The Archive

SIRA framework improves retrieval-augmented agents by modeling expert search priors, reducing retrieval rounds and latency for organizational knowledge bases.

Extension of Venn-Abers predictors to unbounded regression using conformal prediction; narrow technical contribution to probabilistic forecasting.

Chromophore-region 3D mechanism graphs for fluorescent protein quantum-yield prediction; domain-specific protein modeling task, limited AI audience relevance.

Comprehensive benchmark for Multimodal Domain Generalization revealing inconsistent evaluation protocols and fragmented research; validates real-world robustness challenges.

StraTA framework adds explicit strategy sampling to agentic RL, improving credit assignment and exploration over long-horizon LLM decision-making.

Hybrid concept-based and abductive explanations for vision models using formal causal reasoning; advances interpretability beyond single-concept limits.

GlazyBench dataset (23K formulations) for ceramic glaze property prediction and generation; domain-specific multimodal benchmark with limited broad AI relevance.

RAO trains recursive agents to delegate sub-tasks recursively, enabling divide-and-conquer inference scaling for longer contexts and harder problems.

ScaleLogic benchmark isolates proof-planning depth and logic expressiveness to systematically study RL training scaling for LLM long-horizon reasoning.

Framework for parsing and verifying source attribution in LLM research agents; evaluates citation accuracy via AST parsing and reproducible verification.

Method to compress SFT-induced LLM behaviors into sparse, causally necessary subnetworks for selective inference-time control.

Hybrid quantum-classical GAN framework for generating adversarial network traffic using variational quantum generators.

Lightweight asymmetric neural codec for compressing multi-modal sensor data on bandwidth-constrained edge devices.

PianoCoRe: unified piano MIDI dataset with 250k performances, 5.6k pieces, and note-level alignments for music AI.

Human-in-the-loop L2 Korean morphosyntactic annotation using parser agreement as proxy for correctness.

MASPO: framework for joint prompt optimization across LLM-based multi-agent systems with unified evaluation.

Formalization of Algospeak coevolution dynamics between LLM evasion and detection; introduces Majority Understandable Modulation metric.

OpenAI releases three audio models via API enabling voice app development for developers.

Theoretical analysis of why sign-based optimizers (SignSGD, Muon) outperform SGD in foundation model training via ℓ1-norm bounds.

SkillOS: self-evolving LLM-agent framework learning long-horizon skill curation policies from streaming task interactions.

Online Bayesian calibration method for time-evolving digital twins handling gradual drift and abrupt system changes.

Mechanistic analysis traces attention sink phenomenon to variance discrepancy in value aggregation and FFN super-neuron activation.

SoftSAE introduces dynamic sparsity for Sparse Autoencoders, adapting feature activation count per input for improved mechanistic interpretability.

Proves Transformers implement normalized gradient descent for in-context logistic regression, formalizing implicit algorithmic execution.

DARTS optimizes covariate selection in budget-constrained randomized trials using Thompson sampling for causal inference.

AI CFD Scientist demonstrates LLM-based agents automating computational fluid dynamics discovery with physics validation loops.

Dynamic budget allocation for multi-turn LLM evaluation under conformal survival frameworks to predict rare jailbreak events.

Patch2Vuln reconstructs security vulnerabilities from Linux binary patches using LLM agents without source code access.

Functional-analytic proof that L² weight decay on Transformer cross-entropy satisfies Villani coercivity criteria for optimization guarantees.

UniSD unifies self-distillation design choices for LLM adaptation without external teachers, systematizing complementary techniques.