The Archive

Evidence-traceable dynamic interviewer system adapts question depth and context for qualitative research using local LLMs.

SkillContrast improves agent skill retrieval by comparing text differences between similar skills rather than query matching, yielding 54-72% more safe retrievals.

Coding agents fail when co-installed skills conflict; benchmarks miss these failures because task completion passes despite unintended behavior changes.

Phonologically informed tokenization via syllabification achieves competitive performance on German speech recognition using wav2vec 2.0 backbone.

UXBench Pro introduces 1,000 multi-turn dialogue instances with personalized user profiles for evaluating heterogeneous user preferences beyond binary reward models.

LTBD defends LLMs against prompt injection by learning explicit trust-boundary delimiters that distinguish user intent from embedded malicious instructions.

Benchmarks 18 neural architectures on 10 temporal pattern recognition tasks for robotic gesture control across 250+ experimental configurations.

Gaussian differential privacy mechanisms enable continual machine unlearning via Newton updates that match exact retraining under sequential deletion.

Quotient-space exploration improves genome-scale metabolic model repair by discovering biological hypotheses rather than output diversity alone.

Consensus pseudo-label learning improves reliability by constructing complementary decision sources via graph structure variation and node coassignment.

Katie Harbath discusses her book on Facebook's political disruption during the 2010s; tangential to AI development and governance.

Layer-selective fine-tuning reduces LLM capability degradation by identifying which Transformer layers should adapt to specialized tasks.

Model-free policy gradient method for mean-field control using transport maps to capture population distribution effects.

Knowledge distillation framework for spatiotemporal prediction that organizes teacher model knowledge into task-aware memory.

AgentEvolver enables agents to develop reusable capabilities during task execution while keeping foundation models fixed.

CAMEO framework for clinically-aware ultrasound report generation addressing alignment between model outputs and visible evidence.

NanoProof: open-weights, reproducible theorem prover for Lean 4 with released dataset, tools, and weights achieving 50.8% pass@16.

Framework for identifying decisive errors in LLM multi-agent systems by modeling error propagation across agent interactions.

Study shows LLM version turnover undermines manuscript screening detectors: 23 LLM versions tested on 4000 abstracts from June 2023–Aug 2026.

Framework for organizing large coding agent patches into runnable, ordered commits while measuring maintenance impact.

ProloNg benchmark probes long-horizon deductive reasoning in 8 LLM families using Prolog with up to 62k context and depth-22 problems.

MH-INDIC: culturally grounded evaluation framework for maternal-health LLM interactions in North Indian contexts using ten reasoning dimensions.

Multilingual adaptation framework extends English quality classifiers to 100+ languages for LLM pretraining data selection.

Analysis of embedding bias in conditional independence testing shows when learned representations fail to preserve statistical validity.

SDPAD converts ANN-based autonomous driving perception to spike-driven SNNs achieving energy-efficient edge deployment for planning.

Physics-aware uncertainty propagation for autonomous vehicle trajectory prediction quantifies motion variable uncertainty through vehicle dynamics.

Chronos framework enables LLM code agents to leverage historical pull requests via structured experience graphs for software evolution reasoning.

Elastic Expert Routing smooths discrete top-k boundaries in sparse MoE training via stochastic expert budget sampling.

Multi-strategy LLM agent memory framework categorizes and processes diverse memory types with type-specific retrieval approaches.

Loop-native attention residuals stabilize looped Transformers across tens of thousands test-time iterations preventing performance degradation.