan: An Editable Planning State for Diffusion Language Model Agents
SummaryThe paper introduces DiffuPlan, a diffusion search agent that keeps a single editable planning state and rewrites only the affected steps when new evidence arrives, instead of appending text as autoregressive agents do. It also proposes Expansion-Partition Marginalization (EPM) for variable-length revisions and adapts diffusion trajectory RL to multi-turn search. Across four multi-hop QA benchmarks, DiffuPlan reaches 50.24 average F1, and editable planning state improves the diffusion agent by 2.44 F1 versus 0.47 for its autoregressive source model.
In briefWhen Repetition Loops Get Stuck: Sampling Limits and a Gap in DRY
SummaryThis paper studies whether a language model can escape once it has begun repeating the same text verbatim. Because the context is deterministic inside a loop, the authors compute loop survival under different samplers exactly from one forward pass. They identify a blind spot in the widely used DRY sampler and propose DRY-C, which avoids looping to the length limit in their tests.
WhyAnalyzes escape conditions for repetition loops and exposes a DRY blind spot, with DRY-C validated across 11 models and 720 generations.
In briefMultilingual brain decoding shows a shared cortical semantic geometry across languages
SummaryThe study develops a multilingual brain decoding framework and analyses brain recordings from 227 participants speaking Chinese, English and French. Cortical semantic representations showed consistent semantic geometry relationships across languages and individuals. This cross-language consistency increased from sensorimotor to higher-order association cortex and closely aligned with the principal cortical functional gradient.
WhyShows a shared cortical semantic geometry across languages aligned with the principal functional gradient, using 227 participants across three languages.
In briefSQuIREL: A Method for Measuring Entity Richness in Wikidata by Context
SummaryWikidata is a large, varied knowledge graph, and existing "richness" quality metrics mostly operate at the ontology level and miss the structural depth of individual entities. This paper introduces SQuIREL, which measures entity richness across six dimensions: semantic specificity, qualifier density, internal connectivity, reference provenance, external network connectivity, and linguistic accessibility. Case studies of oil paintings (Brazilian vs. French) and writers grouped by citizenship and gender show how the metric reveals representational inequalities that may indicate systemic bias. For example, 48% of Brazilian paintings are low-richness stubs, compared with 0% in the French group.
Michelle Nwachukwu, Albert Mero\~no-Pe\~nuela, and Martim Brand\~ao
In briefPhysics-informed neural operator on the sphere for multi-scale weather downscaling
SummaryThis paper proposes a Spherical Physics-informed Neural Operator for downscaling global meteorological fields to high-resolution regional forecasts. It uses a Spherical Laplacian Decomposition to split atmospheric fields into hierarchical frequency components, and a localized spherical integral operator with geometry-aware attention to evaluate them at arbitrary locations. Differentiable constraints are added during learning to enforce dynamical consistency. The authors report higher accuracy and zero-shot generalization across several meteorological variables and unseen queries.
In briefAtomSD: Procedural Atom-Guided Self-Distillation for LLM Agents
SummaryThis paper introduces Procedural Atoms, fine-grained units capturing Intent, When, Do, and Avoid guidance that can be applied independently. It proposes AtomSD, a framework where an LLM agent continually extracts these atoms from its own successful trajectories and distills them into its policy. Experiments show AtomSD outperforms strong reward-based and distillation baselines, by up to 9.31 points on average.
In briefSuperClip Pyramid for Finding Moments in Videos
SummaryThis paper presents SCVTG, a method for locating moments in videos based on user queries, covering moment retrieval and highlight detection. It groups adjacent clips into semantically coherent SuperClips and builds a multi-scale pyramid from them. A temporally adaptive kernel aggregates these variable-length segments, and relative saliency losses help the model learn importance between clips. The authors report the best performance across five video temporal grounding benchmarks.
In briefNative Probability Outputs for Clinical Triage: Calibration and Decision Accuracy Evaluation
SummaryThis paper evaluates Jev, a decision model that outputs probabilities over predefined actions, for patient triage. Using CARE-Bench, it compares Jev with four general-purpose language models and three input-independent baselines across 925 decision points from 439 public cases. Jev has the lowest observed probability error but ranks fourth of five by macro-F1, and it misses clarification needs in the four-action task.
In briefChecking Whether Human-Written Reasoning References Hold Up
SummaryThe paper proposes a reference-centred method to check whether each step in human-written reasoning chains is entailed by its cited premises, with full-premise rechecking and failure attribution. Evaluated on P-FOLIO using an LLM to translate steps into first-order logic and the Z3 solver for entailment, 202 of 386 derivations (52.3%) passed all criteria. Most initial failures became entailed when complete premises were used, suggesting disagreement may stem from the reference itself.
In briefGeometry-aware interpolation for recovering degraded images
SummaryThis paper addresses recovering missing visual information from incomplete or degraded images. It proposes GAIN, a variational autoencoder framework that uses the local tangent structure of the data manifold to estimate pixel intensities, with entropy-based regularisation guiding spatial transformation parameters. Experiments on image completion and classification benchmarks report better predictive accuracy and lower computational cost than existing methods.
In briefCoLAT: a contrastive adversarial training method for graph neural networks resistant to label poisoning
SummaryGraph neural networks are effective on relational data, but a few corrupted training labels can spread errors across the graph through message passing. The paper proposes CoLAT, a contrastive adversarial training framework that uses contrastive embeddings to detect label-structure inconsistencies and select high-risk nodes for adversarial label flipping during training. The authors report that CoLAT outperforms existing defenses on multiple benchmarks under various poisoning intensities and scales efficiently to large graphs.
Manshika Charvi Bissessur, Melis Ilayda Bal, Michael Muehlebach
In briefMM-LeJEPA: Correlated Gaussian Regularization for Multimodal Joint-Embedding Predictive Architectures
SummaryThis paper extends LeJEPA's task assumptions to multimodal learning, modeling tasks as weighted combinations of components reading independent and shared information across two modalities. Under explicit assumptions, it derives a correlated Gaussian embedding target and adapts Sketched Isotropic Gaussian Regularization to it, yielding a two-term objective. The method adds one scalar hyperparameter, the target cross-modal correlation. On public audio-visual classification benchmarks, results are competitive with representative open-source methods.
In briefMESS: A method that synthesizes multi-exposure sequences to train image enhancement models that generalize better.
