In briefKhyontekAI's CheckThat! 2026 Study on Multilingual Retrieval of Scientific Sources for Web Claims
SummaryThis paper addresses CLEF 2026 CheckThat! Task 1, which aims to automatically find the scientific publications behind claims shared on social media that mention research without linking to it. The authors propose a two-stage multilingual system that combines fine-tuned dense retrieval models (multilingual-E5, BGE, Contriever) with cross-encoder reranking, and compare metadata configurations and multilingual training strategies. They report that metadata enrichment consistently improves retrieval and that cross-encoder reranking adds further precision gains. The final system reaches MRR@5 scores of 0.6222 for English, 0.5112 for German, 0.6254 for French, and an overall average of 0.5862.
In briefLatent Taxonomies: How Models Learn to Generalize Properties During Pre-training
SummaryThe authors build a hierarchical taxonomy of synthetic entities grouped by shared properties and inject its facts into continued pre-training of Pythia and OLMo 2 models from 70M to 7B parameters. They find that taxonomically aligned induction emerges from next-token prediction alone, transferring more strongly within categories. With a fixed data budget, induction weakens as model size grows, suggesting larger models memorize facts rather than compress them into shared structure.
WhyBrings category-based property induction to LM pretraining using synthetic taxonomies across multiple scales, with a novel finding that induction weakens with size; experiments are fairly thorough.
Jan Sobotka, Ekaterina Goliakova, Marcin Detyniecki et al.
In briefRecommending local citations without relying on citation graph signals
SummaryThis paper studies local citation recommendation, which retrieves suitable references for a given sentence. The authors release LCRC, a new dataset of 14,735 scientific statements drawn from astrophysics review papers, each linked to one or more citation targets. Using it, they build a system that avoids citation graph bias and, without fine-tuning, improves Recall@25 by 8.0% and NDCG@25 by 6.9% over dense retrieval baselines. Code and dataset are publicly available.
Benjamin Basseri, Pavlos Protopapas, Kemeng Zhang et al.
In briefSame averages, different adaptation: how hidden second-moment structure shapes learning in shared networks
SummaryThe paper studies small Adam-trained shared-representation networks that match on selected mean readout measurements yet respond differently to further learning. Using state interventions, the authors separate second-moment scale effects from coordinate-pattern effects and examine how they form and change over time. Exchanging second moments reverses the ten-update contrast, while exchanging first moments does not, indicating an effect beyond overall update size.
In briefGuarded block transfers for SVM training: linear convergence with explicit constants and gap-certified screening
SummaryThe paper addresses hard-margin SVM training, which is framed as finding the distance between two convex hulls. It proposes guarded block transfers that combine several MDM transfers into one exact line search, falling back to an away step when needed. The authors prove global linear convergence with explicit constants on the Minkowski-difference polytope and bound the iterations needed to screen out zero-weight points. Experiments show matching accuracy with baselines, but the solver is slower than LIBSVM or LIBLINEAR in several settings, so its main advantage is the certified duality gap.
In briefBenchmark of deep learning models for spotting anomalies in real cloud telemetry data
SummaryThis paper empirically benchmarks deep learning models (GRU, Transformer, TimesNet, Autoformer) for anomaly detection on a real IBM Cloud Console dataset. It uses the Numenta Anomaly Benchmark (NAB), a timing-aware metric that accounts for early detection and false alarms. TimesNet achieves the highest NAB scores, while an improved GRU autoencoder offers competitive results with a more stable false-alarm profile. The authors conclude that real cloud anomalies are mainly short-term deviations, which makes timing-aware evaluation and robust post-processing important.
In briefML-CLAPPS: A Probabilistic Sampler for Prioritized Imbalanced Multilabel Learning
SummaryThis paper introduces ML-CLAPPS, a learnable sampler that optimizes how training instances are drawn, controlling the model's exposure to different classes in multilabel learning. It addresses class imbalance while letting users prioritize classes through a user-defined importance vector, such as high-risk ECG conditions. Experiments on a real-world multilabel ECG image dataset report state-of-the-art performance, and raising a class's importance improves its classification results.
