{"results":[{"record_id":"knows:generated/vig-multimodal-chain-of-thought-compression/1.0.0","profile":"paper@1","title":"VIG: Visual Information Gain as a Reward Signal for Multimodal Chain-of-Thought Compression","summary":"The paper introduces Visual Information Gain, a per-token information-theoretic reward that measures how much an image reduces a multimodal reasoning model's predictive uncertainty. Combined with answer and format rewards in GRPO, VIG improves accuracy and inference efficiency by increasing visual information density rather than directly enforcing shorter chains.","venue":"arXiv preprint","year":2026,"discipline":null,"keywords":["multimodal reasoning","chain-of-thought compression","visual information gain","reinforcement learning","GRPO","conditional mutual information"],"coverage_statements":"exhaustive","coverage_evidence":"key_evidence_only","provenance_origin":"machine","provenance_actor_name":"knows-gen","version_record":"1.0.0","lint_passed":true,"download_count":0,"created_at":"2026-09-03T00:54:30.309706+00:00","stats":{"stmt_count":24,"evidence_count":19,"relation_count":86,"artifact_count":19,"claim_count":14,"method_count":3,"limitation_count":4}},{"record_id":"knows:generated/complexity-induction/1.0.0","profile":"paper@1","title":"Complexity Induction: Compositional Generalization via Structured Training Distortion","summary":"Demonstrates that structured distortion of training data — termed complexity induction — via mixed labels (soft targets from Jaccard similarity over class names) or expanded datasets with structurally motivated false samples can induce compositional generalization in a standard CNN classifier without architectural modification; the two methods act on distinct mechanisms (classifier activation vs. ","venue":"arXiv preprint","year":2026,"discipline":null,"keywords":["compositional generalization","soft labels","complexity induction","zero-shot classification","convolutional neural networks"],"coverage_statements":"exhaustive","coverage_evidence":"key_evidence_only","provenance_origin":"machine","provenance_actor_name":"knows-gen","version_record":"1.0.0","lint_passed":true,"download_count":0,"created_at":"2026-09-03T00:52:21.356562+00:00","stats":{"stmt_count":19,"evidence_count":17,"relation_count":60,"artifact_count":11,"claim_count":7,"method_count":3,"limitation_count":4}},{"record_id":"knows:generated/semantics-saturate-emerge/1.0.0","profile":"paper@1","title":"When Semantics Saturate or Emerge: Adaptation-Conditional Semantic Utility in Source-Free Cross-Domain Few-Shot Learning","summary":"An empirical and diagnostic study of source-free cross-domain few-shot learning showing that frozen zero-shot prompt ranking is not a reliable proxy for adaptation-anchor quality. The paper introduces adaptation-conditional semantic utility and identifies two reproducible regimes: semantic saturation (EuroSAT, CropDisease) where a large zero-shot advantage contracts after LoRA, and semantic emerge","venue":"arXiv preprint","year":2026,"discipline":null,"keywords":["source-free cross-domain few-shot learning","vision-language models","semantic prompting","low-rank adaptation","domain adaptation","CLIP","prompt evaluation"],"coverage_statements":"exhaustive","coverage_evidence":"key_evidence_only","provenance_origin":"machine","provenance_actor_name":"knows-gen","version_record":"2.0.0","lint_passed":true,"download_count":0,"created_at":"2026-09-03T00:52:12.235691+00:00","stats":{"stmt_count":21,"evidence_count":17,"relation_count":83,"artifact_count":16,"claim_count":10,"method_count":6,"limitation_count":1}},{"record_id":"knows:generated/execrubrics/1.0.0","profile":"paper@1","title":"ExecRubrics: Executable Tool-Augmented Rubrics for Verifiable and Efficient Long-Form Evaluation","summary":"Proposes ExecRubrics, a framework that represents evaluation