BENCHMARK RADAR / DISCOVERY DIGEST · 2026-10-30
127 first-seen evaluation-related records, organized by domain and type, with links to their original sources.
AI-assisted brief · not independently reviewed
Type counts reported by the original digest: New benchmarks 66 · New datasets 17 · Evaluation reports 13 · Method papers 20 · Other records 11
These records first appeared in this saved collection and were deduplicated using the project's exact artifact identities. First seen is not a release date, and not every record introduces an AI benchmark.
This page preserves the 127 first-seen records from the 2026-10-30 publication, using the collection history saved at that time. Later additions to historical sources do not rewrite this published brief.
Classification and brief model: deepseek-v4-flash
Categories and Chinese briefs come from the original automated digest. English briefs are translations, with individual corrections noted. Check the original material through each link. Briefs do not become model-score observations in the catalog.
DataCite
What it covers: A benchmark of LLM code reviewers detecting cryptographic API misuse.
Reported finding: Includes 54 Python snippets, 9 misuse classes and 1,890 trials across 7 models.
DataCite
What it covers: A reproducible benchmark of adaptive indirect prompt injection in multi-document LLM screening.
Reported finding: Validates the pipeline with a synthetic responder; no real models were measured.
DataCite
What it covers: A cyber-physical benchmark dataset for intrusion detection in UAV swarms.
Reported finding: Contains 144 simulations, 6 classes (normal traffic and 5 attack types), 2,427 windows of 30 seconds and 74 features.
Brief correction: Clarified against the saved source abstract.
Crossref
What it covers: A reliable Android malware benchmark dataset derived from technical expert reports.
Reported finding: No abstract supplied by the source.
Kaggle Dataset
What it covers: An AI governance benchmark distinguishing operational memory retention from authored signal.
Reported finding: No results reported in the abstract.
Hugging Face
What it covers: A traffic-to-client-code benchmark with hidden execution tests.
Reported finding: No results reported in the abstract.
Hugging Face
What it covers: A benchmark of synthetic SQL queries.
Reported finding: No abstract supplied by the source.
Kaggle Dataset
What it covers: A patch-audit benchmark for Gemma 4 checkers, with 552 labeled patches.
Reported finding: No results reported in the abstract.
Hugging Face
What it covers: A French clinical and radiological speech-recognition benchmark covering 9 imaging modalities and SNOMED CT prescriptions.
Reported finding: WannaMedASR v2 reports 0.48% CER, 99.52% precision and 0.82% WER.
Hugging Face
What it covers: A dental implant brand classification benchmark with 3276 images and patient-level 5-fold splits.
Reported finding: No results reported in the abstract.
DataCite
What it covers: A benchmark of Japanese LLM virtual-patient recovery after contamination, including human evaluation.
Reported finding: No results reported in the abstract.
Hugging Face
What it covers: The Tissues-with-Issues benchmark dataset for issue detection in tissue images.
Reported finding: No abstract supplied by the source.
Crossref
What it covers: A public kidney CT benchmark measuring data leakage without patient identifiers.
Reported finding: No abstract supplied by the source.
Crossref
What it covers: Evaluates general-purpose LLMs for radiology-report quality assurance across imaging modalities and real-world clinical centers.
Reported finding: No results reported in the abstract.
Crossref
What it covers: Evaluates how LLM medical guidance adapts to different emotional sensitivities and literacy levels.
Reported finding: Literacy adaptation is accurate with near-zero error; emotional adaptation varies more with model architecture and prompt design.
Crossref
What it covers: Evaluates 10 LLMs on complex decisions for colorectal cancer with peritoneal metastases.
Reported finding: Full concordance is 56.4% across 1080 primary responses, with no robust between-model difference; semantic adjudication reduces it to 45.5%.
Crossref
What it covers: Evaluates LLM thyroid-nodule biopsy recommendations under clinical equipoise.
Reported finding: LLM biopsy rates range from 46%–74% (lowest for Claude 3.5, highest for Llama 3.2-90B), with moderate clinician agreement (α=0.372–0.453).
Crossref
What it covers: An end-to-end audit of an oral-health screening prototype combining YOLOv8m and other models.
Reported finding: Training near-duplicates account for 17.3% of test images; removing them changes mAP by only +0.001. The router rejects 23% of images, ulcer sensitivity is 0.36, and a label-mapping defect misnames 73% of findings.
