The Archive

Code world models risk exploitation in continuous control when sampling-based acceptance fails to detect rare critical events.

LLM hidden state manifolds organize as small-world networks enabling multi-hop reasoning; bypasses attention artifacts via geometric analysis.

SIGMA uses SHAP guidance for LLM-based automated feature engineering without metadata, reducing context window overflow and local optima.

Continual Abstraction Discovery evolves procedural content generators via LLM mutation, extracting reusable primitives across game level design.

Graph-structured online difficulty estimation optimizes RLVR exploration budgets for LLM reasoning by adaptive sample-difficulty scheduling.

Small models grade open-ended exam answers reliably as frontier models when given explicit rubrics, reducing inference cost 80%+.

EvoTS-Agent self-evolves via validation feedback for autonomous financial time-series change-point detection without manual model selection.

Collective Counterfactual Planning formalizes multi-agent coordination where representational geometry constrains perception and goal verification.

SpeechSense dataset captures paralinguistic cues for speech sentiment analysis, preserving acoustic features beyond ASR transcription.

Adaptive policy portfolios optimize robustness in MDPs by synthesizing offline policies with lightweight online selection, reducing conservative bounds when environment dynamics become identifiable.

Decentralized planning framework for lifelong multi-agent path finding that improves computational scalability over RHCR through group-based MDP theory.

Deep learning classifier for appendicitis severity grading in ultrasound images using Grad-CAM explainability; domain-specific medical imaging application.

Taxonomy-based analysis of question types students ask generative AI during CS coursework, finding distinct inquiry patterns in learning workflows.

Benchmark of YOLO and object detection models on military target detection using civilian datasets, demonstrating transfer learning viability for defense applications.

Hybrid decision-tree/linear-regression model improves pre-routing delay estimates in OpenLane by 80% error reduction for open-source chip design.

CABLE memory retrieval system for LLM agents addresses reachability of semantically-distant evidence in long-term conversational history via antecedent linking.

AutoResearch connects idea generation to experimental execution for autonomous research, integrating research signals with domain knowledge and multi-model validation.

Dynamic compression technique enables recurrent networks to revisit and revise fixed-size history state, improving efficiency on long-context tasks.

BEAR-Bench evaluates multimodal LLMs on bilingual text-dense document reasoning (English/Russian) across enterprise and academic contexts.

LSTM with residual learning for aerodynamic load prediction on NLR 7301 airfoil benchmark using CFD data.

BayesPrompt reformulates prompt optimization as Bayesian inference to generate human-readable prompts with competitive perplexity.

ARASH selects optimal in-context examples for tabular prediction via few-shot prompting without retraining foundation models.

Optimization framework for split federated learning resource allocation and model splitting in edge networks.

MoRAX augments geospatial foundation models with human mobility data for transferable urban representation learning.

Probing study reveals six frozen LLMs encode geometric constraints in hidden states but fail to steer behavior using them.

Theoretical analysis of regret-instability trade-off in multi-armed bandit algorithms with finite-time lower bounds.

Novel white-box attack on LLMs using knowledge editing framework and associative context retrieval to induce unsafe outputs.

AdaLens provides interactive visualization and steering controls for long-running agentic data analysis workflows with LLMs.

LeakGauge detects context-leakage attacks on LLMs through output behavior signals without requiring hidden state extraction.

On Tuesday, Warp introduced Warp Factories, a new infrastrructure system designed to make building AI software factories as easy as possible.