In briefChangeT³: Testing a new vision method for spotting changes in satellite images
SummaryThis paper presents ChangeT³, a Siamese model for detecting changes between two remote sensing images taken at different times. It pairs a ViT³-Tiny encoder with a ChangeFormer decoder, computing fast weights separately for each image. On LEVIR-CD+ it reaches an F1 of 80.61±0.90, versus 78.84±0.21 for ChangeFormer, but neither model transfers zero-shot to SYSU-CD, DSIFN-CD, or WHU-CD. It is also slower than ChangeFormer at small input sizes.
Sharukshan Niranjan, Oshadha Samarakoon, Nichula Wasalathilaka et al.
In briefAdaptive Look, Flexible Thought: Learning How and How Long Multimodal Models Should Reason
SummaryThis paper proposes L2T, a framework that converts visual states produced by external tools into visual tokens, so perception and text reasoning share one autoregressive space. It uses a modality-aware routing reward to choose reasoning modes and a difficulty-aware fast-slow strategy to set reasoning depth. Experiments on perception and reasoning tasks show that, compared with explicit tool-calling methods, L2T stays competitive while reaching 1.7 to 3.6 times higher throughput.
In briefTesting Deep Learning Frameworks by Tracking Data Flow
SummaryThis paper introduces the Tensor Flow Graph (TFG), which captures runtime tensor values and operator input requirements, and builds FlowMuT, a mutation-based testing method on top of it. The authors evaluate FlowMuT on PyTorch and MindSpore, reporting 55 bug reports for cases missed by baselines, of which developers confirmed 41 and fixed the bugs in 7. They report 43.4% higher data flow diversity than the strongest baseline and a legal mutant rate of 96.7%.
In briefTotem: Object-Centric 3D Memory for Consistent Video Worlds
SummaryThis paper presents Totem, a framework that generates videos consistent with a unified 3D world, especially under large viewpoint changes. It stores each selected object as a persistent 3D memory aligned to a shared world frame, and uses a dual-stream architecture that combines these object memories with scene geometry to synthesize target views. Experiments on world-generation and object-centric benchmarks show that Totem keeps objects more consistent under challenging camera trajectories.
In briefRecovering damaged video frames using depth-guided single-reference adaptation
SummaryThis paper studies recovering video content from corrupted bitstreams by combining RGB appearance with depth information. The proposed framework selects only one non-local RGB reference per recovery window and uses Depth-Conditioned Reference Adaptation to integrate it into feature completion. On the BSCV benchmark with oracle masks, it reports 32.30 dB on YouTube-VOS and 28.93 dB on DAVIS, outperforming full-reference B2SCVR. Receiver-side latency and peak memory are reported to be 36.8% and 40.8% lower, respectively.
In briefTests Pass, Meaning Missing: Evidence of Interpretive Lock-In in Code Generation
SummaryThe paper studies whether a mistaken interpretation of a task can persist through planning, implementation, and self-written tests in DataFrame code generation, using 30 tasks derived from TPC-DS. Test-pass counts rose to 29/30 and 30/30, yet no generated solution matched its stored reference output. The authors describe their findings as diagnostic rather than causal, and conclude that increasingly successful self-generated tests did not establish semantic correctness.
In briefAn explainable system detecting offensive content in Nastaliq Urdu, with a new corpus, model comparisons, and word-level explanations
SummaryThe study builds a publicly released corpus of 7,604 Nastaliq Urdu Facebook comments for offensive content detection. Four Urdu-capable transformers are compared with twelve feature-based baselines on identical splits, with MuRIL performing best at 87.05% accuracy. Word-level SHAP analysis shows attributions concentrate on frequent slurs but also reach neutral function words and political vocabulary, suggesting performance may partly reflect lexical and topical patterns.
In briefGeo-vision: A Method Using Geometry and Topology to Classify Vehicle Damage
SummaryThis paper presents Geo-vision, which classifies vehicle-part damage using a fixed representation that combines appearance, multiscale geometry, orientation, and sampled digital topology features. On 55,303 crops from 4,916 source groups, it reaches an average precision of 0.7577, above 0.7458 for a frozen ImageNet-pretrained YOLO11n on the same crops. The authors report that it uses no external encoder pretraining but is slower than YOLO, and note that independent external validation remains necessary.