SummaryThis paper addresses the limited generalization of image enhancement models under adverse lighting, such as low-light, exposure correction, and backlit scenes, which stems from a lack of large, diverse paired training data. The authors derive multi-exposure sequences from high-bit RAW images and train a diffusion-based generator to synthesize such sequences from single 8-bit RGB images. Using this approach, they build a 500K-pair dataset and train a lightweight network that shows better generalization than existing models across several enhancement tasks.
In briefFrom Views to Worlds: Active 3D Exploration for Vision-Language Models
SummaryThe paper notes that vision-language models perceive passively from single views, which limits their 3D spatial reasoning. The authors propose View-to-World (V2W), which reconstructs an explicit 3D world with geometry, camera poses, and semantics from multi-view images, and lets models actively gather spatial evidence through visual and language interfaces over multiple turns. V2W works training-free with existing models and can be further improved with reinforcement learning, reportedly improving results on spatial mental modeling benchmarks.
In briefDataPraxis: Scaling Analytics Tasks via Verified Workspace Evolution
SummaryThis paper introduces DataPraxis, a framework that synthesizes data analytics tasks by progressively enriching workspaces with validated findings and artifacts. The authors build DataPraxis-5K, a dataset of 5,000 verified trajectories, and use it to train DataPraxis-27B and DataPraxis-35B-A3B. The reported DataPraxis-27B results are 55.11 on DABStep and 51.9 on JobBench, matching or surpassing frontier models such as GLM-5.2.
In briefDo Conversational Agents Need Explicit Memory? A Controlled Comparison with Raw-History Retrieval
SummaryThis paper compares six state-of-the-art explicit memory systems for conversational agents with a baseline that retrieves directly from raw conversation history, using LongMemEval-S and BEAM-100K under original and controlled retrieval settings. Explicit memory showed only localized gains for certain question types and tight run-time context budgets, with no consistent advantage over strong raw-history RAG baselines, despite construction costs reaching hours per conversation. The authors find retrieval is a major bottleneck when useful information is preserved, suggesting better access to existing information may matter at least as much as building richer explicit memory.
In briefKnowledge Diversion: A New Pre-training Method for Decomposable Image Generation Models
SummaryPre-trained models often mix knowledge useful for many tasks with task-specific knowledge, making selective reuse difficult. This paper proposes Knowledge Diversion (KDiv), which factorizes initial weights into rank-one units organized into shared learngenes and condition-specific tailors during pre-training. Condition-aware routing keeps knowledge shared across conditions in learngenes and directs condition-dependent variations to tailors. Experiments on class-, text-, and image-conditioned generation report that KDiv matches or exceeds dense baselines at comparable pre-training cost and achieves the best average downstream performance with trainable parameters comparable to low-rank adaptation methods.
In briefWhat RL Brings to Reasoning: Decisions the Base Model Would Not Make
SummaryThe paper studies what reinforcement learning with verifiable rewards (RLVR) adds to a base model's math reasoning. Using generation-time interventions on Qwen3-4B-Base over 120 AIME problems, the authors find that borrowing the RL model's token at roughly 4% of positions, where greedy choices differ, recovers 90.5% of the pass@1 gain. A single opening token recovers about 35% of the gain, but later replacements contribute more, while retaining much of the base model's diversity.
In briefFrom Prediction to Action: A World Model for Population Health in Digital Medicine
SummaryThe paper argues that digital medicine still focuses mainly on task-specific prediction and proposes a population health world model (PHWM) that couples health states, individual trajectories, population dynamics, interventions, constraints, and feedback. The authors contend that building such a model requires three minimal capabilities: representation learning under action-dependent observation, multi-scale intervention intelligence, and validation with explicit decision boundaries. They also highlight the value of social, contextual, and public-health sensing data for action-oriented use.
In briefCan AI agents write law? Testing AI's legal relevance and creativity
SummaryThe study has one language model play several characters that debate in "constitutional assemblies" to draft a hypothetical constitution for a community of AI agents. The assemblies converge on a shared set of core provisions and produce varied proposals, including some AI-specific clauses with no human-law equivalent. The authors introduce a measure of recombination to identify the most novel legal material. They conclude that AI extends familiar legal ideas to new situations but does not work through how those situations would change, so it still depends on human legal reasoning.
In briefLightweight Machine Learning for Classifying Disaster-Related Social Media Posts
SummaryThis paper examines whether lightweight lexical machine-learning methods can classify disaster-related social media posts into humanitarian information labels effectively and efficiently. Using the CrisisBench v1.0 English humanitarian task with 87,455 records, the authors compare TF-IDF representations and lightweight classifiers, with Macro-F1 as the main selection metric. The selected balanced LinearSVC reached 0.8282 accuracy, 0.6578 Macro-F1 and 0.8315 Weighted-F1 on the test split, and the study also analyses per-label behaviour, error patterns and a tropical-cyclone specialist model.
In briefGood Judgment, Bad Advice: Decision-Model Controllers Add Cost Without Better Success
SummaryThis study tests whether Jev 1.13, a typed decision model consulted at fixed decision points, can make a Claude Sonnet 5.5 terminal agent cheaper without hurting results. On 21 held-out Terminal-Bench 2.1 tasks (315 trials), Jev raised cost per attempt by 19% at an identical 73.0% pass rate, and was more expensive than a placebo controller. Jev's completion probabilities were poorly calibrated, often asking the agent to verify work that was already correct.
In briefCaMeWM: A Causal Medical World Model for Tumor Evolution and Treatment Planning
SummaryThe paper introduces CaMeWM, a causal medical world model that simulates post-TACE tumor evolution to support comparing treatment regimens. It admits queried regimens only for comparable, supported histories, generates lesion-local follow-up images with a 3D diffusion transition, and maps the generated outcomes to restricted mean survival time (RMST) with a response-Cox model. Across 79 paired out-of-fold episodes, it reports 0.618 SSIM for follow-up CT transition and a 0.615 C-index for generated-world survival, and factual-only AIPW estimates a 9.60-week RMST gain for the frozen policy (95% CI 0.61 to 18.13).
