The Archive

OSReward establishes standardized evaluation for VLM-based reward models assessing computer-using agent trajectories.

KAISEN five-phase audit pipeline detects subgroup fairness disparities in clinical risk models with stress-tested components.

Safety fine-tuning suppresses LLM consciousness attribution across entities; mechanistic steering restores human values.

Change2Task generates executable coding agent tasks from merged pull requests with verified state reconstruction and validation.

VAD counterfactual algorithm isolates visually attributable components in multimodal on-policy distillation corrections.

MixFrag uses KL-divergence-guided mixed-precision quantization to deploy Vision Transformers efficiently on resource-constrained devices.

PAIChecker reveals 13.6% misalignment between PR-issue pairs in SWE-bench-like benchmarks, questioning LLM code-resolution eval validity.

β-OPSD generalizes on-policy self-distillation via tunable KL regularization to improve reasoning language models with better stability.

DualG-MRAG decouples macro reasoning and micro-matching in multimodal retrieval-augmented generation to handle complex multi-hop tasks.

Study shows repeated sampling outperforms self-refinement and reflexion methods across 1.5B-7B models at equivalent token budgets with confidence intervals.

Computational complexity analysis of committee election winner determination under Thiele voting rules with FPT algorithms.

Empirical study of inference-time scaling strategies for local computer-use agents under hardware constraints across multiple dimensions.

Frontis-MA1 (35B) meta-evolution agent trained for recursive self-improvement in ML engineering via OpenMLE system with program-evolution operators.

DR-FRL applies doubly robust estimation to irregular longitudinal functional data for causal inference with efficient influence function targeting.

APO applies unsupervised policy optimization to 3D atomic structure prediction, eliminating need for ground-truth coordinate alignment.

ShieldFont open-source tool poisons font data to defend against unauthorized AI model scraping.

Apptronik’s Apollo 2 robot takes a baseball glove off of a shelf. | Image: Google Google DeepMind says the latest version of its Gemini Robotics AI model can "control entire humanoid robots." While the previous model focused on controlling a humanoid robot's upper body, Gemini Robotics 2 now supports "whole-body motions" ranging from its feet to fingertips, according to an announcement on Thursday. The new model will allow humanoid robots to perform a wider range of actions, as it allows them to walk, crouch, stretch, and manipulate objects. Videos shared by Google show how Apptronik's Apollo...

ORCA-bench evaluates LLM agents on production oncall root-cause analysis using real telemetry (Prometheus, Jaeger, OpenSearch) and code in realistic incident scenarios.

ScaFE uses LLM-generated feature programs to classify pathological scars (keloids vs. hypertrophic) from clinical photos with limited labeled data and local data governance.

Graph neural network magnetic force fields learn effective energy functionals for itinerant spin dynamics from electronic structure, analogous to ML interatomic potentials.

Study shows LLMs reproduce standard language ideologies, privileging Inner Circle English norms and marginalizing non-dominant World Englishes variants.

MANTA enables LLM-based multi-agent systems to self-adapt communication topology at inference time based on task and real-time collaboration performance.

DAR-Net addresses dual ambiguity (semantic and spatial) in all-in-one image restoration by disentangling degradation cues from scene content.

Proposes leakage-free evaluation protocol for same-graph cross-task transfer in GNNs across node classification and link prediction with fixed splits and negatives.

Introduces selective credibility-limited belief update framework extending Katsuno-Mendelzon model to handle partial compound epistemic inputs.

CS-RNR method enables agents in imperfect-information games to safely exploit flawed opponents with provable certificates on deployed strategy.

Multi-level framework models literary creativity as selective transformation across lexical, semantic, conceptual, structural, and narrative dimensions.

Sociolinguists analyze how generative AI impacts academic writing norms and linguistic diversity in global scholarly publishing.

Live benchmark measuring LLM resilience to state-backed information operations using 2,100+ real disinformation campaigns.

Target-conditioned abstraction method for scientific inspiration retrieval using analogical reasoning to transfer problem-solving principles.