rubrics as compact executable Python scoring programs rather than natural-language criteria interpreted by black-box LLM judges, demonstrating that generated executable rubrics can match or exceed NL rubric baselines on HealthBench, ArgQuality, and HelpSteer while reducing evaluation latency by orders of magnitude.","venue":"arXiv preprint","year":2026,"discipline":null,"keywords":["rubric-based evaluation","LLM-as-judge","executable rubrics","long-form evaluation","tool augmentation","preference accuracy"],"coverage_statements":"exhaustive","coverage_evidence":"key_evidence_only","provenance_origin":"machine","provenance_actor_name":"knows-gen","version_record":"1.0.0","lint_passed":true,"download_count":0,"created_at":"2026-09-03T00:51:49.456605+00:00","stats":{"stmt_count":24,"evidence_count":13,"relation_count":43,"artifact_count":12,"claim_count":8,"method_count":4,"limitation_count":5}},{"record_id":"knows:generated/ankhdjet-open-source-compiler/1.0.0","profile":"paper@1","title":"Ankhdjet: An Open-Source Compiler for Mask-Programmed Ternary Compute-in-ROM on an Open PDK","summary":"Presents Ankhdjet, an open-source compiler that lowers HuggingFace ternary checkpoints (BitNet b1.58) into via-mask programs of a fixed compute-in-ROM macro on the open SKY130 PDK, verified end-to-end with open tools (KLayout DRC zero, netgen LVS zero, clean timing) and submitted for fabrication on the TinyTapeout/ChipFoundry ttsky26c shuttle.","venue":"arXiv preprint","year":2026,"discipline":null,"keywords":["compute-in-memory","mask ROM","ternary neural networks","BitNet","open-source EDA","SKY130","hardware compilers"],"coverage_statements":"exhaustive","coverage_evidence":"key_evidence_only","provenance_origin":"machine","provenance_actor_name":"knows-gen","version_record":"1.0.0","lint_passed":true,"download_count":0,"created_at":"2026-09-03T00:51:15.021039+00:00","stats":{"stmt_count":32,"evidence_count":25,"relation_count":38,"artifact_count":13,"claim_count":11,"method_count":8,"limitation_count":5}},{"record_id":"knows:generated/same-model-different-harness/1.0.0","profile":"paper@1","title":"Same Model, Different Harness: Different Coding-Agent Results","summary":"A paired experiment on three coding benchmarks shows that changing the harness configuration (a closed-loop working-view rule plus stall detector plus command safeguards) changes what the same frozen model weights can accomplish, especially under tight context pressure; treatment raises mean per-task F2PF on all three pressured cohorts and complete solutions on SWE-bench Verified and Pro, while re","venue":"arXiv preprint","year":2026,"discipline":null,"keywords":["coding agents","harness design","context window","SWE-bench","agentic evaluation"],"coverage_statements":"key_claims_and_limitations","coverage_evidence":"key_evidence_only","provenance_origin":"machine","provenance_actor_name":"knows-gen","version_record":"1.0.0","lint_passed":true,"download_count":0,"created_at":"2026-09-03T00:51:12.401577+00:00","stats":{"stmt_count":24,"evidence_count":20,"relation_count":71,"artifact_count":11,"claim_count":10,"method_count":4,"limitation_count":4}},{"record_id":"knows:generated/stereo-reranker/1.0.0","profile":"paper@1","title":"A Reranker for Orchestrating Heterogeneous Speech and Text Retrievers","summary":"STeReO is a cross-modal reranker that aggregates heterogeneous speech and text retrieval candidates for Retrieval-Augmented Generation, using a Z-score normalized fusion for training data construction and a LoRA-fine-tuned audio language model supporting pointwise, pairwise, and listwise objectives.","venue":"arXiv preprint","year":2026,"discipline":null,"keywords":["multimodal reranking","heterogeneous retrievers","speech retrieval","retrieval-augmented