Live topic stream

Models & Research

Name: LDS Mentor
Availability: InStock

Model releases, research results, benchmarks, evaluations, and methods. Reader coverage is updated continuously; Google Search eligibility is decided separately.

Stream archive

Latest coverage

Browse all topics

7.3
convolutional-networks
Jul 15, 20263 sources
NVIDIA Opens an AI Pre-Decoder for Quantum Color Codes
NVIDIA researchers have released an open training pipeline and model resources for an AI pre-decoder targeting triangular quantum color codes. The system uses a local three-dimensional convolutional network to simplify error syndromes before the Chromobius decoder handles the remaining work. At code distance d=31 and a physical error rate of 0.3%, the authors report a 347x improvement in logical failure rate and a 7.33x runtime reduction versus raw Chromobius. Those are simulation and implementation results from the authors, not evidence from a fault-tolerant quantum computer. The architecture is promising because local pre-decoding can fit parallel space-time workflows, but practical value depends on independent reproduction, hardware-specific noise tests, end-to-end latency, and accuracy under drift.
Read story
6.1
bioinformatics
Jul 15, 20262 sources
CPP2Vec Learns Efficient Sequence Representations for Peptide Delivery Screening
A PLOS Computational Biology article presents CPP2Vec, a Word2Vec-based system for prioritizing cell-penetrating peptides. The authors combine experimentally curated peptides with computationally generated sequences, learn peptide embeddings, and train separate supervised models for peptide classification, uptake efficiency, and enhanced PMO delivery. They also compare the approach with embeddings from pretrained protein language models. The authors report competitive predictive performance and strong computational efficiency across their tasks. CPP2Vec is an in-silico screening tool, not evidence that a candidate works in cells, animals, or patients. LDS recommends strict train-test separation, external peptide datasets, calibrated uncertainty, sequence-similarity controls, and prospective wet-lab validation before predictions influence therapeutic development decisions.
Read story
6.8
bayesian-optimization
Jul 15, 20261 source
LLM Evolution Generates Multi-Objective Bayesian Optimization Algorithms
A new arXiv preprint extends LLaMEA to generate complete multi-objective Bayesian optimization algorithms. LLMs act as mutation and crossover operators inside an evolutionary strategy, while SMAC tunes hyperparameters during the search. Across nine evolutionary runs, the authors generated about nine hundred algorithms and evaluated them on twelve synthetic and three engineering problems against a qParEGO baseline. The authors report 0.971 mean normalized hypervolume and roughly 60x lower wall-clock time for the strongest synthetic-suite result. They also report transfer to unseen engineering problems. LDS sees the work as a promising algorithm-design method, but recommends controlling benchmark access, compute budgets, implementation validity, and repeated-run variance before treating generated optimizers as better than expert designs.
Read story
6.4
ai-scientists
Jul 15, 20261 source
HEP Makes AI Scientist Hypothesis Changes Explicit and Auditable
A new arXiv preprint proposes the Hypothesis Evolution Protocol, or HEP, as an agent harness for making scientific reasoning inspectable. Current research agents may propose hypotheses, run tests, and revise beliefs, but those transitions often disappear inside unstructured logs. HEP represents hypothesis generation, evaluation, evidence, and evolution as explicit operations that researchers can inspect. On materials-science tasks, the authors report that HEP-equipped agents followed the hypothesis-test-evidence-belief cycle more completely than planning-style agents and used the protocol more effectively as the base model improved. LDS sees auditability as a meaningful advance, but the protocol still needs independent reproduction, tamper-resistant records, falsification checks, and human challenge points before research teams rely on its conclusions.
Read story
6.3
concept-based-xai
Jul 15, 20261 source
ConceptSMILE Audits Whether Concept-Based AI Explanations Are Trustworthy
A new arXiv preprint introduces ConceptSMILE, a model-agnostic audit layer for concept-based explainable AI. The framework perturbs image regions, measures how concept responses change, applies locality weighting, and fits an XGBoost surrogate around the prediction being explained. It then evaluates attribution accuracy, surrogate fidelity, faithfulness, stability, and consistency rather than assuming that human-readable concepts are reliable. In retinal fundus experiments, the authors report that MedSAM concepts led spatial attribution and produced fidelity scores of 0.8503 and 0.8465 while the vision-language pathway performed better on selected vessel-faithfulness and artefact-stability tests. LDS views the work as a useful audit template, but the evidence remains author-run, task-specific, and unvalidated for clinical use.