Usef Faghihi, Amir Saki, Serge Jonathan Kogbeti-Wongo
In briefEstimating the Value of Policy Improvement in Adaptive Experiments
SummaryThis paper proposes an asymptotically normal estimator for the value of policy improvement (VPI), the gap between the best treatment policy in a class and a baseline policy, under adaptively collected data. The estimator sequentially fits a weighted empirical welfare maximizing policy on historical data and evaluates it with inverse propensity weighting on future samples. The authors test coverage on synthetic and semi-synthetic data and apply it to an adaptive experiment on Jordanian job assistance programs. Their estimates are more conservative than the original hierarchical Bayesian analysis, yet still support benefits of personalization among Syrian refugees.
WhyProposes an asymptotically normal VPI estimator under adaptive data and non-unique optima, with faster regret rates shown, validated on synthetic and real experiments.
Justin Whitehouse, Emma Brunskill, and Vasilis Syrgkanis
In briefAgent Sociodrama: Social-Dynamics-Guided Persona Recasting for LLM Agent Teams
SummaryThis paper presents Agent Sociodrama, a framework where an external LLM judge updates each agent's emotional state and directed, asymmetric relationships every round to model social dynamics in LLM agent teams. The authors assess social fidelity across six LLM backbones and five social scenarios using classical social-theory patterns and agreement with human judgments. They also introduce Trajectory-Guided Persona Recasting (TPR), which diagnoses group-dynamic failures from interaction trajectories and rewrites personas before rerunning tasks, improving performance on four collaborative benchmarks.
In briefRSI-Nav: Policy-Synchronized Recursive Self-Improvement for Navigation Data
SummaryThis paper presents RSI-Nav, a framework for vision-language navigation in which one shared model both follows instructions and generates instructions for unlabeled trajectories. It processes trajectories in small shards, keeps generated instructions that pass goal-completion and path-efficiency checks, and updates the model before the next shard. The authors report success rates of 68.6% on R2R-CE and 72.4% on RxR-CE, along with better instruction generation and modestly improved spatial understanding.
In briefHardware-aware pruning that picks a sparsity structure per layer for efficient LLMs
SummaryThe paper proposes Heterogeneous Structured Pruning (HSP), a training-free framework that considers latency and memory on target hardware when pruning large language models. It chooses among dense, unstructured, 2:4, block-sparse, and head/channel pruning options for each sublayer through a global solver that minimises proxy pruning loss under a resource budget. Evaluated on OPT and LLaMA models on an NVIDIA Jetson Orin Nano Super edge device, HSP is reported to improve quality-resource trade-offs over training-free baselines.
Cameron Barker, Andreas Paraskeva, Jan N. van Rijn et al.
In briefWho Sets the Pace of Judge Hacking: Judge Capability or Policy-Judge Gap
SummaryThis paper trains small Qwen2.5 policies with GRPO on GSM8K against frozen LLM judges to study how reinforcement learning exploits judge false positives. The depth of exploitation correlates with judge AUC, with larger judges fooled less deeply, while speed results partly contradict the capability-gap account. With only 12 cells, the authors cannot formally separate judge AUC from the policy-judge gap for either endpoint.
In briefRSIPO: A self-evolving reinforcement learning method where agent skills and policy improve together
SummaryThe paper proposes RSIPO, an agentic reinforcement learning framework in which the policy model extracts skills from its own completed trajectories and improves its skill generation through trajectory replay. Skills and the policy co-evolve during training. The method also uses empirical Bayes shrinkage to adjust how much the student trusts teacher scores. With Qwen2.5-7B, RSIPO reports a 98.6% average success rate on ALFWorld, outperforming all compared baselines.
In briefHand2Robot: object-centric 4D hand reconstruction from human videos to supervise robot policy training
SummaryThe paper presents Hand2Robot, a framework that turns videos of humans performing a task into robot action supervision. It extracts interaction clips, reconstructs object geometry, and refines hand motion with hand-object interaction optimization and retargeting. Experiments on DexYCB and HO3D-v3 report improved object pose accuracy and interaction quality. On three real-robot dexterous manipulation tasks, the reconstructed demonstrations are reported to effectively supplement limited robot data for policy post-training.