generation","cross-modal"],"coverage_statements":"exhaustive","coverage_evidence":"key_evidence_only","provenance_origin":"machine","provenance_actor_name":"knows-gen","version_record":"1.0.0","lint_passed":true,"download_count":0,"created_at":"2026-09-03T00:50:48.782182+00:00","stats":{"stmt_count":18,"evidence_count":14,"relation_count":34,"artifact_count":14,"claim_count":8,"method_count":3,"limitation_count":3}},{"record_id":"knows:generated/hamdani-2026-sms-ai-software-quality/1.0.0","profile":"paper@1","title":"Challenges and Contributions in Quality of AI-Based Software: A Systematic Mapping Study","summary":"A systematic mapping study synthesizing 33 primary studies (2020-2026) on the quality of AI-based software, identifying six recurring challenge categories and eleven contribution categories, with limitations in existing quality assessment models (e.g., ISO/IEC 25059) being the most prominent challenge.","venue":"arXiv preprint","year":2026,"discipline":null,"keywords":["AI-based software","software quality","systematic mapping study","quality assurance","machine learning systems","ISO/IEC 25059"],"coverage_statements":"exhaustive","coverage_evidence":"key_evidence_only","provenance_origin":"machine","provenance_actor_name":"knows-gen","version_record":"1.0.0","lint_passed":true,"download_count":0,"created_at":"2026-09-03T00:50:23.761010+00:00","stats":{"stmt_count":16,"evidence_count":13,"relation_count":31,"artifact_count":5,"claim_count":9,"method_count":3,"limitation_count":2}},{"record_id":"knows:generated/exam2-multilingual-multimodal-audio/1.0.0","profile":"paper@1","title":"EXAM^2: Extending Audio Understanding in Multilingual and Multimodal Analysis","summary":"Introduces EXAM^2, a benchmark for multilingual and multimodal audio understanding spanning six languages (EN, DE, ES, JA, MS, ZH) and four audio domains (speech, sound, music, mix), with 5,667 MCQs, 22,614 image instances, and 135,684 multilingual translations, plus a lightweight Gemma3n-EXAM^2 fusion model and OmniLoRA method that yield up to 12.4% multilingual and 21.7% multimodal gains over ba","venue":"arXiv preprint","year":2026,"discipline":null,"keywords":["multilingual audio understanding","multimodal audio benchmark","large audio language models","LoRA fine-tuning","multiple choice question answering","cross-modal reasoning"],"coverage_statements":"exhaustive","coverage_evidence":"key_evidence_only","provenance_origin":"machine","provenance_actor_name":"knows-gen","version_record":"1.0.0","lint_passed":true,"download_count":0,"created_at":"2026-09-03T00:49:37.984643+00:00","stats":{"stmt_count":23,"evidence_count":14,"relation_count":47,"artifact_count":15,"claim_count":14,"method_count":3,"limitation_count":3}},{"record_id":"knows:generated/4dstreamctrl/1.0.0","profile":"paper@1","title":"4DStreamCtrl: Interactive Video Generation with Online 4D Control","summary":"4DStreamCtrl unifies camera motion, object trajectories, and depth into a single 3D point-track conditioning interface for video diffusion, trained on a new in-the-wild 3D motion dataset (OpenVidHD-Motion3D) and distilled into a causal streaming student that runs at 20 FPS on a single GPU with memory independent of video length.","venue":"arXiv preprint","year":2026,"discipline":null,"keywords":["controllable video generation","4D control","video diffusion","streaming generation","3D point tracks","causal distillation"],"coverage_statements":"exhaustive","coverage_evidence":"key_evidence_only","provenance_origin":"machine","provenance_actor_name":"knows-gen","version_record":"1.0.0","lint_passed":true,"download_count":0,"created_at":"2026-09-03T00:49:37.984611+00:00","stats":{"stmt_count":20,"evidence_count":16,"relation_count":50,"artifact_count":12,"claim_count":9,"method_count":4,"limitation_count":4}},{"record_id":"knows:generated/ucag-p-camera-centric-embodied/1.0.0","profile":"paper@1","title":"One Policy, Many Embodiments: Unified