Read story
7.2
coding-agents
Jul 15, 20262 sources
Researchers Map Structural Failures in LLM-Generated Code
A new research paper calls locally plausible but repository-incoherent AI code the patchwork problem. The authors define eight structural failure categories covering imports, internal APIs, dependencies, configuration, resources, control flow, cross-file contracts, and security wiring. They evaluated 336 generated changes from two frontier models under four prompting strategies. The paper reports 67 structural failures, with 65 missed by type checking, tests, static application security testing, and regex baselines. Those results are author-reported and need independent replication. LDS argues that teams using coding agents should add repository-wide invariant checks and deployment-contract tests, because a patch that compiles and passes local tests can still violate configuration, schema, routing, or resource assumptions elsewhere in the system.
Read story
6.3
generative-planning
Jul 15, 20261 source
Shortcut Models Simplify Generative Planning for Offline RL
Researchers propose Shortcut Trajectory Planning, a single-stage generative planner for offline reinforcement learning with an adjustable inference budget. The system generates possible future state trajectories, ranks them with a learned critic and a feasibility penalty, and converts the selected plan into an action through inverse dynamics. The authors evaluate STP across locomotion, navigation, manipulation, and dexterous-control tasks in the D4RL benchmark suite. They report competitive results while avoiding the separate teacher and student training used by consistency-based planners. The evidence is author-reported and has not been independently reproduced. LDS sees the main deployment question as whether lower-step generation preserves control quality when latency, candidate selection errors, and environment shifts are measured together.
Read story
7.0
computational-pathology
Jul 15, 20261 source
ALICE Distills Specialist Pathology Models Into One Foundation Backbone
A new arXiv preprint introduces ALICE, a pathology foundation model built through staged distillation from vision, vision-language, and slide-level experts. Rather than forcing one objective to absorb every capability at once, the method agglomerates eight teacher models into dedicated modules on a shared backbone. ALICE distills eight specialist teachers and is pretrained on 24,985,184 tile images plus 155,604 high-resolution images. The authors evaluate 21 scenarios, 96 tasks, and 48 sources, reporting the best average rank across their three settings. Those results are broad but author-run. LDS views external-site validation, rare-disease stress tests, calibration, and workflow-specific error analysis as essential before the model is considered dependable for pathology research or clinical decision support.
Read story
6.2
digital-twins
Jul 15, 20261 source
LDT-Coord Uses Digital Twins to Cut Embodied-Agent Coordination Traffic
A new arXiv preprint proposes LDT-Coord, a coordination layer for heterogeneous LLM embodied agents operating across computing and network resources. Each agent reports an intended action plus structured temporal constraints to a lightweight digital twin instead of maintaining repeated natural-language negotiation with every peer. A rule-based resolver handles ordinary conflicts, while a constrained reinforcement-learning controller decides when agents should report updates. The authors report comparable task success with more than 70x lower communication overhead in simulation. The result is promising but author-run and simulation-bound. LDS sees the main engineering question as whether compact state reports remain complete, timely, and trustworthy when agents, environments, and network conditions change outside the benchmark.
Read story
6.5
llm-reasoning
Jul 15, 20261 source
Anchor Retrieval and LLM Reranking Target Source Recovery From Binaries
A new arXiv preprint proposes recovering source functions from stripped binaries by searching a source database rather than generating approximate pseudocode. The pipeline extracts strings, constants, external calls, and surviving function names with Ghidra, retrieves candidate files through an inverted index, narrows them to function snippets, and asks an LLM to rerank the candidates against binary context. The authors report 95.2% assembly instruction coverage on a controlled, high-fidelity database and 35.5% average coverage on a GitHub-based database. That gap makes retrieval quality the central operational risk. The results are author-run and have not been independently reproduced, so practitioners should separate retrieval recall from reranking accuracy before judging the method.