In briefSurgiCLAW: an interpretable robot agent that picks precise surgical micro-skills
SummarySurgiCLAW is a hierarchical framework for precision surgical robots, where a multimodal language model handles visual understanding, explanation, and selecting atomic surgical skills. Physical execution is delegated to verifiable deterministic control programs. The authors built and expanded a state-to-skill dataset for brain-machine-interface implantation and trained the model with supervised fine-tuning plus three rewards. On a fixed test set it reached 94.6% skill-selection accuracy and 93.0% Macro-F1, and it achieved an 88.0% task success rate across 50 closed-loop real-robot trials.
In briefA Decade in a Day: Building Agent-Friendly Desktop Interfaces from Software Evolution
SummaryThis paper examines how desktop software interfaces, designed for humans, affect the performance of large language model agents. The authors introduce DeskIF-Bench, a benchmark of 255 verified tasks across six applications, and find that newer software releases consistently lead to better agent results. They propose ANIS, a framework that synthesizes agent-native interfaces, which improves success rates and reduces online token usage on an expanded evaluation of 368 tasks.
WhyNovel use of software evolution as a natural variable for agent-native interfaces, backed by a 255-task benchmark and multi-model, multi-application experiments showing clear gains.
In briefSlotAVS: Object-Centric Audio-Visual Segmentation via Bi-Modal Slot Attention
SummaryThis paper presents SlotAVS, a method for audio-visual segmentation that outlines pixel-level masks of sounding objects in video. It creates object slots from audio sequences and has them compete over visual tokens, which reduces merging of multiple sound sources, and adds Hungarian matching and confidence-aware mask fusion. On the AVSBench dataset, it reports 88.6% mIoU on the Single-Source (S4) split and 71.7% mIoU on the Multi-Source (MS3) split.
In briefRobust Marathi Sentiment Analysis Using Domain-Adaptive Pretraining and Data Augmentation
SummaryThis paper studies sentiment classification for Marathi, a widely spoken but low-resource language, across multiple domains. It uses the pretrained MahaBERT model, applies domain-adaptive pretraining with masked language modeling, and uses back-translation to augment training data. The authors report consistent accuracy gains in all domains, ranging from about 0.12% to 1.74% depending on the dataset.
In briefHyperbolic Time Series Forecasting with Learnable Decomposition and Manifold Dynamics
SummaryThis paper proposes HyperForecast, a long-term time series forecasting framework that works in hyperbolic space on the Poincaré ball. It decomposes signals into components ordered from low to high frequency, placing them at increasing radial distances. Two manifold-dynamics variants, autoregressive and non-autoregressive, are used for prediction. The authors report it outperforms baselines on 9 of 11 benchmarks, with up to 58% lower mean squared error at long horizons.
In briefImproving Zero-Shot Cross-Domain Sentiment Analysis for a Low-Resource Language: A Marathi Case Study
SummaryThis paper studies whether training sentiment classifiers for Marathi on multiple domains improves accuracy on domains the model has not seen. The authors compare models trained on one, two, and three domain datasets. Training on two domains improved accuracy on unseen domains by about 8% over single-domain training. A third domain brought further but smaller gains, averaging around 2% for some domains.
In briefFrozen History: Why Temporal Losses Can Stop Accumulation in Learned Supersampling
SummaryThis paper studies how a temporal consistency loss affects learned real-time supersampling, which combines jittered low-resolution frames along motion vectors. The authors find that with accurate registration, the network can copy its warped history on static content at almost no temporal cost, suppressing accumulation. They identify four registration errors that training can hide. Under exact registration, a higher temporal weight reduces pixel variance but lowers accuracy and perceived quality, losing 0.7 to 0.9 dB on static frames at weight 0.25.
WhyShows registration errors let temporal loss suppress accumulation, validated by controlled training and a model-free oracle across architectures and scales.
