The Archive

DIGEST combines LLM-guided search with adaptive generation for symbolic regression, reusing prior search insights to improve mathematical expression discovery.

FlashScript compiles successful agent runs into reusable neuro-symbolic programs, replacing repeated reasoning with parameterized code and LLM decision points.

92B coding LLM trained via SE expertise and RefactorForge pipeline for repository-level refactoring, deployed to 100k+ developers.

AutoTavern LLM research agent with typed memory and verification gates builds Hearthstone Battlegrounds AI matching top-human performance.

OPAL LLM agent dynamically rewrites optimizer code mid-training to adapt schedules and update rules without meta-training.

Proposes instruction-tuned LLMs implement a distinct genre combining conversational and formal discursive patterns.

Foyer compares typed decision models, LLMs, and baselines on 739 multilingual emails using stratified sampling and human labels for inbox routing.

BF16 RL training instability in LLMs solved by aligning rounding locations in RoPE, SwiGLU, and softmax across inference-training stack.

Game-theoretic analysis of LLM platform pricing under mixed ad-supported and subscription models with endogenous quality.

Empirical study across 11 open-source LLMs shows LLM-based code self-refinement may not reduce energy costs or offset training overhead despite performance gains.

Multi-step transition prediction in finite-horizon MDPs under block-feedback yields Bayesian-optimal value functions via Bellman theory.

LEMS and KFAC-SVD improve LLM compression via layer-wise error modeling and Fisher-based rank allocation for dynamic compression.

Study reveals LLM vulnerability auditors exhibit asymmetric deference to static-analysis evidence, accepting false assurances more readily than questioning false alarms.

LLM-agent opinion simulations on Qwen3, Llama-3.2, Gemma-3 fail complement consistency when reasoning over own vs. neighbor states.

J-Comm enables latent communication between LLM agents via GEAR token reweighting and H+J interface exposing hidden states and Jacobian gradients.

Proposes five-level communication architecture claiming LLM context alone cannot ground utterance formation for ASI; requires external schemas.

Study quantifies demographic biases in LLM-emulated voter representation and evaluates alignment techniques to improve fairness across voter groups.

Self-training framework enables post-training LLM improvements without external supervision using curriculum guidance.

ActionRAG couples entity indexing with graph diffusion retrieval to maintain factual consistency in LLM agent long-term memory.

LGMT: open-source tool detects logical inconsistencies in LLM reasoning using metamorphic testing grounded in first-order logic.

LLM agents reuse docking probes and edits across protein pockets for structure-based drug design via experience replay.

PHYRD benchmark evaluates LLM and agent performance on 106 theoretical physics reproduction and discovery tasks, achieving 19-74% accuracy.

MCP-Injector fault-injection framework reveals LLM agent vulnerabilities to tool failures via Model Context Protocol; scale does not guarantee resilience.

Accordion enables variable-cost LLM inference via mixture-of-experts layer looping, eliminating need to maintain separate model checkpoints.

Framework decomposes LLM agent failures into model vs. harness contributions using Valid Completion Rate and Model-Harness Fit metrics.

ALLMTS grounds LLM medical triage decisions in telehealth guidelines to reduce over-escalation and improve auditability in emergency dispatch.

Sharpness-aware minimization combined with Muon optimizer improves LLM training generalization under heavy-tailed gradient noise.

SCI-AGENT-EXAM benchmarks LLM agents on 120 expert-curated scientific research tasks across 11 fields, revealing capability gaps in realistic workflows.

Async-RBench benchmarks asynchronous task execution in LLM agents with 200 executable tasks across eight scenarios testing state consistency.

ANT-A benchmark measures LLM annotator consistency against human agreement across 7 datasets, identifying gaps in multi-agent annotation systems.