Read story
6.0
ai-agents
Jul 15, 20261 source
Agora Routes Agent Tasks Through Confidence-Calibrated Auctions
University of Birmingham researchers have proposed Agora, a preprint framework that decomposes a request into task units and allocates each unit among candidate models or tools through a confidence-calibrated auction. Bids combine predicted competence with monetary and latency costs, while feedback can update calibration over time. The authors report improvements over matched routing and cascade baselines across text, scientific-code, and multimodal benchmarks, plus a tunable cost-quality tradeoff. These results remain author-controlled and do not establish production reliability or economic incentive compatibility. LDS recommends replaying real workload traces, calibrating on held-out data, testing distribution shift, and measuring end-to-end quality, cost, latency, and failure propagation before deployment.
Read story
6.5
bioinformatics
Jul 15, 20261 source
TheBioCollection Unifies Biological Data for Language-Model Training
Researchers have released TheBioCollection, a preprint corpus that converts heterogeneous biological databases into training-ready language-model records spanning molecules, proteins, genomes, cells, pathways, and cross-domain relationships. The release also includes instruction tasks, an evaluation suite, and a model trained with the corpus. The authors report broad gains when holding the base architecture fixed, but the benchmark and training corpus were designed together and have not been independently replicated. LDS sees the main contribution as data engineering: traceable source identifiers, deterministic biological features, structured tags, and evaluation-overlap controls. Teams considering the resource should audit licensing, provenance, contamination, biological validity, and performance on external benchmarks before adoption.
Read story
6.0
convolutional-neural-network
Jul 15, 20261 source
3D CNN Models Subgrid Turbulence in Quiet-Sun Simulations
Researchers developed a 3D convolutional neural network to estimate unresolved turbulent transport in realistic quiet-Sun hydrodynamic simulations. The accepted ApJS manuscript compares the network with a multilayer perceptron and physics-based gradient and Smagorinsky baselines. The authors report average improvements of about 31% on diagonal Reynolds-stress components and about 8% on off-diagonal components, with a logarithmic target transformation helping on skewed values. These are author-reported simulation results, not independent validation or a production closure model. LDS explains the deployment test that matters: a surrogate should preserve conservation behavior, remain stable across resolution and solar regimes, and improve end-to-end simulation outcomes rather than only pointwise error.
Read story
6.6
adversarial-machine-learning
Jul 15, 20264 sources
Avida Study Shows Near-Perfect Life Classifier Can Be Spoofed Out of Distribution
A study accepted for an artificial-life conference shows how a classifier with 99.97% held-out accuracy can still assign extreme confidence to adversarial examples outside its training distribution. The experiment used short programs in the Avida digital-life environment, not extraterrestrial chemistry, rover instruments, or an operational space-mission model. Across 1,560 targeted search runs, the authors report that model confidence could reach 100% within 150 queries. That is an attack result, not a real-world false-positive rate or proof that every astrobiology model fails. LDS examines the engineering lesson: scientific classifiers need out-of-distribution detection, calibrated abstention, adversarial testing, independent physical confirmation, and human review before high-confidence outputs can trigger extraordinary claims.
Read story+1
5.8
llm-as-judge
Jul 15, 20261 source
NL-PAC Preprint Derives Prompt-Specific Risk Floors for LLM Supervision
A new arXiv preprint proposes NL-PAC, a framework for certifying a minimum unavoidable error when natural-language supervision permits multiple interpretations but observed labels do not reveal which interpretation is intended. The author distinguishes identification failure from ordinary sampling uncertainty: collecting more labels cannot resolve a hidden rubric meaning when the data-generation process looks compatible with several targets. The accompanying experiment is a limited probe using one frozen judge and a specified moderation prompt, not a benchmark of production LLM evaluators. LDS translates the result into an audit rule: freeze the model, prompt, threshold, decoding mode, and input distribution; test prompt paraphrases; validate human interpretations externally; and escalate rather than treating a zero certificate as proof of safety.