In briefEdgeKernelBench: A Benchmark and Layered Agent for LLM-Generated Kernels on Edge Hardware
SummaryThis paper introduces EdgeKernelBench, which pairs a unified IR deployment framework with a hierarchical benchmark for LLM-generated kernels across diverse edge backends. With a single attempt, the best model reached 64.5% Correct@1 but only 25.5% Fast@1.0, showing correctness does not reliably yield speedups. The authors also propose LEKA, a layered agent that achieves 43.5% Fast@1.0 on the agent subset, versus 13.0% for direct generation and 26.1% for Claude Code.
In briefModeling Uncertainty Propagation in 3D Morphable Models
SummaryThis paper studies how to map the parameter-space variance of a 3D Morphable Model (3DMM) to per-vertex variance. It proposes Closed-form Local Uncertainty Estimation (CLUE), a closed-form approximation derived from the law of total variance that is computationally inexpensive and accurate. When used in training for in-the-wild human mesh recovery, CLUE reportedly outperforms probabilistic and deterministic baselines and achieves the highest uncertainty-to-error Spearman correlation.
WhyClosed-form approximation maps 3DMM parameter variance to per-vertex variance, outperforming baselines in mesh recovery experiments.
Megh Shukla, Edo Collins, Prashanth Chandran et al.
In briefFrequency-domain membership inference does not beat a confidence threshold
SummaryThis paper tests a widely repeated claim that memorized samples depend on fine detail, so degrading input resolution should expose membership through how fast predictions collapse. The authors apply a family of zero-phase low-pass filters to each image, track the target's confidence decay, and score it in a dozen ways. On two federated ResNet-18 targets, no profile score beats a single confidence threshold, and the authors attribute the failure to the spread growing faster than the mean gap.
In briefSpectral-aware proposal enhancement for diffusion-based multispectral drone object tracking
SummaryThis paper addresses multispectral UAV multi-object tracking with oriented bounding boxes, where small targets, cluttered backgrounds, and unstable associations make tracking hard. The proposed DSATrack framework adds a lightweight spectral encoder that feeds multispectral information into the RPN branch to produce better initial proposals, and recalibrates confidence using local multispectral saliency contrast. On the MMOT benchmark, DSATrack outperforms the adapted DiffusionTrack baseline by 4.6 HOTA, 9.0 MOTA, and 7.7 IDF1 points.
In briefSelm: A Self-Evolving Library Management Framework for Multi-Skill Agent Optimization
SummaryLLM agents often use skills to acquire task-specific knowledge, but existing methods lack systematic management of evolved skills, making reuse across domains difficult. This paper presents SELM, which treats skill evolution as library-level management, jointly handling skill retrieval, updating, and lifecycle using confidence and utility signals. It was evaluated on four agent benchmarks covering diverse tasks and execution environments, where it consistently outperformed baselines in most settings.
In briefMuse Glimmer 30B: A Local Model for Autonomous AI Agents
SummaryThis paper introduces Muse Glimmer, an open-source 30-billion-parameter language model with a perception encoder, built for autonomous agent tasks on consumer hardware. It accepts text and image inputs, produces text outputs, and is quantized to about 4-bit precision for single-GPU inference. The authors report state-of-the-art results for its size class on agentic benchmarks such as tool use and multi-step planning. The model is released under the Apache 2.0 license.
Arash Ebrahim Abtahizadeh, Muna Aghamelu, Adil Ahmed et al.
In briefLearning hierarchical meaning structures in text-linked graphs by combining LLMs and GNNs
SummaryThis paper proposes SHiFT, which combines large language models and graph neural networks within an EM framework to learn hierarchical semantic structures in text-attributed graphs. In the E-step, an LLM induces an explicit semantic tree, updating it only when anomalies are detected to reduce cost. In the M-step, the GNN encoder is aligned with this hierarchy via clustering, topology, and semantic concepts. The authors report state-of-the-art results and human-readable taxonomies.
SummaryThis paper presents a diagnostic framework for visual dynamics models, separating visual accuracy, geometric validity, and agreement with applied forces. The authors evaluate several models on planar and spatial assemblies and find that a model can improve image error while still failing to produce valid geometry. They also report persistent actuator-force residuals even after geometry is restored, helping distinguish coordinate, numerical, and force-response errors.
In briefSafety Kernel: a shared set of neurons for defending LLMs against cross-lingual attacks
SummaryThis paper finds a small set of safety-related neurons shared across languages in large language models, called the safety kernel (SK). The authors identify language-specific safety neurons using contrastive activation scores on parallel safe and unsafe prompts, then intersect them to obtain the SK and test it with activation interventions. They propose SK-Masked LoRA, which restricts LoRA adaptation to the feed-forward output dimensions tied to the SK. The authors report that this substantially lowers attack success rates across multiple languages, including some not used in SK discovery, with generally limited effects on general capabilities.
In briefHow Much Do Activation Monitors Gain from Original Model Computations?
SummaryThis paper studies how much activation-based monitors for AI systems depend on access to the original computations that generated a model's behavior. Across up to four models and two counterfactual monitoring tasks, monitors recover much of their performance without the original trace. Monitors trained on native computations keep a mild but consistent advantage, outperforming the strongest controls by up to 0.014 AUROC.
In briefHSPR: Recovering Software-Agent Benchmark Scores from a Small Task Subset
SummaryRunning a software agent on a full benchmark is costly, and small task subsets can give inaccurate scores. This paper proposes Hierarchical Score Posterior Recovery (HSPR), which combines response factors fitted from source agents with uncertain group success rates, and uses posterior-guided task selection to update the estimate. At a 10% task budget on four SWE-bench-family task banks, HSPR achieved the lowest mean absolute error on three banks and the lowest maximum absolute error on all four among compared methods.
In briefIntriAdapt: Using Camera Intrinsics to Adapt Geometry Models at Test Time
SummaryThis paper tackles the ill-posed problem of recovering metric 3D geometry from images, which is ambiguous due to perspective projection. IntriAdapt is a test-time adaptation method that uses known camera intrinsics as a geometric check, minimizing the residual between the predicted point cloud's viewing rays and those defined by the intrinsics, without depth sensors or external priors. It optimizes a lightweight LoRA adapter on a frozen geometry foundation model and extends to video, where adapting on a short sequence generalizes to other sequences from the same camera. Reported experiments on ten single-image and seven sequence benchmarks show gains over unadapted and oracle-intrinsics baselines.