Crossref
What it covers: Evaluates 15 LLMs’ ability to detect methodological errors in randomized clinical trials.
Reported finding: Major-error detection scores on 2026 trials range from 58%–99%. Opus 5.5 leads at 99%; GPT-6.1 Sol scores 94% and GPT-6 Luna scores 86%.
Crossref
What it covers: Evaluates LLM robustness, from ModernBert to LLaMA, for depression detection in social-media text.
Reported finding: No abstract supplied by the source.
DataCite
What it covers: A telemetry benchmark for multi-agent file-sharing workflows through the LimeWire MCP server.
Reported finding: No results reported in the abstract.
GitHub
What it covers: An end-to-end Elixir/OTP CLI engineering benchmark for coding agents and LLMs.
Reported finding: No results reported in the abstract.
DataCite
What it covers: A dataset of MCP multi-agent file-sharing telemetry.
Reported finding: Contains 180 empirical records covering Claude, ChatGPT, Cursor, VS Code and Gemini CLI.
DataCite
What it covers: A benchmark suite evaluating how frontier LLMs and agent runtimes use governed registry skills.
Reported finding: No results reported in the abstract.
GitHub
What it covers: A benchmark of agents completing tasks in a simulated software company.
Reported finding: No results reported in the abstract.
GitHub
What it covers: A repository collecting game-agent evaluation benchmarks.
Reported finding: No abstract supplied by the source.
Hugging Face Papers
What it covers: A benchmark of computer-use agents serving users with diverse persona requirements.
Reported finding: Across 44 personas, mean task completion falls by 36%. Even fully completed requests earn the top satisfaction rating in only 62.7% of cases, versus 84.1% for original requests.
Hugging Face Papers
What it covers: A benchmark jointly evaluating CLI-agent task execution and human-agent collaboration quality.
Reported finding: The strongest configuration averages only 49.2%; the joint collaboration score peaks at 37.6%, showing a substantial gap between task effectiveness and collaboration quality.
Hugging Face Papers
What it covers: An action-oriented benchmark of conversational agents following banking procedures.
Reported finding: Contains 1,000 Indian retail-banking tasks. Checking final records alone incorrectly marks 6.5%–12.0% of sessions as successful.
Hugging Face
No usable brief was supplied. Open the original source for details.
Hugging Face Papers
What it covers: WorkFactory constructs executable professional tasks through information support graphs for scalable agent training.
Reported finding: Builds 112,686 tasks across 612 occupations; WorkPilot-35B-A3B improves from 879 to 1565 Elo on GDPval-AA v2.1.
Hugging Face
What it covers: Raw human evaluation data for Tech-Graphics-Benchmark, with one JSON file per generation.
Reported finding: No results reported in the abstract.
DataCite
What it covers: A LimeWire benchmark dataset of generative AI and 3D reconstruction models.
Reported finding: Contains 350 empirical records evaluating 7 generative AI and 3D reconstruction models across 50 prompts.
Hugging Face
What it covers: A pilot dataset for fine-grained diagnosis of generated-video failures, with 33 failure subclasses across 8 families.
Reported finding: Its 33 failure subclasses and 8 families support failure discovery, subclass classification and temporal diagnosis.
Hugging Face
What it covers: Test prompts for Tech-Graphics-Benchmark, evaluating image models’ ability to generate technical graphics.
Reported finding: No results reported in the abstract.
Hugging Face
What it covers: A benchmark dataset named MiRAGeNews-Splicing-Benchmark.
Reported finding: No abstract supplied by the source.
Hugging Face Papers
What it covers: IntReal is a real-world dataset and benchmark for multimodal interleaved generation.
Reported finding: IntReal-100K covers four sample categories. Evaluation agrees with human judgments at 83.3%, inter-metric ρ≤0.42, and four UMM families show distinct strengths and weaknesses.
Hugging Face Papers
What it covers: EditAgent expands the action and observation spaces of agentic image editing using verification records.
Reported finding: A general editor produces no usable result in over 40% of cases. EditAgent enables off-the-shelf VLM judges to outperform fine-tuned reward models; experts prefer its parallel, listwise approach to sequential refinement in 78.5% of decided comparisons. It improves two open-weight editors on three benchmarks.