Read story
6.5
bioinformatics
Jul 15, 20264 sources
BINND Deep Learning Model Predicts Complex DNA-DNA Binding
A peer-reviewed Nature Communications paper introduces BINND, a deep-learning model for predicting whether pairs of DNA sequences bind to each other. The authors and NC State report training and evaluating the approach across 144 million sequence pairs, with 83.5% proof-of-concept accuracy and inference described as 50x faster than current comparison methods. The public paper reports accuracy above 80% across diverse sequences, but false negatives were more common than false positives. The result is a research milestone, not a deployed DNA-computing system, and the application claims remain prospective. Three authors disclose that they co-founded DNAli Data Technologies. LDS examines the evaluation risks that matter next: sequence-family leakage, calibration, false-negative cost, out-of-distribution generalization, and reproducibility from the public code.
Read story+1
6.0
combustion
Jul 15, 20261 source
Preprint Tests Entropy-Constrained ML for Faster Combustion Simulation
A new arXiv preprint tests a neural surrogate for chemical kinetics that trades some one-step prediction accuracy for better stability when coupled back into a combustion simulation. The authors combine a feed-forward network, radial-basis interpolation, residual data augmentation, and a soft penalty intended to discourage negative entropy production. In their two-dimensional turbulent methane-air test, the unconstrained surrogate diverged before completing the target rollout, while the constrained version remained bounded through the training horizon and later diverged outside it. The result is promising but narrow: the work is not peer reviewed, does not enforce mass conservation, has no visible public code or data release, and evaluates one simulation setting. LDS focuses on the gap between strong offline metrics and stable behavior inside a dynamical system.
Read story
7.0
contextual-bandits
Jul 15, 20261 source
CABS Router Uses Surrogate Scores to Reduce LLM Exploration
A new arXiv paper proposes correlation-aware contextual bandits for routing each prompt to an LLM under an accuracy-cost objective. The system uses predicted surrogate rewards from related models to learn from more than the selected model's observed result. Its decoupled variant, CABS-D, keeps the ordinary bandit and surrogate-aware learner separate, then adapts between them when proxies become unreliable. Across three offline routing benchmarks, the authors report stronger utility than standard implementable online bandits. In one low-cost-sensitivity setting, linear CABS-D scored 0.5328 versus 0.4876 for LinUCB. The result is promising evidence for faster simulated learning, not a deployment guarantee: it depends on historical benchmark labels, fixed model pools, learned correlations, and offline reward predictors that may drift in production.
Read story
6.5
anomaly-detection
Jul 15, 20262 sources
Pareto-DQN Trades Precision for Recall in Financial Anomaly Tests
A new arXiv paper proposes Semantic Pareto-DQN, a multi-objective reinforcement-learning system that converts tabular financial records into narrative embeddings and separately optimises detection value, customer friction, and semantic novelty. On the paper's e-commerce test split, the model reached 0.267 recall at 0.069 precision, producing 371 true positives and 5,026 false positives. On the UCI credit-default test, it reported 0.419 recall, 0.563 precision, and a 0.481 F1 score. The results show a configurable recall-friction trade-off, not a production fraud breakthrough: the evaluation uses manually calibrated rewards, single chronological splits, no confidence intervals, no deployed cost study, and no independent reproduction. The real test is whether the method beats calibrated supervised baselines at the same operational review budget.
Read story
6.6
foundation-models
Jul 15, 20262 sources
Foundation Force Fields Face Lunar Regolith Transferability Test