In briefSCOPE-VSR: A Broad-Coverage Training Dataset for Video Super-Resolution
SummaryThis paper addresses semantic imbalance and limited diversity in training data for open-domain video super-resolution. It introduces SCOPE, a curation pipeline combining task-oriented video selection with hierarchical semantic rebalancing, used to build the roughly 27k-video SCOPE-VSR training set and the SCOPE-test set. The authors also propose GraCo-VSR, a one-step video restoration framework with reliability-aware design and explicit supervision to reduce degradation-induced distortions. The authors report strong reconstruction fidelity and perceptual quality, and the code is publicly available.
In briefSelf-ensemble vision transformer with residual tokens for small-data image classification
SummaryThis paper studies how Vision Transformers (ViT) perform on small image datasets, where data is often limited, such as in medical imaging. The authors propose ReSEViT, a multi-stage residual ViT that combines self-ensemble learning across stages with Dynamic Learner Selection to adaptively weight each stage's importance. Residual token embeddings, which resemble ResNet blocks, are added to model local structure. The paper reports that this design reduces redundancy, prediction error, and variance while keeping the model size efficient.
Stephen Ekaputra Limantoro, Yan-Kai Wang, Ching-Chun Huang et al.
In briefTCMScience: An Autonomous Research Agent for Traditional Chinese Medicine
SummaryThis paper presents TCMScience, an autonomous agent for knowledge discovery in traditional Chinese medicine research. It separates language-model proposals from execution and release authority, using a trusted kernel, a scientific compiler, and claim-support contracts to check findings against provenance and claim scope. In a network-pharmacology analysis of Gegen Qinlian Tang, switching to an assayed-protein background reduced significant pathways from 66 to zero. In a governance ablation, all 121 constructed erroneous outputs were blocked when both layers were active, while all 10 compliant controls were retained.
In briefSeverity-channel conditioning: input-level study for smartphone image deblurring
SummaryThis paper studies whether a blur-severity descriptor, partly derived from the paired sharp target, helps adapt motion-deblurring models to smartphone imagery on the iPhoneBlur benchmark. Under a shared update budget, several conditioning methods are compared against an unconditioned spectral-loss control. Only input concatenation of the severity map improves overall peak signal-to-noise ratio, from 31.23 to 31.30 dB, and the authors note this is an oracle diagnostic using target-derived information at inference, not evidence of a reliable deployment advantage.
Abdullah Al Shafi, Kazi Saeed Alam, Sumaiya Rahim Suma et al.
In briefAdamPM: A memory-saving variant of Adam for training large language models
SummaryThis paper proposes AdamPM (Adam with Pooled Moments), a simple variant of the Adam optimizer that groups adjacent moment entries along rows or columns and represents each group with one shared mean. The authors find that pooling only the second moment is the effective design, reducing its memory to 1/s of the original for group size s. Experiments on language model pretraining, fine-tuning and multimodal learning show that model performance and training speed are maintained while optimizer-state memory is substantially reduced.
In briefCADS: Measuring How Much LLM Proposers Add to Sequential Biological Screens
SummaryThe paper introduces Certified Agentic Discovery Screens (CADS), a framework that controls error rates while measuring how much a large language model proposer adds beyond an identical loop without an LLM. In DepMap-based simulations, the LLM-free loop alone passes the feedback gate, and raw hit counts overstate LLM contributions. On the initial 20 seeds, giving Qwen2.5-7B tested-candidate history raised its net contribution from +0.35 to +2.90, but this effect did not persist in a 50-seed matched replication.
In briefTraining with ciphers to improve in-context term learning for rare-language translation
SummaryThe paper examines whether translation models fine-tuned with in-context examples actually learn the translations of rare terms from those examples, defining this as in-context term learning. It proposes RareCipherFT, which replaces rare terms with character-shuffled ciphers so they can only be recovered by copying from context, and anneals the ciphered share of training data to preserve translation quality. Across seven translation directions involving English and five rare languages, and four models from 1B to 7B parameters, the method improves term recall over standard in-context fine-tuning without lowering COMET scores.
In briefScene-Adaptive VLA: efficient self-driving via dynamic layer routing
SummaryThis paper proposes Scene-Adaptive VLA, a framework for vision-language-action models in autonomous driving that adjusts which parameters are active based on real-time scene complexity. Complexity is estimated from frame-to-frame scene changes and decision deviations, and a budget-conditioned routing mechanism selects layer combinations, with the whole system optimized end-to-end. On the Bench2Drive benchmark, applied to a state-of-the-art VLA model, the approach reduces inference-time computation by 53.5% with negligible loss in driving and language performance.
In briefDeliMF: Early ICU delirium prediction fusing structured time series with clinical notes
SummaryThis paper presents DeliMF, a model that predicts delirium in intensive care unit patients within the first day of admission. It encodes structured vital signs, lab values and medications with a temporal Transformer, and radiology reports with a clinical text encoder, then fuses them via gated cross-attention. On an incident-only MIMIC-IV cohort, it significantly outperforms baselines for patients with clinical text and does not lose performance for those without.
In briefConsistency check for judging whether digital twins of patients can credibly predict untried treatments
SummaryThis paper addresses human digital twins that predict outcomes of treatment alternatives not observable in the same patient, noting that ordinary predictive accuracy cannot establish counterfactual credibility. The authors propose the Concordance Plausibility Check (CPC), a claim-centered framework combining real-world evidence, qualified simulation, and clinical knowledge. CPC matches evidence to the causal claim, audits contradictions and shared dependencies, and reports Pass, Fail, or a reason-coded Indeterminate result.
Saeed Farzi, Christopher Kipps, John McBeth et al.
In briefPersonalizing speech recognition for pathological voices: accuracy, efficiency and utterance length
SummaryThis paper studies how to personalize Whisper, a speech recognition model, for pathological speech using LoRA adaptation when only one speaker's limited data is available. It compares how adapter placement interacts with model size, training efficiency and utterance duration. The results show that scaling the model reduces errors but does not eliminate speaker mismatch. Full-attention LoRA gives the most reliable accuracy and best time-accuracy trade-off, with medium-length recordings benefiting most while short and long utterances remain challenging.
Giuseppe Giaimis, Sergei Katkov, Muhammad Azfar Yaqub et al.