Camera-Centric Action Geometry Pre-training for Heterogeneous Embodied Manipulation","summary":"UCAG-P introduces a camera-centric unified action formulation that aligns heterogeneous embodied data (robot, simulation, human) into a shared geometric action space via wrist/end-effector and grasp-center anchor motion, enabling a single VLA policy checkpoint to be competitive across LIBERO, RoboTwin, RoboCasa GR-1, and real-world Piper platforms without benchmark-specific fine-tuning.","venue":"arXiv preprint","year":2026,"discipline":null,"keywords":["vision-language-action","cross-embodiment learning","human demonstrations","robot manipulation","camera-centric action space","unified policy"],"coverage_statements":"exhaustive","coverage_evidence":"key_evidence_only","provenance_origin":"machine","provenance_actor_name":"knows-gen","version_record":"1.0.0","lint_passed":true,"download_count":0,"created_at":"2026-09-03T00:49:37.984598+00:00","stats":{"stmt_count":27,"evidence_count":21,"relation_count":56,"artifact_count":13,"claim_count":11,"method_count":7,"limitation_count":4}},{"record_id":"knows:generated/pre-training-visual-dexterity-simulation/1.0.0","profile":"paper@1","title":"Pre-training Visual Dexterity in Simulation","summary":"We introduce SPD (Simulation Pre-training for Dexterity), a framework that uses VR teleoperation in a physics simulator to collect 75 hours of multi-task bimanual dexterous manipulation data, then pre-trains a 222M-parameter diffusion transformer policy that significantly improves real-world fine-tuning on five bimanual dexterous tasks compared to behavior cloning from scratch.","venue":"arXiv preprint","year":2026,"discipline":null,"keywords":["dexterous manipulation","simulation pre-training","VR teleoperation","diffusion transformer","behavior cloning","bimanual robots"],"coverage_statements":"exhaustive","coverage_evidence":"key_evidence_only","provenance_origin":"machine","provenance_actor_name":"knows-gen","version_record":"2.0.0","lint_passed":true,"download_count":0,"created_at":"2026-09-03T00:49:37.984588+00:00","stats":{"stmt_count":16,"evidence_count":16,"relation_count":47,"artifact_count":8,"claim_count":5,"method_count":6,"limitation_count":2}},{"record_id":"knows:generated/univvt/1.0.0","profile":"paper@1","title":"UniVVT: A Unified End-to-End Framework for High-Fidelity Video Virtual Try-on","summary":"UniVVT is a unified end-to-end framework that reformulates video virtual try-on as semantically conditioned video generation, eliminating the need for masks, pose estimation, and garment warping at inference by using an MLLM-based Scene-Task Perceiver and a lightweight Semantic Bridge to provide implicit task-aware guidance to a diffusion-based video generator.","venue":"arXiv preprint","year":2026,"discipline":null,"keywords":["video virtual try-on","multimodal large language model","semantic conditioning","diffusion transformer","implicit guidance","end-to-end generation"],"coverage_statements":"exhaustive","coverage_evidence":"key_evidence_only","provenance_origin":"machine","provenance_actor_name":"knows-gen","version_record":"2.0.0","lint_passed":true,"download_count":0,"created_at":"2026-09-03T00:49:37.984572+00:00","stats":{"stmt_count":26,"evidence_count":21,"relation_count":45,"artifact_count":7,"claim_count":7,"method_count":9,"limitation_count":3}},{"record_id":"knows:generated/legaria-santiago-2026-remote-traffic-air-pollution/1.0.0","profile":"paper@1","title":"Leveraging Remote Traffic Data for Local Air Pollutant Estimation: A Scenario-Based Machine Learning Study Across London Monitoring Sites","summary":"This study evaluates four tree-based machine learning models (Random Forest, Extra Trees, LightGBM, XGBoost) under six predictor scenarios combining remote traffic, meteorological, temporal variables, and measurements from neighbouring monitoring