A v1 arXiv preprint submitted on July 10, 2026 benchmarks six universal machine-learning force fields on four minerals relevant to lunar regolith. The authors ran all 24 model-mineral combinations for 1,000 molecular-dynamics steps at 300 K, a trajectory of 1 picosecond, and compared local structures with crystallographic references. Every run remained stable, but iron-oxygen and titanium-oxygen environments varied more across models than silicon, magnesium, aluminum, or calcium coordination. On one NVIDIA RTX 4090, SevenNet-0 delivered the shortest author-reported runtimes, while MatterSim used the least peak memory. The result is a useful screening benchmark, not validation for lunar reaction chemistry: it lacks direct DFT force or energy errors, long trajectories, defects, amorphous regolith, and independent reproduction.
Read story
7.0
ai-agents
Jul 15, 20263 sources
SETA Releases 4,567 Verifiable Environments for Training Terminal Agents
A 22-author team submitted the first arXiv version of SETA: Scaling Environments for Terminal Agents on July 12, 2026. The project releases 4,567 executable training environments, with 3,255 synthesized tasks and 1,312 evolved tasks, plus code and dataset artifacts. SETA-Synth converts grounded technical material into containerized tasks, while SETA-Evol adjusts context and difficulty using model performance. The authors report that a Qwen3-8B model trained on a selected 560-environment subset averaged 10.7 percent on Terminal-Bench 2.0 and the base model averaged 3.1 percent. The widely cited 12 percent result is the best run, not the average. This is an unreviewed v1 preprint with author-run evaluations, not independent replication.
Read story
6.6
ai-agents
Jul 14, 20262 sources
Richard Sutton Launches Oak Lab for Continual Learning Research
Richard Sutton and Khurram Javed founded Oak Lab in July 2026, leaving Keen Technologies to pursue continual-learning systems built from real-time experience. The lab's official mission says its algorithms learn without storing or replaying data and target far lower compute demand than mainstream batch-trained systems. Oak Lab's stated long-term goal is an agent that learns and plans in real time using 20 watts. For ML practitioners, the immediate significance is methodological rather than product-ready: the venture is betting on batch-size-one credit assignment, event-driven networks, and temporal abstractions instead of ever-larger curated datasets. BetaKit independently confirmed the launch and the founders' move. The technical claims remain research goals, so evidence should come from reproducible benchmarks showing stable continual learning, selective forgetting, planning quality, and energy use outside controlled demonstrations.
Read story
6.4
ai-research
Jul 14, 20262 sources
Anthropic commits $10 million CAD to Canadian AI research
Anthropic committed $10 million CAD to Canadian AI research institutions, pairing model access with support for projects in safety, health, languages, and core research. The program covers Amii, Mila, and Vector plus CHEO, CAMH, Universite Laval, the University of Toronto, and the University of Saskatchewan. Across the partnerships, Claude credits will support work ranging from reinforcement learning and AI trust to children's health, psychiatric fairness, low-resource languages, and biomedical research. Anthropic also said the three regional AI institutes will join its startup program this summer. The commitment broadens researcher access to a frontier model, but Anthropic did not publish a detailed split between cash funding and model credits, leaving allocation and measurable outcomes as the main details to watch.
Read story
6.8
ai-research
Jul 14, 20263 sources
Soofi S Brings Open German-English Foundation Model to Preview
Soofi S is a German and English foundation model released as an open research preview by the Soofi consortium, pairing a hybrid mixture-of-experts design with sovereign European infrastructure. The project report and independent coverage describe a base model intended for downstream adaptation rather than a finished consumer assistant. Its practical appeal is traceability: the team says it will publish weights, selected checkpoints, training and evaluation code, and detailed data accounting. For practitioners, the meaningful test is not the consortium's benchmark lead but whether independent evaluations reproduce German-language quality, long-context efficiency, and code performance on real workloads. Licensing details, deployment tooling, safety post-training, and production reliability still need careful review before enterprise use.
Read story
7.2
ai-safety
Jul 14, 20262 sources
Clinical RAG Study Finds Citations Can Support Wrong Entities