In briefNeural network representations follow distinct geometric trajectories during training
SummaryThis paper studies how the geometry of neural network representations changes throughout training rather than only at convergence. A large-scale empirical study identifies three recurring trajectories: persistent high-rank, collapse-and-recovery, and persistent low-rank. It also names Destructive Representation Collapse (DRC), where representations collapse to rank-one and fail to recover. The authors report that collapse-and-recovery is consistently linked to successful optimization in controlled settings, but on real-world models the link depends on configuration.
In briefAutoSwarm: A Framework for Automating Collaborative AI Agent Teams
SummaryThis paper presents AutoSwarm, a multi-agent framework in which large language model agents solve problems collaboratively. It defines each agent's responsibilities and acceptance conditions, and coordinates execution through a shared state, dependencies, and supervisory relationships, adapting actions to review feedback. The authors report experiments on long-horizon programming, interactive planning, and financial analysis where it outperforms single-agent and existing multi-agent baselines. Source code is publicly available.
In briefMaximizing Market Maker Revenue Under Inventory Limits With Regret Minimization
SummaryThis paper studies how a market maker can maximize revenue over T rounds when buyer and seller valuations are unknown, while keeping inventory balanced. With only trade success or failure feedback, the proposed algorithm's regret and inventory bounds scale with T to the power of 2/3. With full order book data for failed trades, these bounds improve to the square root of T. A lower bound shows that tighter per-round inventory limits necessarily increase the minimum achievable regret.
Davide Maran, Marcello Restelli, Francesco Emanuele Stradi
In briefControllable multi-object insertion to augment object detection datasets
SummaryThis paper proposes SceneP2F, a two-stage trainable framework that inserts objects into images at requested categories and counts to augment object detection datasets. ScenePlan allocates a labeled query per instance, predicts scene-compatible object centers, and estimates box sizes, while SceneAlign-Fill renders the planned objects while limiting unintended changes to source content. Experiments on UAVDT, VisDrone, and BDD100K reportedly show it consistently outperforms representative general-purpose editors and object-insertion methods.
In briefFrom Stories to Features: Scaling Concept Vectors for AI Alignment and Interpretability
SummaryThis paper studies feature vectors, directions in a language model's activation space that match human-interpretable concepts such as honesty or deception. Instead of recovering the whole feature space, the authors focus on AI safety and alignment concepts, creating synthetic concept and antagonist story pairs and using the difference in their mean activations as a feature vector. The approach scales to 1,036 concepts and 2 million stories, which are publicly released. The vectors generalize across six models to detect concepts in chat dialogues and web text.
WhyUses activation differences from synthetic story pairs as concept vectors, scaling to 1,036 alignment concepts validated across six models for generalization and steering.
Dmitry Scherbakov, Oleg Chernikov, Victoria Khamash et al.
In briefCodec-compatible learned compression for 3D Gaussian Splatting scenes
SummaryThis paper proposes CCT-GS, a framework that compresses 3D Gaussian Splatting data by mapping spherical harmonics coefficients into compact maps suited to standard video codecs. It first learns a generalized transform across scenes, then adds LoRA-based scene adaptation to recover scene-specific appearance with lightweight side information. On MPEG GSC sequences, it reportedly beats representative video-based baselines, with an average BD-rate reduction of 51.30% over the video-based anchor.
In briefCARVS: A congestion-aware algorithm for online multi-agent pickup-and-delivery with humans in warehouses
SummaryThis paper introduces the Multi-Agent Pickup-and-Delivery with Human Collaboration (MAPDwHC) problem, which extends standard multi-agent pickup-and-delivery by adding human unpredictability, temporary path blockages from robot-human interaction, and real-time requirements. The authors propose the CARVS algorithm, which improves task assignment and path planning through task swapping, validation checks, and random initialization. According to simulation results, CARVS significantly reduces path-planning timeouts in large-scale scenarios where traditional MAPD algorithms fail.
In briefLabel-free few-shot rock thin-section classification via cross-polarization pairing
SummaryThe paper studies recognizing unseen rock types from thin-section images without any identified training slides. It proposes cross-polarization pairing (XPP), which treats each slide as a pseudo-class and pairs plane-polarized and cross-polarized views of the same field so matching relies on grain shape and texture rather than colour. On a 108-rock-type Nanjing University collection, XPP is the best label-free method on average over three random splits and matches a supervised model trained on 78 identified rock types.
In briefTheory-guided early stopping for training examples, based on second-order loss dynamics
SummaryThis paper provides a theoretical justification for the second-order loss-difference criterion used in instance-dependent early stopping, which skips backpropagation for examples whose losses have stabilized. It derives a noise-calibrated mastery threshold and a conditional stability bound, and the proposed TG-IES method sets its threshold and patience from training-run statistics. On CIFAR-10 and CIFAR-100 convolutional settings, it saves 53-58% of backpropagation instances with accuracy within 0.1-0.9 points of full training. It also transfers to ViT-Tiny with smaller savings and a modest accuracy cost.
Ahad Bin Islam Shoeb, Jamal Uddin Tanvin, Kamrul Hasan et al.
In briefEGIP: An Evidence-Grounded Interaction Protocol for Clinical Agents
SummaryThis paper proposes EGIP, a protocol that controls how clinical agents gather evidence through patient dialogue and examinations. It enforces role-specific information limits, shows available diagnostic actions without revealing their results, retrieves results from stored clinical records, and organizes the collected evidence for the final decision. Across two agent frameworks, five datasets, and six language models, EGIP improved diagnostic accuracy over native baselines. A clinician review found that the share of automatically scored correct responses judged evidence-supported rose from 10.20% to 36.55%.
In briefFaster finite-sum bilevel optimization using only gradient oracles.
SummaryThis paper studies nonconvex, strongly convex bilevel optimization with finite-sum objectives, using only first-order gradient information. It proposes SICBO, which combines a penalty approximation with variance-reduced corrections for lower-level variables. The method achieves square-root dependence on the number of components and keeps the best known condition-number exponent for smooth problems.
In briefWrite once, query many: evolving compact reusable memory for long videos
SummaryLong videos often exceed the context limits of multimodal large language models, and agentic search repeats costly video processing for every query. This paper proposes Video-EMem, an evolvable video memory built from time-anchored observations, where a Writer encodes each video once and a Reader retrieves evidence only from that memory. Across three long-video QA benchmarks, it improves Qwen3-VL-30B's average accuracy by 2.6% over direct input of 2,048 frames while reducing storage by about 99.86% on average.