Hugging Face
What it covers: Corrected videos and tables for the mirohand-egohand-umetrack benchmark.
Reported finding: Includes 54 mesh-overlay videos and complete ARCTIC/H2O/EgoRenderer tables; EgoHand is converted from SOMA to MANO.
Hugging Face
What it covers: Generated images and manifest files for each round of Tech-Graphics-Benchmark.
Reported finding: Images are organized by round; shuffled manifest rows and opaque generation IDs preserve blind evaluation.
Hugging Face
What it covers: A benchmark of LAION CLAP fused and unfused checkpoints for audio-text retrieval.
Reported finding: Under the short-clip protocol, AudioCaps differences are inconclusive. On Clotho, unfused scores higher in both directions, with 95% intervals excluding zero; equivalence is not established.
Crossref
What it covers: Defines correct-option surface profiles and audits composition sensitivity in industrial MLLM evaluation.
Reported finding: High-Low risk differences are 5.20–5.90 percentage points, with only 19/90 cells strictly ordered. Composition shifts standardized accuracy by -0.46 to +0.48 points and causes four local ranking swaps; Top-3 membership stays stable.
DataCite
What it covers: A benchmark dataset for RNAflash.
Reported finding: No abstract supplied by the source.
DataCite
What it covers: SomBench is a lunar-science machine-learning benchmark dataset with 30+ layers and a task suite.
Reported finding: ResNet-50 and SwinV2-B baselines show that each benchmark task is learnable from the released inputs, establishing reference points.
Crossref
What it covers: A benchmark of 18 co-occurrence network inference methods on synthetic 16S rDNA-seq data.
Reported finding: Methods work only in restricted, often unrealistic settings, with performance declining under more realistic parameters.
DataCite
What it covers: A de novo peptide sequencing benchmark based on MassIVE MSV000103779.
Reported finding: No abstract supplied by the source.
DataCite
What it covers: GEMS-GER is a groundwater machine-learning benchmark dataset with water levels, meteorological forcings and static features.
Reported finding: Includes weekly data for 3,207 German monitoring wells from 1991–2022, with NSE/RMSE and other results for single-well, global dynamic and dynamic-plus-static models.
Crossref
What it covers: AMBER is an AI-ready downhole microseismic benchmark database with waveforms and manually verified phase-pick labels.
Reported finding: Contains over 4,000 events, 150,000 P/S wave picks and 10,000 signal-free noise examples, all manually verified.
Crossref
What it covers: Compares conceptual hydrological models with deep learning and proposes benchmark-informed model selection based on catchment attributes.
Reported finding: PB models average 0.82 modified KGE across 70 catchments; LSTM averages 0.68 across 82 catchments.
Crossref
What it covers: Proposes an optimizer-equivalence evaluation protocol for tuning deep-learning models of global solar irradiance.
Reported finding: With a 10-epoch objective, optimizers do not differ significantly; optimizer choice explains 0.8% of variance versus 7.9% for seeds. At 50 epochs, two guided methods improve on random search by 0.74 W/m².
Crossref
What it covers: A decision-centric benchmark of enrichment and potency ranking in CNN-rescored molecular docking.
Reported finding: Potency ranking is negligible (ρ=0.020). Enrichment depends on the benchmark setup; a structure-free classifier reaches ROC-AUC 0.959–0.993.
Hugging Face
What it covers: The OpenNLI leaderboard’s 44 public test sets and evaluation scripts for reproducing scores.
Reported finding: No results reported in the abstract.
Crossref
What it covers: A Chinese text-length-control benchmark with 5 topics × 4 word-count levels, introducing the LAPP post-processing framework.
Reported finding: DeepSeek-V3, Qwen2.5 and Kimi have initial deviation rates of 11.93%–38.70%. LAPP reduces these to about 5%, raises pass rates to 85%–100% and yields a BERTScore of 0.97.
Hugging Face
What it covers: An NLI benchmark leaderboard reporting accuracy, calibration and speed.
Reported finding: No results reported in the abstract.
Kaggle Dataset
What it covers: A dataset of 100 architectural QTO cases for a multilingual LLM anti-hallucination benchmark.
Reported finding: No results reported in the abstract.