In briefGoogle Earth AI Geospatial Reasoning Agent: Capabilities and Systematic Evaluation
SummaryThis paper presents the Geospatial Reasoning agent (GRA), a Gemini-powered system from Google Earth AI that combines geospatial data sources, foundation models, and tools to answer questions about the planet. It breaks complex queries into multi-stage workflows through upfront planning. The authors introduce two benchmarks, GeoExpertQueries and GeoKTools, and compare four configurations. GRA's skills and tools give the largest gain, and its planning and execution guidance further improves open-ended problems, letting Gemini 3.8 Flash match or exceed Pro at lower cost.
In briefSpikeOcc: Spiking experts for more resilient multimodal 3D occupancy perception in driving
SummaryThis paper studies multimodal 3D occupancy prediction for autonomous driving, focusing on robustness when sensors are missing. It proposes SpikeOcc, a spiking neural network with a Spiking Mixture-of-Experts and Sensor Availability Learning. Experiments show it keeps competitive accuracy with all sensors while raising performance under sensor outages.
In briefRSAJEV: A training-free, abstaining prototype memory that removes silent hallucination in MLLM-based remote sensing classification
SummaryThis paper studies scene classification of remote sensing images with multimodal large language models that pick answers from lettered options. The authors find this method is accurate within its training vocabulary but unstable on unseen classes, where confidence is higher for unseen than seen classes, which they call silent hallucination. They keep the answer-slot representations and replace the readout with a training-free prototype retrieval method using shrinkage-Mahalanobis correction. It reaches 89.43% on the UCM cross-dataset test with 16 references per class and 96.90% on AID, with open-set AUROC of 0.91 to 0.99.
In briefSeen but not avoided: repeated proposals of rejected candidates in LLM optimization loops
SummaryThis paper studies whether LLM proposers avoid re-proposing candidates they already tried and rejected in catalog-based optimization loops. In experiments with three open LLMs improving a GUI agent, static probes underpredicted re-proposal within the proposers' own loops, and already evaluated harness-edit pairs consumed 28-63% of valid proposer iterations. Stating the rejection status in one sentence lowered re-proposal across all models.
In briefScenePlacer: Estimating Object Poses in Scenes as Visibility Field Learning
SummaryScenePlacer is a framework that places 3D objects produced by state-of-the-art shape generation models into scenes, using a partial geometric observation, without changing the generation pipeline. It separates object generation from pose estimation by predicting a visibility field in object space and then aligning the partial observation to visible mesh parts to get a rigid transformation. On 3D-FUTURE it outperforms existing baselines, reaching a Chamfer distance of 0.0077 and F-score of 92.40 with estimated depth, and 0.0023 and 98.27 with ground-truth geometry.
In briefNepPlural: A Human-Anchored Benchmark for Labelling Personas in Nepali Migration Debate
SummaryThis paper introduces NepPlural, a benchmark for labelling competing socio-cultural personas in Nepali online discussions of youth migration and brain drain. The authors annotated 2,148 code-mixed YouTube comments on four persona axes plus a No-Persona class, and checked them against a 300-comment gold set labelled by three native speakers. They report that large language model annotation reliability depends heavily on protocol, and that allowing models to abstain lets them match or exceed individual native speakers.
In briefA dependency-aware multi-GPU training framework that speeds up temporal graph neural networks
SummaryTemporal graph neural networks store node memories to track graph events, but multi-GPU training is hard because later predictions depend on earlier memory updates. The paper proposes DepBloom, which uses temporally bounded spatial partitioning, intra-batch memory refinement, and critical-path optimization to reduce cross-GPU synchronization. The authors report an average speedup of 2.84x and up to 5.16x over state-of-the-art methods, with comparable or better accuracy.
In briefVisually readable yet adversarial: watermark-based attacks on document-understanding vision-language models
SummaryThe paper proposes AR-AWM, a white-box attack that optimizes only the pixels of a small, spatially distributed watermark-shaped text overlay to disrupt end-to-end vision-language models, without using questions or language-model signals. On Qwen2.5-VL-7B, covering 7.98% of page pixels drops DocVQA exact match from 0.899 to 0.150 and ANLS from 0.939 to 0.212. Unoptimized and random-perturbation controls retain performance, indicating adversarial optimization is the main cause. Cross-model transfer is weak.