In briefA Multi-Step Inertial ADMM Method for Federated Learning
SummaryThis paper proposes MiFedADMM, a multi-step inertial ADMM algorithm for federated learning, which aims to reduce communication overhead and handle system heterogeneity. After each server-side aggregation, it uses multiple historical iterates to build a higher-order extrapolation that speeds up convergence. Under mild assumptions, the authors establish convergence for a single-step variant and derive an O(1/K) rate. Experiments on synthetic and real-world datasets show improved performance over existing approaches.
In briefFrom Score to Diagnosis: A Search-Independent Post-Diagnosis Method for Symbolic Regression
SummaryMost symbolic regression solvers rely on a low-information scalar fitness score that offers little insight into why a formula fails or how to revise it. This paper proposes PostDiag, a feedback mechanism independent of the search process that can be shared across different solvers, using patch-wise features and two risk measures for diagnosis. Experiments on multiple LLM-based and GP-based solvers show it consistently improves accuracy and speeds up convergence.
In briefStructured latent dynamics to keep long-horizon robot world-model predictions physically consistent
SummaryThe paper proposes HWM, a framework that represents a world model's latent states in a phase space of coordinates and momenta, propagates them with dissipative Hamiltonian dynamics, and maps them back through a learnable inverse interface. It can be added to existing world models without changing their perception or decoding parts. In controlled simulations, it gives more accurate velocity, acceleration, and jerk estimates and substantially reduces long-horizon energy drift compared with unstructured latent dynamics. Across three embodied world-model settings and real AgiBot World trajectories, it also reports gains such as a 2.7% relative task-success improvement on RoboCasa.
In briefGTSM: Gaussian Splatting with Thermal State-Space Dynamics for Reconstructing and Predicting Spatio-Temporal Temperature Fields
SummaryThis paper proposes GTSM, a Gaussian Thermal State-Space Model that recovers a four-dimensional spatio-temporal thermal field from two-dimensional observations, without a predefined 3D model or thermodynamic equation. It models thermal evolution with a shared scene-level latent process and per-Gaussian local thermal states, propagated by differentiable RK2 integration. GTSM works with both static and dynamic Gaussian Splatting, enabling novel-view thermal rendering and future thermal-state prediction. Experiments on real infrared and synthetic data report better reconstruction accuracy and temporal extrapolation than representative baselines.
In briefSource-free adaptation of object detectors to adverse weather using contrastive hypergraphs (extended version)
SummaryFog, rain, snow, and low light degrade visibility and can impair object detectors trained on clear-weather data. When source data cannot be accessed for privacy or transmission reasons, source-free domain adaptation is needed. The paper proposes REUNION, which builds a contrastive hypergraph over target-domain object proposals to capture high-order relations, and optimizes it with a group-wise Hyper-InfoNCE objective. The authors report that it reduces domain shift across diverse benchmarks and achieves state-of-the-art results in source-free settings.
In briefA Music-Sheet Reader Tool for Reading String Quartet Scores
SummaryThis paper presents an optical music recognition system for PDF scores of string quartets. It first uses a SegNet-based model to locate musical regions, then applies convolutional neural networks and image processing to classify and extract symbols. Dynamic programming enforces time signature constraints and handles tuplets to improve rhythmic precision. On a benchmark dataset, it shows robust accuracy for pitch, onset, duration, and accidentals; code and data are public.
In briefRethinking EEG spatial super-resolution as dynamic state filtering
SummaryThis paper treats EEG spatial super-resolution, recovering dense brain signals from sparse electrodes, as a sequential state filtering problem. The proposed Dynamic State Filtering (DSF) separates predictable evolution from newly observed changes and fuses them using a topology-aware transition operator and uncertainty-aware correction. The authors report that DSF improves reconstruction fidelity and downstream utility on SEED, SEED-IV, and BCI2000, and remains robust under changing or unreliable sparse observations.
In briefVisConf: Quality-Aware Consensus for Best-of-N in Vision-Language Model Reasoning
SummaryThis paper proposes VisConf, a score for judging the quality of reasoning rollouts in vision-language models (VLMs) by multiplying visual engagement with calibrated Self-Certainty, so rollouts weak in either signal are penalized. It also introduces Rank-Weighted Consensus (RWC), a tuning-free rule that turns within-question VisConf ranks into answer vote weights. The authors report that VisConf with RWC consistently outperforms strong Best-of-N baselines across three VLMs on MathVista, MMMU-Pro, and MMStar.
Muhammad Usman, Muhammad Qasim Ayub, Usman Ahad et al.
In briefReconstructing unobserved core electron density in tokamak plasmas using a neural operator surrogate with uncertainty quantification
SummaryThis paper addresses the difficulty of observing the electron density profile in tokamak plasmas, where diagnostic access is limited. It uses a Fourier Neural Operator as a fast, differentiable surrogate of a transport solver to infer coefficients of a 1D particle transport equation from reflectometry and interferometry data. Four uncertainty quantification methods are compared, scored with CRPS and coverage calibration on observed and unobserved regions. The approach is shown to outperform standard analytical function fitting on synthetic and DIII-D experimental discharges.
In briefMOMO: Generating Human Motion from Moving Sounds
SummaryThis paper introduces a new task of generating human motion from dynamic spatial audio, where a sound source moves relative to the listener. The authors release two motion-capture datasets, DySAM and DySAL, and propose MOMO, a diffusion-based framework that estimates source direction from raw binaural audio and conditions motion on sound semantics and spatial cues. Experiments report that MOMO outperforms existing baselines in motion fidelity, audio-motion alignment, and physical plausibility.
In briefSparsified MLPs vs. Transcoders: Many Transcoder Features Are Disguised Neurons
SummaryThe paper examines whether MLP neurons in transformer language models, often seen as polysemantic, can serve as a basis for interpretability instead of transcoders. The authors define "neuron features" as transcoder features strongly aligned with individual MLP neurons, and find they make up a substantial fraction of active features in late layers across 13 open-source transcoders. The authors argue MLP neurons may be a stronger interpretability basis than commonly assumed and call for better methods to attribute non-neuron structure.