stations to estimate NO2, PM10, PM2.5, and O3 concentrations across London traffic-type monitoring sites, finding that traffic-related variables can cont","venue":"arXiv preprint","year":2026,"discipline":null,"keywords":["air pollution estimation","machine learning","traffic data","NO2","PM2.5","PM10","ozone","London","SHAP analysis"],"coverage_statements":"exhaustive","coverage_evidence":"key_evidence_only","provenance_origin":"machine","provenance_actor_name":"knows-gen","version_record":"1.0.0","lint_passed":true,"download_count":0,"created_at":"2026-09-03T00:49:37.984567+00:00","stats":{"stmt_count":21,"evidence_count":18,"relation_count":48,"artifact_count":8,"claim_count":8,"method_count":4,"limitation_count":4}},{"record_id":"knows:generated/eduriskx/1.0.0","profile":"paper@1","title":"EduRiskX: A Neuro-Symbolic Framework with F-Logic Reasoning for Early Academic Risk Prediction","summary":"EduRiskX is a neuro-symbolic framework that integrates a temporal Transformer-based predictor with F-Logic symbolic reasoning, achieving accuracy 0.900, F1-score 0.894, and an average early detection week of 9.32 on the OULAD dataset through logistic-regression-based fusion of neural and rule-based signals.","venue":"arXiv preprint","year":2026,"discipline":null,"keywords":["neuro-symbolic AI","F-Logic","academic risk prediction","learning analytics","temporal Transformer","educational data mining","OULAD"],"coverage_statements":"exhaustive","coverage_evidence":"key_evidence_only","provenance_origin":"machine","provenance_actor_name":"knows-gen","version_record":"1.0.0","lint_passed":true,"download_count":0,"created_at":"2026-09-03T00:49:36.925372+00:00","stats":{"stmt_count":23,"evidence_count":17,"relation_count":66,"artifact_count":8,"claim_count":10,"method_count":6,"limitation_count":2}},{"record_id":"knows:generated/behaviorworldgen/1.0.0","profile":"paper@1","title":"BehaviorWorldGen: Closing the Loop between Action Models and World Simulators via Controllable Behavior-Aware Structured World Generation","summary":"BehaviorWorldGen is a framework that closes the loop between action models and world simulators through controllable behavior-aware structured world generation. Its core BehaviorFlow module injects interpretable meta-action controls over agent lifecycles and behaviors, producing interaction-consistent multi-agent rollouts that are rendered into realistic observations and used to refine diverse act","venue":"arXiv preprint","year":2026,"discipline":null,"keywords":["autonomous driving","world model","traffic simulation","action model","self-improvement loop","BehaviorFlow","meta-action"],"coverage_statements":"exhaustive","coverage_evidence":"key_evidence_only","provenance_origin":"machine","provenance_actor_name":"knows-gen","version_record":"1.0.0","lint_passed":true,"download_count":0,"created_at":"2026-09-03T00:49:36.794894+00:00","stats":{"stmt_count":26,"evidence_count":17,"relation_count":58,"artifact_count":8,"claim_count":10,"method_count":8,"limitation_count":3}},{"record_id":"knows:generated/trace-retrospective-streaming/1.0.0","profile":"paper@1","title":"TRACE: Retrospective Streaming Generation of Physical Fields under Sparse Structured Sensing","summary":"TRACE is a retrospective streaming generative framework that reconstructs continuous physical fields from temporally sparse and spatially localized streaming observations by fusing per-frame diffusion-posterior evidence with a Matérn state-space Gaussian process via Kalman filtering and Rauch–Tung–Striebel smoothing in a learned continuous-coordinate latent space.","venue":"arXiv preprint","year":2026,"discipline":null,"keywords":["physical field reconstruction","generative reconstruction","structured sensing","Kalman filtering","RTS smoothing","state-space Gaussian process","diffusion posterior sampling","functional Tucker