In briefExploring the generative range of diffusion models with spectral-guided quality-diversity search
SummaryStandard diffusion sampling tends to favor high-likelihood regions, which limits output diversity. This paper combines Quality-Diversity optimisation with diffusion models, directly optimising latent noise and conditioning embeddings, and introduces an h-space Spectral Mutation operator that searches along directions of maximal semantic sensitivity derived from principal singular vectors of the model's Jacobian. On text-to-image generation, the combined method achieves the lowest FID score against 12 baselines, showing a good balance between image quality and diversity.
Hannah Janmohamed, Joao Correia, Penousal Machado et al.
In briefRobust adaptive confidence sets in high dimensions: statistical and computational limits
SummaryThis paper studies adaptive confidence sets for the mean of an identity-covariance multivariate Gaussian under Huber contamination with unknown contamination proportion. It shows that an information-theoretically optimal adaptive set can be built by quantifying uncertainty across all one-dimensional projections, but this construction is computationally infeasible. A Statistical Query lower bound shows that achieving diameter O(m^{-1/2}) requires exponentially many queries or tiny query tolerance, and a low-degree moment-matching algorithm using d^{O(m)} samples and time is also given.
Ilias Diakonikolas, Chao Gao, Daniel M. Kane et al.
In briefPhySimCode: A Benchmark for Turning Physics Videos into Runnable Simulation Code
SummaryThis paper introduces PhySimCode, a benchmark that tests whether AI models can watch a raw video, infer the physics behind it, recover its parameters, and write runnable Python code that reproduces the video. The dataset has 160.6K procedurally generated samples across 162 physics phenomena. Evaluations of ten models found they can state physical laws fairly well but struggle to infer them from video, and the best model recovered only 7.7% of ground-truth parameters within ±20%.
In briefHistory Replays Itself: Unlocking Long-Context Ability in Linear Attention via Sparse Memory Replay
SummaryLinear attention models use a fixed-size recurrent state that must both support computation and store long histories, which has been seen as limiting their long-context ability. The authors propose NoteBank, a sparse episodic memory that archives selected text before queries are known and reads it back through the model's native forward computation, without extra training or external retrievers. At one million tokens, the paper reports that Gated DeltaNet's mean accuracy on six RULER retrieval tasks rises from 1.7% to 70.3%.
WhyTraining-free sparse episodic memory with native replay lifts Gated DeltaNet's million-token retrieval accuracy from 1.7% to 70.3%, validated across several backbones and benchmarks.
In briefMedHarnessBench: A Unified Benchmark for Evaluating Healthcare Agents
SummaryThis paper introduces the Medical Harness Protocol, which independently assesses healthcare agents across seven process dimensions, and builds MedHarnessBench with 342 executable tasks. Across more than 16,000 controlled executions, healthcare-specific runtime support improved task success, subtask completion, and process quality. The authors also show the evaluation can reveal failures that terminal outcomes alone would miss.
In briefTL;QR: A query tool for finding how language changes over time in text collections
SummaryThe paper introduces TL;QR, a declarative query language and open-source engine for finding patterns in text that unfold over time. Queries combine natural-language predicates judged by a Large Language Model with regular expressions and temporal logic operators, while a symbolic engine aggregates the LLM's local judgments and links each match to its supporting passages. On controlled synthetic benchmarks, the authors report higher accuracy than LLM-only baselines. In a case study of ten years of European Parliament debates, it identifies candidate shifts in political groups' stances on the GDPR.
Fran\c{c}ois Hublet, Vladimir Krsmanovic, Jesus Solano et al.
In briefWhat DNA foundation models learn beyond basic sequence patterns
SummaryThis paper tests whether five frozen DNA foundation models provide predictive information beyond normalized k-mer frequencies, using 57 genomic classification datasets and two gene-expression regression tasks. Simple k-mer features remain competitive with most models, while NTv3 shows a consistent advantage and improves 46 of 57 datasets when combined with 6-mer features. The authors also find that linear recoverability from k-mer frequencies does not reliably indicate downstream usefulness.