In briefHABIT-Bench: A benchmark testing whether agent memory can infer user habits from weak evidence
SummaryThis paper introduces HABIT-Bench, a benchmark for evaluating whether LLM agent memory systems can infer user habits from longitudinal weak evidence and apply them to later requests. It builds synthetic user histories across food, finance, software, and travel from task-oriented dialogue data and controlled habit graphs, with four-choice tasks covering temporary exceptions, habit changes, and unaccepted suggestions. Eight memory methods were evaluated, and the best achieved 33.61% macro accuracy, with performance remaining limited even when annotated evidence was supplied directly.
In briefAxiomatic World Modeling (RLAWM): penalizing wrong physics models during training to improve LLM physics reasoning
SummaryThe paper identifies "modeling drift" as a failure mode in LLM physics reasoning, where models alter given quantities, fabricate laws, or add unstated assumptions before reasoning in that distorted world. The authors propose RLAWM, a reinforcement learning method that penalizes divergence between the model's inferred axiomatic world model and the true physical world, using modeling and solving rewards. According to the abstract, it improves performance by up to 12.2 points on PhysReason and 5.2 on PHYSICS, with no added inference cost.
In briefImproving analytical light transport prediction in turbid media with physics-guided neural network corrections
SummaryThis paper proposes a physics-guided residual learning framework that uses the diffusion approximation as a physical prior and trains a neural network only on the residual between the analytical prediction and Monte Carlo reference results. Compared with the diffusion approximation, the framework reduces scalar-reflectance RMSE and spatially resolved reflectance RMSLE by roughly 62 and 14 times, respectively. Using only 1% of the training data, it achieves about 87% and 93% lower errors than a direct neural-network surrogate.
In briefM²Net: A Spatial-Temporal Forecasting Model Combining Scale-Specific Structure and Prototype Decomposition
SummaryThis paper proposes M²Net, a framework for spatial-temporal forecasting in urban management. It uses a multi-scale masked attention mechanism to learn spatial dependencies that change across time scales. It also adds a prototype residual learning module that separates shared global patterns from node-specific deviations. The authors report that M²Net outperforms state-of-the-art baselines on real-world datasets.
In briefProbePO: Sparse probes redistribute reward credit to reasoning steps in critic-free LLM reinforcement learning
SummaryThis paper addresses a problem in reinforcement learning from verifiable rewards, where a single final reward is given to a whole reasoning trajectory, making it hard to tell which steps helped. The proposed method, ProbePO, uses a reward-free uncertainty score to choose partial solutions to evaluate under a limited budget, then redistributes GRPO's trajectory-level advantage across tokens. Tested on DeepSeekMath and Qwen models from 1.5B to 7B parameters, it outperforms GRPO and, against the dense-probing VinePPO, achieves higher accuracy with 76% less training time.
In briefTopoCompile: Molecular generation that exactly meets global count targets
SummaryThis paper introduces TopoCompile, a method for generating novel molecules that exactly satisfy global count constraints such as molecular size, elemental composition, and ring content. Molecules are represented as trees of intact ring systems and individual atoms, and an autoregressive Transformer is combined with a dynamic-programming compiler that masks actions with no target-satisfying completion. The authors report 100% Exact-valid yield on Common and Rare four-count target groups across four benchmarks, and the highest validity or V.U.N. scores in unconditional generation.
In briefWhat Kind of Uncertainty Does Conformal Prediction Measure?
SummaryThis paper examines what the size of conformal prediction sets reveals about uncertainty. It shows that the difference in set size between full and split conformal prediction is directly tied to Bayesian posterior variance, the model's epistemic uncertainty. This difference combines two opposing effects: better accuracy from extra calibration data, and the model's formability toward each possible label. The authors propose that formability, estimated via posterior variance without the quantization of prediction sets, can measure epistemic uncertainty, performing on par with state-of-the-art methods in many setups.
Soroush H.Zargarbashi, Mohammad Sadegh Akhondzadeh, Rebekka Burkholz et al.
In briefDenoising Implicit Variational Inference: learns the marginal score directly to make ELBO gradients cheaper and more stable.
SummaryThis paper addresses the difficulty of optimizing the Evidence Lower Bound in semi-implicit variational inference, where the marginal score is generally intractable. It proposes Denoising Implicit Variational Inference (DIVI), which learns this score directly via denoising score matching and uses it as a plug-in estimator in the pathwise ELBO gradient, replacing inner reverse-conditional sampling. The authors analyze DIVI as inexact stochastic gradient ascent and bound the averaged stationarity measure under stated assumptions. Experiments on synthetic distributions and real-data Bayesian inference tasks show DIVI improves over current UIVI baselines while reducing ELBO-gradient estimation cost.
In briefBEM: Learning Bregman Divergences for Concept Erasure in Flow Matching Models
SummaryThis paper introduces BEM (Bregman erasure matching), a method for removing harmful, private, or copyrighted concepts from flow matching text-to-image models such as FLUX and Stable Diffusion 3. BEM learns the divergence through bi-level optimization so the model forgets the target concept on unseen prompts while unrelated generations remain unchanged. Under the GEM evaluation protocol, BEM removes more of the concept than GEM on most settings across nudity, violence, real identities, and copyrighted characters, while retention holds for some concepts and degrades for others.
In briefGazeAct: A Human-Gaze-Annotated Robot Manipulation Dataset for Attention-Regularized Policies
SummaryThis paper introduces GazeAct, a real-robot manipulation dataset with human gaze annotations and a training framework for multi-view vision-language-action policies. Operators' gaze on the robot camera feed was recorded during screen-based teleoperation, yielding 12,000 demonstrations across 40 tasks and five domains. Gaze-derived targets regularize language-conditioned visual attention during training, while inference remains gaze-free. Across 18 real-world tasks, average success rose from 48.1% to 58.6%.
In briefSAP-Depth: A scene-adaptive probability framework for self-supervised depth estimation from single images
SummaryThis paper proposes SAP-Depth, a new architecture for self-supervised monocular depth estimation, which predicts depth from a single camera image without labeled depth data. It uses a Dynamic Bins mechanism with a learnable parameter predictor to adapt depth distributions to each scene, and a decoder that fuses multi-scale features to predict dense depth probabilities. Experiments on KITTI show better accuracy than prior state-of-the-art methods, especially at preserving complex edges, and tests on Make3D indicate good generalization and robustness.