decomposition"],"coverage_statements":"exhaustive","coverage_evidence":"key_evidence_only","provenance_origin":"machine","provenance_actor_name":"knows-gen","version_record":"1.0.0","lint_passed":true,"download_count":0,"created_at":"2026-09-03T00:49:36.670621+00:00","stats":{"stmt_count":31,"evidence_count":26,"relation_count":61,"artifact_count":13,"claim_count":14,"method_count":6,"limitation_count":3}},{"record_id":"knows:generated/tailorcopilot-uist26/1.0.0","profile":"paper@1","title":"TailorCoPilot: Enabling Agentic Pattern Making with Version-Controlled State Tracking","summary":"We present TailorCoPilot, an agentic pattern-making system built on a version-controlled backend (TailorTrace) that models sewing patterns as discrete validated states connected by explicit geometric operations, capturing tacit expert knowledge as reusable traces to support novice apprenticeship and downstream AI training.","venue":"arXiv preprint","year":2026,"discipline":null,"keywords":["Garment Pattern Making","Expert Process Knowledge","Version Control","Human-AI Collaboration","Traceable Workflows"],"coverage_statements":"exhaustive","coverage_evidence":"key_evidence_only","provenance_origin":"machine","provenance_actor_name":"knows-gen","version_record":"1.0.0","lint_passed":true,"download_count":0,"created_at":"2026-09-03T00:49:34.923082+00:00","stats":{"stmt_count":23,"evidence_count":20,"relation_count":51,"artifact_count":11,"claim_count":8,"method_count":6,"limitation_count":4}},{"record_id":"knows:generated/biomimetic-joint-covering-skin-proprioception/1.0.0","profile":"paper@1","title":"Design of a Biomimetic Joint-Covering Skin with Tissue-Like Structure to Enhance Proprioception in a Musculoskeletal Humanoid","summary":"Designs a biomimetic joint-covering skin with a three-layer tissue-like structure (epidermis, dermis, subcutaneous tissue) embedding strain gauge and conductive filament receptor-like sensors, implemented on the musculoskeletal humanoid Musashi-W, and demonstrates that the skin alone can estimate elbow joint angle with ~3 degrees error and that encoder-based fusion with muscle sensing further impr","venue":"arXiv preprint","year":2026,"discipline":null,"keywords":["biomimetic skin","proprioception","musculoskeletal humanoid","multimodal sensing","tactile sensors","tissue-structured design"],"coverage_statements":"exhaustive","coverage_evidence":"key_evidence_only","provenance_origin":"machine","provenance_actor_name":"knows-gen","version_record":"1.0.0","lint_passed":true,"download_count":0,"created_at":"2026-09-03T00:49:34.923079+00:00","stats":{"stmt_count":23,"evidence_count":16,"relation_count":49,"artifact_count":6,"claim_count":8,"method_count":5,"limitation_count":4}},{"record_id":"knows:generated/katok-adaptive-tokenizer/1.0.0","profile":"paper@1","title":"Keep-or-Drop? Adaptive Tokenizer for Compact Video Representation","summary":"KATok is an adaptive transformer-based VAE for video that learns per-token keep-or-drop probabilities via Gumbel-Softmax, discarding uninformative tokens to enable content-adaptive compression, and pairs the sparse tokens with cascaded or joint content-position diffusion for state-of-the-art generation quality at substantially fewer tokens than dense tokenizers.","venue":"arXiv preprint","year":2026,"discipline":null,"keywords":["VAE","Adaptive Tokenization","Video Diffusion Model","Latent Diffusion","Content-Adaptive Compression"],"coverage_statements":"exhaustive","coverage_evidence":"key_evidence_only","provenance_origin":"machine","provenance_actor_name":"knows-gen","version_record":"1.0.0","lint_passed":true,"download_count":0,"created_at":"2026-09-03T00:49:34.923079+00:00","stats":{"stmt_count":20,"evidence_count":17,"relation_count":40,"artifact_count":18,"claim_count":10,"method_count":5,"limitation_count":3}}],"total":338885,"next_cursor":"20"}