Vincenzo Yuto Civale, Andrew David Bagdanov, Alberto Magi
In briefSuggested Is Not Superseded: Pre-registered test of a fast decision model for memory retirement
SummaryThis paper studies when a new utterance should retire an old stored claim in assistant long-term memory. The authors define a three-way contract (SUPERSEDED, NEEDS-CONFIRMATION, COMPATIBLE), label 500 test pairs with LLM annotators and human adjudicators, and evaluate a model called Jev. A binary gate prematurely retired 39.2% of merely suggested changes, which fell to 15.5% under the three-way decision at a 15.3-point detection cost.
In briefKeeping small updates in low-bit recurrent state quantization
SummaryLinear attention and recurrent models store past context in a fixed-size state that changes at every step, and keeping this state in high precision is costly at large batch sizes. The paper finds that quantization errors in these states accumulate and are highly coherent across steps, mainly because small updates are lost during rounding. The authors propose Low-Rank Bypass, which buffers small updates and periodically merges large ones, and Bounded Directional Rounding, which rounds reversed updates toward their direction without extra storage. Their training-free method substantially outperforms INT4 on Gated DeltaNet, GLA, Mamba2, and RetNet, reducing state memory by up to 5.3x and speeding up decoding by up to 1.9x relative to FP32 states.
In briefSpecForge: Recursive Self-Improvement for Coding Agents via Behavioral Specs
SummaryThis paper studies how frozen-model coding agents can improve by refining natural-language behavioral specifications, without updating model weights. It proposes SpecForge, a multi-frontier tree search framework with three mechanisms that handle interdependent refinements and output coupling. On DevEval+, it achieves the best results across five backbone models, and the improved specifications transfer to cross-language translation and new-feature implementation.
In briefGeneralizable model editing may go beyond direct generalization
SummaryThis paper argues that model editing has aimed for direct generalization, treating factual updates as input-output mappings, but this paradigm faces theoretical identifiability problems and practical limits. The authors propose compositional generalization, enabled by the reasoning and composition abilities of modern LLMs, as a theoretically grounded and empirically validated alternative. They report that this behavior appears across multiple models and editing methods. The paper calls for a shift toward more generalizable and deployable knowledge update techniques.
In briefBharatSetu: A 1M-Row by 22-Language Bridge Corpus for Indic Translation
SummaryThis paper releases BharatSetu 22-Indic 1M, a table of 984,764 English sentences each machine-translated into all 22 languages in India's Eighth Schedule, yielding about 21.7 million Indic entries. Because every language column translates the same English sentence, any two columns form a translation pair, enabling pivot-derived pairs for languages lacking direct parallel data such as Bodo, Dogri, Santali and Kashmiri. The corpus is machine-generated and not human-verified, so the authors propose a confidence layer and a human-correction layer keyed to row identifiers, along with an evaluation protocol. No translation-quality results are reported.
In briefRespFocus: Learning Temporal Evidence for Weakly Supervised Respiratory Sound Classification
SummaryThis paper presents RespFocus, a weakly supervised framework that uses multiple instance learning to learn inspectable temporal acoustic evidence from cycle-level labels and integrate it into classification. The authors also introduce PedResp2026, a multicenter pediatric respiratory sound dataset with adjudicated cycle-level annotations. RespFocus achieves four-class Scores of 69.55% on ICBHI 2017 and 72.91% on PedResp2026.
In briefTabMosaic: One Diffusion Model for Many Table Generation Tasks
SummaryThis paper treats tasks such as synthesizing records, filling missing values, and conditional generation on tabular data as one problem: modeling the joint distribution of any target cells given observed cells of an unseen table. The authors propose TabMosaic, a unified in-context diffusion model that generates mixed-type target cells without table-specific training. Across 48 real-world tables, it ranks first on all evaluated generative tasks against 20 baseline approaches.
WhyUnifies tabular generative tasks as joint conditional diffusion over arbitrary target cells, linking the objective to posterior predictive; ranks first across 48 real tables versus 20 baselines.