Hugging Face
What it covers: A machine-readable mirror of LLM benchmark and local GPU evaluation data.
Reported finding: Covers 100 global models, 56 benchmarks, 2540 table cells, 5069 measurement variants and 24 local GPU variants.
Hugging Face
What it covers: Audit logs and scorecards for the Sovereign G4 Hermetic benchmark.
Reported finding: Includes interactive evaluation data for RouterBench and Aider Polyglot.
DataCite
What it covers: A benchmark-score snapshot and reproduction package for Mind the Gap.
Reported finding: Includes reproduction data for 520 models, 359 benchmark versions, 6180 measured scores and 14314 estimates.
Crossref
What it covers: Introduces a tiered evaluation framework for structured scientific evidence extraction and benchmarks multiple LLMs.
Reported finding: Models perform substantially better on low-complexity extraction. Scaling yields uneven gains and does not eliminate the weaknesses.
Hugging Face
What it covers: A Tibetan OCR benchmark with 472 page images, manual transcriptions and annotations.
Reported finding: Includes 472 transcribed page images and per-page metadata for evaluating OCR systems.
Brief correction: Clarified against the saved source abstract.
GitHub
What it covers: A proof-carrying geometric-art and model-evaluation benchmark with replayable constraint checks and an OpenEnv reinforcement-learning environment.
Reported finding: No results reported in the abstract.
Crossref
What it covers: A multi-source benchmark dataset evaluating nine deep-learning models for cropland parcel boundary extraction.
Reported finding: Spatial resolution has a nonlinear effect. BsiNet reaches 86.58% F1 in plains/hills but falls to 36.56% in mountains, where RCF leads at 70.85%.
GitHub Organization
What it covers: Benchmark data evaluating models’ ability to predict an object’s hidden sixth view.
Reported finding: No results reported in the abstract.
Kaggle Dataset
What it covers: A dataset named CW-SSN AiOIR Clean Benchmarks.
Reported finding: No abstract supplied by the source.
Hugging Face
What it covers: A benchmark testing whether VLM spatial understanding transfers to VLA behavior, with three components built on BEHAVIOR-1K/OmniGibson.
Reported finding: No results reported in the abstract.
Hugging Face
What it covers: A robot manipulation benchmark for factory packing, with teleoperation demonstrations for developing and evaluating manipulation policies.
Reported finding: No results reported in the abstract.
GitHub
What it covers: A public RoboDojo agent benchmark atlas.
Reported finding: No results reported in the abstract.
Crossref
What it covers: A multimodal benchmark dataset for driver behavior analysis in Pakistan’s urban and highway traffic.
Reported finding: No abstract supplied by the source.
Hugging Face
What it covers: A GR1 six-task mini leaderboard.
Reported finding: No abstract supplied by the source.
Hugging Face
What it covers: A continuous-dynamics and stiff-ODE benchmark for LLMs.
Reported finding: No results reported in the abstract.
Hugging Face
What it covers: A benchmark of 280 mathematical problem pairs studying how reasoning models describe reasoning traces.
Reported finding: Each pair changes one numerical value, preserving the solution method but changing the answer, enabling controlled reasoning-trace transfer experiments.
Hugging Face
What it covers: A speech-to-text leaderboard for Indian languages and Hinglish.
Reported finding: No results reported in the abstract.
Brief correction: Clarified against the saved source abstract.
Hugging Face
What it covers: A speech-to-text leaderboard for Indian languages and Hinglish.
Reported finding: No results reported in the abstract.
DataCite
What it covers: A benchmark dataset suite for evaluating transferable recommendation.
Reported finding: No abstract supplied by the source.
DataCite
What it covers: A benchmark suite of lithium-ion battery storage operational time series and semantic reports.
Reported finding: Covers 2023-12-01 to 2024-05-31 at one-minute intervals, with three downstream tasks including SOC prediction.
DataCite
What it covers: A source-verified benchmark dataset of fully grouted concrete block masonry prisms.
Reported finding: Contains 159 benchmark groups from 32 test programmes and at least 870 prisms, with data graded A/B.
DataCite
What it covers: A benchmark dataset for train unit shunting with service scheduling and workforce assignment.
Reported finding: Includes the paper’s instances and two annexes: instance-generation details and complete classical-solver results.
