Infoblock alpha

Public evidence graph

Facts stay small. Connections stay visible.

Atomic, primary-sourced findings linked as a graph. Refuted results remain on record, and every public omission is declared.

Blocks
212
Edges
23
Domains
22
Refuted kept
1

Evidence register

Public blocks

212 visible

E3 cond-mat/superconductivity Active In rhombohedral hexalayer graphene a field-induced superconducting state SCH emerges above 0.8 T and persists up to 1.6 T. B_on_T 0.8 · B_off_T 1.6 Zhenqi Hua et al. · arXiv:2607.06520v1 0 links
E3 cs-ai/agents Active Using Echoverse worlds as reinforcement-learning environments with a reward combining the grounded verifier and a dense per-step judge raises held-out score from 58.8% to 68.0%. held_out_score_before_pct 58.8 · held_out_score_after_pct 68.0 Yash Pandya et al. · arXiv:2607.28074v1 0 links
E3 cs-ai/agents Active On MLE-Bench Lite under a 12-hour per-task budget on a single RTX 4090 capped at 12 GB VRAM, Frontis-MA1 (35B) with OpenMLE-Evo improved Medal Average from 39.39% (base model) to 60.61%. model_params_b 35 · medal_average_base_pct 39.39 · medal_average_evo_pct 60.61 · budget_hours_per_task 12 · gpu_vram_gb 12 Junlin Yang et al. · arXiv:2607.28568v1 0 links
E3 cs-ai/agents Active On MLE-Bench Lite (12-hour per-task budget, single RTX 4090 at 12 GB VRAM), Frontis-MA1 (35B) with OpenMLE-Evo-Max reached 71.21% Medal Average, exceeding GPT-5.5 + Codex. model_params_b 35 · medal_average_evo_max_pct 71.21 Junlin Yang et al. · arXiv:2607.28568v1 0 links
E3 cs-ai/agents Active On held-out NatureBench Lite with the model held fixed, swapping in OpenMLE-Evo raised Match-SOTA from 20% to 50%. match_sota_base_pct 20 · match_sota_evo_pct 50 Junlin Yang et al. · arXiv:2607.28568v1 0 links
E3 cs-ai/agents Active On held-out NatureBench Lite with the framework held fixed, swapping in the trained Frontis-MA1 model raised Match-SOTA from 50% to 70%. match_sota_base_pct 50 · match_sota_trained_model_pct 70 Junlin Yang et al. · arXiv:2607.28568v1 0 links
E3 cs-cv/vision Active The complete Chimera system (11B parameters, 2B activated, with sparse Mixture-of-Experts) reached 7.3x the compute efficiency of a matched full-attention Wan-2.1 2B baseline, measured by pretraining diffusion loss. compute_efficiency_multiple 7.3 · total_params_b 11 · activated_params_b 2 · baseline_params_b 2 Chongjian Ge et al. · arXiv:2607.28611v1 0 links
E3 cs-se/software-engineering Active For inter-service communication recovery, OpenAI o3 achieved F1 = 0.61 under zero-shot prompting (high recall, low precision) and F1 = 0.82 under few-shot prompting (two systems; one execution per condition). f1_zero_shot 0.61 · f1_few_shot 0.82 · num_systems 2 · runs_per_condition 1 Danyllo Albuquerque et al. · arXiv:2607.28307v1 0 links
E3 cs-se/software-engineering Active Using OpenAI o3 to synthesize microservice architectures from textual requirements, service-identification F1 versus reference architectures was 0.79 under zero-shot prompting and 0.97 under few-shot prompting (two systems: Bookstore and PetClinic; one execution per condition). f1_zero_shot 0.79 · f1_few_shot 0.97 · num_systems 2 · runs_per_condition 1 Danyllo Albuquerque et al. · arXiv:2607.28307v1 0 links
E3 cs-se/software-engineering Active Change2Task achieves 79.6% verified task construction success from 1,130 eligible source changes across five coding task families (Bug Fix, Feature Addition, Test Generation, API Migration, Security Repair). source_changes 1130 · verified_success_pct 79.6 · task_families 5 Haomin Qi et al. · arXiv:2607.28591v1 0 links
E3 eess/signal-processing Active Interfacial-potential lateral-flow immunoassays quantified estradiol, progesterone, and luteinizing hormone in human plasma with r² > 0.97 versus clinical analyzers. r2_min 0.97 Hyun-June Jang et al. · arXiv:2603.23775v1 0 links
E3 physics/earth-environment Active Using 1.3 million Clay v1.5 embeddings and 10,015 withheld forest-inventory plots spanning 4,672 plant species, a continuous softmax biome representation predicted species occurrence at mean per-species AUC 0.618 versus 0.570 for discrete biome labels across 10 spatial cross-validation folds. embeddings_millions 1.3 · n_plots 10015 · n_species 4672 · n_folds 10 · auc_continuous 0.618 · auc_discrete 0.57 Maxwell B. Joseph et al. · arXiv:2606.11510v1 0 links
E3 q-bio/clinical-genomics Active In TCGA-cRCC (n = 162) of a three-cohort TCGA radiogenomic sweep (n = 340 total), an Evo 2 severity-by-radiomics analysis recovered established renal-cancer drivers and identified 46 additional genes reaching FDR significance that were absent from curated cancer-gene panels. n_crcc 162 · n_total 340 · n_additional_genes 46 Frederik Hauke et al. · arXiv:2607.20583v1 0 links
E3 q-bio/neuroscience Active In a VR head-pointing task, patients with central visual field loss (n = 25) needed a mean pointer-activation-zone diameter of 3.48° to reach best performance at a 2° target, versus 1.32° for normally sighted controls (n = 26). n_patients 25 · n_controls 26 · paz_patients_deg 3.48 · paz_controls_deg 1.32 · target_deg 2 Camille Bordeau et al. · arXiv:2605.19816v1 0 links
E3 q-bio/single-cell Active Nephrobase Cell+, pretrained on ~39.5 million kidney single-cell and single-nucleus profiles from 4,319 samples, achieved >90% zero-shot annotation accuracy for major kidney lineages in both human and mouse. n_profiles_millions 39.5 · n_samples 4319 · zeroshot_acc_pct 90 Chenyu Li et al. · arXiv:2509.26223v1 0 links
E3 q-bio/systems-physiology Active In stretched single collagen fibrils from a pair of bovine leg tendons (same animal; divalent- vs trivalent-crosslink rich), the relative density of SHG emitters versus molecular strain follows a two-state transition with a free-energy barrier between 6 and 10 kBT. barrier_kbt_low 6 · barrier_kbt_high 10 · n_animals 1 · n_tendons 2 MacAulay Harvey et al. · arXiv:2608.18898v1 0 links