In briefHiPPO Functional Memory: An attention method that answers time queries at any point in irregular data
SummaryThis paper introduces HiPPO Functional Memory (HiPPO-FM), which compresses an irregular time-series history into a reusable state that can be queried at arbitrary time coordinates. The method is evaluated on continuous-coordinate reconstruction, irregular representation learning, real-world extrapolation, and event-sequence prediction. The authors report consistent gains over strong irregular time-series baselines, and the compact state can be reused across query sets without reprocessing the observations.
Zhaodong Guo (Imperial College London), Haoyun Feng (Imperial College London)
In briefLength bias in vision-language models answering multiple-choice questions
SummaryThis paper examines length bias, where large vision-language models tend to prefer longer answer options in multiple-choice questions, even when those options are visually inconsistent. Accuracy shifts by 2–38% when correct answers are shorter and rises by 1–70% when they are longer, and the bias persists across model scales. Layer-wise analysis attributes it to late-layer suppression of visual evidence in favor of language priors, and the authors find that the inference-time mitigation strategies they tested only partially help.
In briefTideKV: Layer-wise delayed KV eviction to reduce peak memory in long-context LLM inference
SummaryThis paper addresses the high peak memory that long-context LLM inference needs during prompt prefill. It proposes TideKV, an on-GPU KV cache construction method that processes all prompt chunks within each Transformer layer, keeps full-prompt KV only for the active layer, and evicts at layer end. Experiments on several platforms report that TideKV preserves inference quality and final cache size of the compared methods, cuts construction-time transient memory substantially, and enables 128K prefill where baselines run out of memory.
In briefChecking whether chain-of-thought stays monitorable when hidden reasoning needs computation
SummaryThe paper replicates Emmons et al. (2025), testing whether language models reveal hints in their chain of thought, across eleven models from six families. Models follow simple hints above a no-hint baseline, while concealed complex-hint following mostly falls back near baseline. The authors extend the study to five cue families and three difficulty levels, finding the link between chain-of-thought necessity and unfaithfulness depends on task and model.
In briefCodecMotion-OAD: Fast Online Action Detection Using Video Codec Motion Vectors
SummaryThis paper addresses online action detection in streaming video, where dense optical flow is a major computational bottleneck. The proposed CodecMotion-OAD replaces optical flow with motion vectors provided by the video decoder while keeping existing temporal detector architectures, trained in two stages with a frozen RGB+Flow teacher. On THUMOS'14 with MiniROAD, it reaches 68.3 mAP compared with 70.9 mAP for matched-distillation RGB+Flow, and reduces decoder-to-prediction processing time by 11.48-fold on CPU and 2.04-fold on GPU.
In briefConCI: A controlled counterfactual benchmark testing whether LLMs judge context in privacy decisions
SummaryThis paper introduces ConCI, a benchmark for testing whether large language models judge the appropriateness of sharing personal information based on context. Each test unit pairs an anchor information flow with two counterfactuals that each change a single contextual parameter, and credit is given only when all three are judged correctly. Across seven LLMs, many correct anchor predictions fail on at least one counterfactual, and models perform better on inappropriate flows than on appropriate ones, often continuing to label flows inappropriate after context changes. Larger models reduce but do not eliminate this non-disclosure bias, and reasoning gives only modest gains.
In briefBackend-Frontend Converter: Turning .NET Computational Models into TypeScript Components
SummaryThis paper presents Backend-Frontend Converter, a tool that generates TypeScript code from component-based .NET computational models so they can run in web browsers. The authors evaluate the generated code on four orbital scenarios and a recorded trading-strategy comparison against C# results. They report that paired outputs match in timing and remain very close numerically, and that their comparison procedure detects deliberately introduced perturbations.
Petr Ivankov (1), Max Shestov (2) ((1) Independent researcher, (2) Georgia Institute of Technology)
In briefVBO-MI: A differentiable black-box optimization method using variational mutual information
SummaryThis paper proposes VBO-MI, a gradient-based Bayesian optimization framework for expensive black-box functions that can only be evaluated with noisy measurements. It uses an actor-critic design, with an action network that explores the input space and a variational critic that estimates an information signal, plus a lightweight surrogate trained on a replay buffer so each iteration needs only one batch of real queries. The authors report theoretical convergence results and experiments on high-dimensional synthetic functions and tasks including rover trajectory planning, Lunar Lander, and pest control, where it matches or outperforms baselines.
In briefVocal-tract acoustics and aerodynamics in DYNARTmo, a model that synthesizes speech
SummaryThis paper extends DYNARTmo, a dynamic articulatory model for visualizing and synthesizing speech, with acoustic and aerodynamic components. Gesture trajectories set the vocal-tract shape, which is modeled as an acoustic tube with nasal and lateral branches, while a lumped network links glottal motion and lung pressure to airflow. Three German words are used as examples, and the author notes that the implementation was tuned mainly for recognizability, independent perceptual evaluation is still pending, and fricative noise needs refinement.
In briefModeling Discourse Profiles in Climate Misinformation Discourse
SummaryThis paper argues that climate misinformation is spread through recurring rhetorical patterns, not just individual comments. Using a large TikTok climate dataset, the authors cluster comments into discourse profiles and add them to classifiers via late fusion. Profile-aware models significantly improve macro-F1 over text-only baselines for several representations, though the RoBERTa-large gain is not statistically significant.
In briefLearning Adjustable Multi-Preference AI Policies from Anonymous Crowd Feedback
SummaryThis paper proposes P3L (Preference-PCA Pluralistic Learning), a framework that lets AI systems learn from diverse user preferences and personalize behavior for individuals. It learns trajectory embeddings and a shared basis of preference directions via PCA from crowd feedback without needing annotator identities. The authors report that experiments on embodied control tasks show generalization to new out-of-distribution users, competitive RL performance, and substantially lower computational cost.
Akansha Kalra, Ellen R. Novoseller, Vernon J. Lawhern et al.
In briefMedPAO-Assist: An Edge-Deployable, Modality-Aware AI Agent That Structures Radiology Reports
SummaryThe paper presents MedPAO-Assist, an agentic AI framework that runs locally on edge devices and turns radiologist dictation and voice commands into structured radiology reports. It uses a Small Language Model of about 0.6B parameters, needing roughly 3.3 GB peak RAM and averaging about 31 seconds end-to-end on a standard laptop CPU. The system supports chest X-ray, chest CT, abdomen CT, and brain MRI, and works without keyboard input, external APIs, or internet access.