Crossref
What it covers: A systematic review and experimental benchmark of deep learning for water-network leakage detection.
Reported finding: Reviews 244 records. Time-transformer reaches the highest detection accuracy at 96.1%; the 1D-CNN autoencoder offers the best practical trade-off at 92.6%.
Crossref
What it covers: A vibration-based structural anomaly detection benchmark dataset with a multiscale pattern-learning method.
Reported finding: The proposed TimeMixer-OCSVM method achieves higher anomaly-detection accuracy than other typical methods.
Hugging Face
What it covers: A benchmark dataset named observatory-relational-recovery.
Reported finding: No abstract supplied by the source.
Hugging Face
What it covers: A poker ELO leaderboard dataset.
Reported finding: No abstract supplied by the source.
GitHub
What it covers: A benchmark of open-weight LLMs running through Ollama on an 8GB laptop GPU.
Reported finding: No results reported in the abstract.
Hugging Face
What it covers: The Slayer1B benchmark suite: 50 distinct benchmarks, one protocol and contamination checks.
Reported finding: No results reported in the abstract.
Crossref
What it covers: Introduces RAC-qRON ranking and evaluates it on an industrial-process benchmark with 21 records and 4 criteria.
Reported finding: At q=3, day 21 ranks first (closeness 0.6885). Under 10% random masking, RAC-qRON’s Kendall correlation is 0.750 versus 0.806 for TOPSIS, showing no superior missing-data stability.
Hugging Face
What it covers: An Indonesian game top-up benchmark dataset.
Reported finding: No abstract supplied by the source.
Kaggle Dataset
What it covers: A benchmark of LLM token-generation speed per second on edge boards.
Reported finding: No results reported in the abstract.
DataCite
What it covers: Solution files and full records for improved upper bounds on 17 IHTC-2024 benchmark instances.
Reported finding: All 17 solutions improve the best-known IHTC-2024 upper bounds with zero hard-constraint violations; Claude Opus 5.5 generated each solver in a 1-hour session without human-written code.
DataCite
What it covers: An open benchmark dataset for transaction latency, throughput and fulfillment reliability in Indonesian game top-ups.
Reported finding: No results reported in the abstract.
Hugging Face
What it covers: A benchmark dataset covering 213 AI affiliate programs, including commissions, payout duration and cookie length.
Reported finding: Contains 213 programs, 206 active; active programs have a median commission of 30%.
Hugging Face
What it covers: Operational metrics for water utilities in Berlin, Paris, Singapore and Rio’s CEDAE, plus a cross-utility comparison matrix.
Reported finding: No results reported in the abstract.
Hugging Face
What it covers: A benchmark of how often AI search tools mention local service providers, covering 20 metropolitan areas and 13 service categories.
Reported finding: No results reported in the abstract.
Hugging Face
What it covers: A benchmark of brand mentions in AI answer engines’ responses to B2B buying questions, covering 94 brands, 50 questions and 5 platforms.
Reported finding: No results reported in the abstract.
Hugging Face
What it covers: A local AI server performance benchmark with 273 data rows, including self-measurements and clearly marked external figures.
Reported finding: No results reported in the abstract.
Brief correction: Clarified against the saved source abstract.
Hugging Face
What it covers: A job-posting salary benchmark covering 55 countries and 1675 cells.
Reported finding: No results reported in the abstract.
Kaggle Dataset
What it covers: A GPU rental-price index dataset with 22 models across 26 cloud platforms.
Reported finding: No results reported in the abstract.
Kaggle Dataset
What it covers: A benchmark dataset of AI GPU/CPU specifications, performance, power consumption and prices for 2024-2026.
Reported finding: No results reported in the abstract.
The original digest only reported aggregate counts for method papers and other records. Individual labels were not supplied; every source link is kept below.
Hugging Face
Crossref
Crossref
DataCite
Crossref
Crossref
Crossref
DataCite
GitHub
GitHub
Crossref
Crossref
Crossref
Crossref
Crossref
Crossref
Crossref
Crossref
GitHub Organization
GitHub Organization
Crossref
Crossref
Crossref
Hugging Face Papers
Hugging Face Papers
Hugging Face Papers
Hugging Face Papers
Hugging Face Papers
Hugging Face Papers
Hugging Face Papers
Hugging Face Papers