A new preprint from Lunit researchers identifies deceptive grounding, a clinical RAG failure in which a response cites real evidence but attaches it to the wrong drug or disease. In controlled testing across 13 models, reported failure rates ranged from 8% to 87% under the strongest adversarial conditions, while specialized medical models reached 86.7%. A production measurement across 740 drug-disease pairs found a 7.8% overall rate, rising to 13.6% for recently approved drugs. The authors report that an entity-attribution check reached 97.0% precision and 98.7% recall for the failure. For teams building high-stakes retrieval systems, the paper argues that citation validity and faithfulness are incomplete unless evaluators also verify that each source concerns the exact entity named in the answer.
Read story
6.4
ai-research
Jul 13, 20263 sources
Study Shows Emotional Authenticity Drives AI Afterlife Interactions
A University of Colorado Boulder study found that people interacting with AI simulations of deceased loved ones generally preferred first-person versions, while familiar emotional tone and speech patterns mattered more than strict factual precision. Researchers compared a reincarnation mode that spoke as the deceased with a representation mode that spoke about them. Participants often treated both as direct conversation, showing that interface labels alone did not control the experience. For AI teams, the central safety lesson is that persuasive emotional realism can emerge from limited personal data, so consent, dependency safeguards, provenance, and clear boundaries should be designed before these systems reach grieving users.
Read story
5.8
agentic-ai
Jul 13, 20262 sources
QAgent Uses Multi-Agent Workflow for OpenQASM Code Generation
QAgent is a multi-agent system for OpenQASM code generation, and its authors report a 71.6% improvement in accuracy over static LLM approaches. The system routes regular circuits toward example-based generation and harder parameterized tasks toward tool-assisted planning, execution, and reflection. For quantum-software practitioners, the useful signal is a workflow that tests syntax and functional behavior instead of trusting one model response. The paper also reports weaker performance as tasks combine more quantum algorithms, making decomposition and executable validation central to safe use. QAgent therefore offers a research blueprint for LLM agents that generate low-level quantum programs, but the reported results still need independent replication and human oversight before production deployment.
Read story
7.4
evaluation
Jul 10, 20262 sources
MLCommons Adds Agentic Inference Benchmark To MLPerf
MLCommons added Agentic Inference to MLPerf Inference on July 8, 2026, giving serving teams a benchmark for multi-turn agents rather than isolated prompts. The benchmark measures how LLM systems handle growing context, KV-cache pressure, tool-mediated turns, and closed-loop task progress across coding and enterprise workflow traces. MLCommons says the datacenter workload combines 490 coding trajectories and 500 workflow trajectories, while a July 9 edge call adapts the idea to single-accelerator deployments with Qwen3.6-27B and BFCL v4. For practitioners, the practical signal is that agent reliability now includes progress per user, output-quality gates, and latency under dependent turns, not just aggregate tokens per second.
Read story
5.4
ai-research
Jul 10, 20261 source
PLoS Computational Biology Reviews Two Decades of Systems Biology
PLOS Computational Biology published a July 10, 2026 perspective marking its 20th anniversary and reviewing how systems biology matured from early-2000s institutions into a data-rich, model-driven field. The authors frame the field as broader than a journal milestone: standards such as SBML and BioModels, reproducible modeling practices, single-cell data, uncertainty analysis, and AI/ML are now part of the same toolkit. For data-science practitioners, the useful signal is that biological modeling work is moving toward hybrid approaches that combine mechanistic models, shared artifacts, and machine-learning methods, while still depending on careful provenance and reproducibility rather than model scale alone.
Read story
6.5
protein-design
Jul 9, 20264 sources
Embedding Guide Navigates Sequence Space for Enzymes
PLOS Computational Biology published "Mind the gap" on July 9, 2026, presenting an embedding-guided method for navigating enzyme sequence space while avoiding implausible mutation paths. The study uses protein-language-model embeddings with Monte Carlo sampling to propose enzyme mutants, giving computational-biology teams a more structured way to explore variants than blind local search. For ML practitioners, the useful point is that embeddings are not just similarity maps; they can become constraints for safer design-space traversal. The result is still research-stage, but it connects representation learning to practical enzyme redesign workflows where experimental validation is expensive.
Read story+1

