The Archive

LLM-based agents autonomously conduct quantum-spin experiments with peer review and theory validation to prevent unsupported conclusions.

Token-adaptive self-distillation assigns per-token teachers during on-policy LLM post-training, improving over fixed or response-level teacher approaches.

Comparative study of Mixtral 8x7b vs. human role-players in online counselling training using communication pattern analysis.

ClinGeX LLM agent with 105 genomic skills achieves 71.7% concordance on clinical variant interpretation via task-specific retrieval and orchestration.

Theoretical analysis reveals fundamental limits on predicting decode token length during LLM prefill phase, questioning benchmark validity.

Ridge regression identifies brain-predictive subspace in LLM representations during naturalistic language listening via EEG and fMRI.

Agentic University deploys multi-agent LLM coordination for personalized tutoring, content generation, and adaptive learning path tracking.

rTaste benchmark evaluates frontier LLM research agents' ability to judge experiment value; finds sound MVP proposals but weak novelty discrimination.

Orthogonal projection method enables continual LLM unlearning of sensitive data without utility degradation across sequential deletion requests.

NVIDIA Nemotron 3.5 Super VL is a 120B-parameter mixture-of-experts model that pairs reasoning and tool use with multimodal understanding. This post explores its architecture and use cases, then shows how to deploy it on Amazon EKS, EKS Auto Mode, and Amazon SageMaker HyperPod EKS with NVIDIA Dynamo and vLLM.

LLM Voice product or feature announcement lacks sufficient technical or business detail.

Sometimes when I read a post, I get the growing feeling that I'm reading the output of an LLM. I can't necessarily define why, although there are the common tells (it's not the em-dashes, it's the load-bearing tapestries). As I do this my desire to read on atrophies, sometimes to the point of revulsion There are tools out there that claim to be able to tell the difference between LLM-written prose and the old-fashioned biological product. I'm wary of them, partly due skeptism about their accuracy, but also because it rather misses the point. The reaction to LLM-Voice is a human one, that sens...

HEMO: self-evolving LLM-based operator design for multi-objective combinatorial optimization treating the search harness as an optimization target.

Thirteen design patterns cataloging LLM-rule engine coupling methods across agent frameworks, guardrails, and neuro-symbolic AI with applicability guidance.

MTG-AD dataset and evaluation framework for multimodal LLM-generated inline audio descriptions in user-generated video.

Framework measuring knowledge leakage in persona-conditioned LLM responses during user simulation tasks.

H²R: hierarchical hint-based reward framework using reinforcement learning for multi-step mathematical reasoning in LLMs.

Turing LLM post-training integrates MoE specialists via calibrated multi-teacher distillation and bounded-width dispatch on Turing-20B-A2B.

Enactive theory framework analyzes human-LLM interactions as embodied linguistic encounters, challenging autonomy and social cognition.

DWISE proposes workload-informed scheduling for diffusion LLM serving with heterogeneous decoding cost awareness replacing FCFS.

Evidence-gated continual learning prevents LLMs from absorbing falsehoods at higher rates than random selection via surprise gating.

EvoTeam framework optimizes multi-agent LLM team structure and interactions through collective execution across heterogeneous tasks.

InferD honeypot measures security vulnerabilities in exposed LLM inference APIs, vector databases, and agent orchestration systems.

Representation learning analysis shows harmful and harmless queries are distinguishable at intermediate LLM layers, enabling over-refusal reduction.

Fine-tuning open-weight models with five self-concept variants reveals how LLMs interpret and express identity.

Framework for evaluating LLM agent memory, re-engagement, and context reconstruction in persistent social environments.

Activation probes detect rare AI misalignments with few labeled examples via LLM-assisted upsampling on safety tasks.

Late-fusion module uses supervised LLM to aggregate heterogeneous echocardiography model outputs for cardiac imaging.

ATLAS-Bench: 1,200-query benchmark for evaluating LLM agents on real-world local place search tasks.

LLM-guided SMT solver for quantified array logic with verifiable focal term extraction.