In briefDRanker: A Transformer-Based Multi-Task Ranking Model for Disney+ Homepage Recommendations
SummaryThis paper presents DRanker, a transformer-based multi-task model that ranks personalized recommendations on the Disney+ homepage. It uses an encoder-only architecture over enriched engagement-history sequences, with a lightweight causal debiasing mechanism. The authors report that simpler designs outperformed more complex alternatives on offline metrics, behavioral evaluation, and in production. After global deployment, DRanker produced over 1% lift in a key engagement metric and 0.5% lift in daily active users.
Priya Nirmal Singh Khokher, Daniel Nemirovsky, Benjamin P. Whitesell et al.
In briefDo LLM Tutors Really Teach? A Process-Based Review of the Tutoring Loop
SummaryThis paper reviews large language model tutoring as a sequence of stages, from inferring learner state to pedagogical action, tutor training, evaluation, and student simulation. A structured audit of studies finds that pedagogical evaluation is common, but evidence is much thinner for learner-conditioned intervention, human learning gains, and especially transfer or delayed retention. Human studies show that LLM tutors can improve learning under conditions, with benefits depending on pedagogical structure, learner engagement, and how learning is measured. The authors argue that validating the interfaces between tutoring stages is needed, not just improving individual components.
In briefFARL: Fully Automated Pipeline for Designing RL Interfaces and Solvers
SummaryThis paper introduces FARL, a framework that automatically designs reinforcement learning setups by jointly searching observation and reward programs, learning algorithms, and training settings. It diagnoses failures from training statistics and observed behavior, proposes changes, and tests whether they improve learning. On the Adroit Hand Relocate task at 5M training steps per policy, FARL reaches 97.1% success, versus 0% for Eureka and LiMEN and 1.0% for native PPO.
In briefNotes on uniform score bounds and Lipschitz continuity for several probability-based decision policies
SummaryThis paper studies truncated Gaussian, Cauchy, Boltzmann, and softmax policies under explicitly stated parameterizations. It establishes uniform bounds on their score functions and shows that these policies are Lipschitz continuous. The source abstract consists of this single sentence, so no further details are available.
In briefWhen Explicit Penalties Help Flow Matching
SummaryThis paper studies whether adding constraint penalties to flow matching generative models improves results, using physics-informed flow matching as a motivating example. The authors compare penalised and unpenalised models by their expected squared 2-Wasserstein error, covering vector-valued and functional data in a Hilbert-space setting. They find penalisation improves accuracy exactly when the reduction in estimation variance outweighs the increase in squared bias, a trade-off depending on sample size, model class, noise schedule, and penalty weight.
Lea Kunkel, Alberto Gonz\'alez-Sanz, Sophie Langer
In briefToward Automated Labeling for Short Video Content Governance
SummaryThis paper proposes a multi-stage framework that progressively increases automation in labeling short videos for content governance. Stages include a fine-tuned discriminative multimodal LLM labeler (L1), a domain-adaptive model used for zero-shot labeling (L2), and an agentic workflow with tools for complex or vaguely defined cases (L3). The authors report that accuracy improves at each stage and that manual labor costs in real-world applications were reduced by up to 85%.
In briefRadioactive Feedback: Tracing Models Trained on Score-Only LLM Judge Signals
SummaryThis paper studies a form of knowledge distillation in which an LLM acts as a judge that gives only scalar scores, while the student model generates all its own responses and learns from those scores alone. The authors propose Radioactive Feedback, which embeds a registered pattern of behavioral preferences into the rewards so it can carry over to a student trained by another party. Experiments on code and mathematics show the preferences transfer through scores alone across datasets, enabling black-box detection against a fixed reference pool.
In briefRethinking Human Intent-to-CAD: Generating Parametric CAD Models from Sketches and Text
SummaryThis paper addresses generating parametric CAD models directly from incomplete human intents such as hand-drawn sketches and text, without requiring existing target models. The authors build HiCAD, a large dataset aligning sketches, text descriptions, and parametric CAD code. Their two-stage method combines cooperative multi-task alignment with spatial-aware reinforcement learning to improve geometric and topological consistency. Experiments report gains over existing baselines across multiple tasks.