Models & Research

Latest coverage

NVIDIA Opens an AI Pre-Decoder for Quantum Color Codes

CPP2Vec Learns Efficient Sequence Representations for Peptide Delivery Screening

LLM Evolution Generates Multi-Objective Bayesian Optimization Algorithms

HEP Makes AI Scientist Hypothesis Changes Explicit and Auditable

ConceptSMILE Audits Whether Concept-Based AI Explanations Are Trustworthy

Researchers Map Structural Failures in LLM-Generated Code

Shortcut Models Simplify Generative Planning for Offline RL

ALICE Distills Specialist Pathology Models Into One Foundation Backbone

LDT-Coord Uses Digital Twins to Cut Embodied-Agent Coordination Traffic

Anchor Retrieval and LLM Reranking Target Source Recovery From Binaries

Agora Routes Agent Tasks Through Confidence-Calibrated Auctions

TheBioCollection Unifies Biological Data for Language-Model Training

3D CNN Models Subgrid Turbulence in Quiet-Sun Simulations

Avida Study Shows Near-Perfect Life Classifier Can Be Spoofed Out of Distribution

NL-PAC Preprint Derives Prompt-Specific Risk Floors for LLM Supervision

BINND Deep Learning Model Predicts Complex DNA-DNA Binding

Preprint Tests Entropy-Constrained ML for Faster Combustion Simulation

CABS Router Uses Surrogate Scores to Reduce LLM Exploration

Pareto-DQN Trades Precision for Recall in Financial Anomaly Tests

Foundation Force Fields Face Lunar Regolith Transferability Test

SETA Releases 4,567 Verifiable Environments for Training Terminal Agents

Richard Sutton Launches Oak Lab for Continual Learning Research

Anthropic commits $10 million CAD to Canadian AI research

Soofi S Brings Open German-English Foundation Model to Preview

Clinical RAG Study Finds Citations Can Support Wrong Entities

Study Shows Emotional Authenticity Drives AI Afterlife Interactions

QAgent Uses Multi-Agent Workflow for OpenQASM Code Generation

MLCommons Adds Agentic Inference Benchmark To MLPerf

PLoS Computational Biology Reviews Two Decades of Systems Biology

Embedding Guide Navigates Sequence Space for Enzymes

Models & Research

Latest coverage

NVIDIA Opens an AI Pre-Decoder for Quantum Color Codes

CPP2Vec Learns Efficient Sequence Representations for Peptide Delivery Screening

LLM Evolution Generates Multi-Objective Bayesian Optimization Algorithms

HEP Makes AI Scientist Hypothesis Changes Explicit and Auditable

ConceptSMILE Audits Whether Concept-Based AI Explanations Are Trustworthy

Researchers Map Structural Failures in LLM-Generated Code

Shortcut Models Simplify Generative Planning for Offline RL

ALICE Distills Specialist Pathology Models Into One Foundation Backbone

LDT-Coord Uses Digital Twins to Cut Embodied-Agent Coordination Traffic

Anchor Retrieval and LLM Reranking Target Source Recovery From Binaries

Agora Routes Agent Tasks Through Confidence-Calibrated Auctions

TheBioCollection Unifies Biological Data for Language-Model Training

3D CNN Models Subgrid Turbulence in Quiet-Sun Simulations

Avida Study Shows Near-Perfect Life Classifier Can Be Spoofed Out of Distribution

NL-PAC Preprint Derives Prompt-Specific Risk Floors for LLM Supervision

BINND Deep Learning Model Predicts Complex DNA-DNA Binding

Preprint Tests Entropy-Constrained ML for Faster Combustion Simulation

CABS Router Uses Surrogate Scores to Reduce LLM Exploration

Pareto-DQN Trades Precision for Recall in Financial Anomaly Tests

Foundation Force Fields Face Lunar Regolith Transferability Test

SETA Releases 4,567 Verifiable Environments for Training Terminal Agents

Richard Sutton Launches Oak Lab for Continual Learning Research

Anthropic commits $10 million CAD to Canadian AI research

Soofi S Brings Open German-English Foundation Model to Preview

Clinical RAG Study Finds Citations Can Support Wrong Entities

Study Shows Emotional Authenticity Drives AI Afterlife Interactions

QAgent Uses Multi-Agent Workflow for OpenQASM Code Generation

MLCommons Adds Agentic Inference Benchmark To MLPerf

PLoS Computational Biology Reviews Two Decades of Systems Biology

Embedding Guide Navigates Sequence Space for Enzymes