The Archive

Models systemic risk from compromise of AI vendors serving banking (fraud, credit, AML), demonstrating cascade through interbank networks.

Sample-adaptive routing for vision token pruning in MLLMs, using complementary strategies per input to reduce inference cost.

TimeCues Studio: open-source annotation and prototyping workspace for music detection algorithms, enabling team-scale corpus labeling.

Framework combining symbolic geometry parser with LLM solver to match LMM performance on plane geometry problems with reduced opacity.

OnPoKD: on-policy distillation framework for vision-language model adaptation under class/domain shifts with dynamic target construction.

TRACE: RL framework using synthesized simulator rewards for causal reasoning agent training when ground-truth verification is costly.

Combines active learning and lottery ticket hypothesis: reuses single training loop to discover sparse subnetworks while labeling.

View-Structured Conformal Prediction adds uncertainty quantification to 3D Gaussian Splatting via spatial calibration and view-difficulty factors.

Riemannian Language Models reduce sub-1M parameter LLM size by replacing output matrix with geodesic distance decoding on manifolds.

Sparse autoencoders exhibit feature merging dominance across parameter sweeps; systematic study shows near-zero full-dictionary recovery in MAIS-O43 benchmark.

Instinct’s new email feature lets the AI agent create and manage accounts, contact businesses, handle support requests, and do more on users' behalf.

POMDP-based automated intrusion response for OT systems using PPO; applies RL to industrial cyberattack mitigation.

Brain2Semantics2Text decodes speech from non-invasive neural recordings via semantic embeddings rather than acoustic/lexical targets.

GANDR: two-agent system for per-claim verification in LLM legal answers; Drafter writes structured reasoning, Auditor verifies citations.

Annealable soft-prior Transformers: smooth annealing preserves retrieval circuits after positional prior removal; hard switching fails.

Anthropic researcher Jacob Coxon resigned over AI extinction fears, calling for pacing agreements between labs.

Genetic algorithm approach for consensus Bayesian Network fusion under treewidth constraints; reconciles multiple input networks.

KVShareArena enables KV-cache reuse across contexts and model checkpoints via position/value repair for RAG and multi-agent workloads.

Maverick: private and verifiable LLM inference via matrix-vector multiplication delegation; reduces server overhead vs. prior cryptographic approaches.

Users can ask the assistant to do things like "Create a cart for my Saturday tailgate for 25 people and include some brunch items," or "Build a cart for easy school lunches and after-school snacks," Shipt says.

RD-Forget framework separates stored vs. queried memories in persistent agents, using frozen LM curator to condition evidence retrieval without retraining.

DiSCo benchmark measures cultural preference bias in LLMs via distribution-first evaluation, avoiding single-correct-answer assumptions across culturally grounded tasks.

A-JIT replaces static software binaries with AI-agent-driven dynamic systems that evolve at runtime based on execution traces, analogous to JIT compilation.

4B VLM fine-tuned as per-token classifier with two-token features, ensembled with 400B zero-shot judge for multimodal hallucination detection on SHROOM-Visions task.

LiteRAG uses query-conditioned graph traversal instead of LLM-controlled retrieval, cutting latency 100x while achieving 0.798 quality on multi-hop QA over distributed-systems papers.

Systematic evaluation of four RAG pipeline design choices (answer-path inclusion, triple syntax, order, grounding instruction) across 6 LLMs and 2 KGQA benchmarks.

Φ-Bench evaluates LLM capability on open-ended LLM infrastructure engineering tasks—kernel optimization, operator design, compiler tuning—beyond isolated benchmarks.

Policy-guided embedding search for tabular feature transformation learns hierarchical, permutation-invariant transformations via gradient-free optimization.

Exact policy optimization for genomic tool selection replaces GRPO sampling with enumeration over combinatorially small tool-subset space, improving scientific reasoning tasks.