The Archive

Analysis and fix for token repetition failures in diffusion-based multimodal LLMs using information flow tracing and cache optimization.

DiffPT pretraining method using diffusion models to improve classifier performance on limited domain-specific data.

Annotrace: tool for structured step-level annotation of LLM agent execution traces across multiple trace formats.

VERITAS method detects chain-of-thought reasoning faithfulness by tracing layer-wise model computation evolution during generation.

DeDiCa: causal inference framework for discrete variables with hidden confounding using tensorized Bayesian networks.

NestedVLA consolidates multi-task robotic skills via hypernetwork-based parameter generation instead of parameter merging.

JointCore method and JointCoreBench benchmark for interaction-aware memory selection in LLM agents, identifying jointly necessary evidence.

ADT decouples token granularity from reconstruction resolution in 3D molecular diffusion tokenization for proteins and small molecules.

Self-Instructed CoT Controllability shows reasoning models can control chain-of-thought properties when given internal instruction constraints rather than user prompts.

Online multi-agent preference learning embeds preference reasoning into response models rather than using scalar reward scores for RLHF alignment.

Framework for optimizing inference-time compute allocation between sampling, mixing, and verification given imperfect verifier acceptance rates.

Audit of 11 open-weight agents reveals prompt injection failures are ignored (76-80%) rather than refused, masking security vulnerabilities in benchmarks.

Agentic AI system for optical network troubleshooting integrates data references to prevent hallucinations and improve root-cause analysis.

FLARE: physics-grounded benchmark for evaluating LLM agents on electrical grid feeder restoration with incomplete observations across 78,750 episodes.

Systematic evaluation of confidence-based small-to-large LLM routing across 72 configurations and 42K predictions shows mixed multilingual reliability.

Counterfactual pairing cycles benchmark method isolates relational OOD errors in vision-language models by controlling caption style and image statistics.

Token-budget theory formalizes read-vs.-reason trade-off in reasoning systems with shared context, modeling accuracy as function of input/reasoning token allocation.

MGE-PO improves RLHF credit assignment for math reasoning by pairing correctness with generation entropy to reward low-entropy successes and penalize high-entropy failures.

Mixture-Gated Theory of Anisotropic Representation Dynamics explains how Transformer self-attention layers grow or shrink representation differences via value transport and gating mechanisms.

RIVET enables turn-level routing in multi-turn LLM interactions by using post-call response evidence during training while keeping deployment decisions pre-call.

MultiAgent-OPD combines on-policy distillation with RL to jointly train hierarchical multi-agent systems (Leader-Searcher-Answerer) end-to-end using LLM judge feedback.

Conceptual framework characterizing and analyzing RAG techniques across multiple dimensions to guide selection for specific use cases.

ReLaDrift enables iterative refinement in one-step diffusion-based language generation by recurrently refining drafts in continuous latent space with gradient stopping.

Bayesian framework for Process Reward Models jointly addresses PRM uncertainty from insufficient data and policy uncertainty from distribution mismatch in test-time scaling.

qGAMMA framework optimizes mixed-precision DNN mapping onto spatial accelerators using genetic algorithms and analytical performance models.

DIA shows test-time compute from single models provides correlated rather than independent evidence; proposes disagreement-triggered independent arbitration across models.

MAS-Bai uses linguistically informed multi-agent routing for Bowen-script Bai-Chinese translation, addressing phonetic character borrowing hallucinations.

Non-parametric stochastic optimization framework using Malliavin calculus for time-varying decision problems.

Evaluation framework validating LLM-generated social networks against survey data from ANES 2020-2022 study.

Free energy principle model for spatial hearing and sound localization under dynamic acoustic conditions.