🇰🇷 한국어

• [Paper Review] Accurate Decoding of Natural Sentences from Non-Invasive Brain Recordings

  Meta AI's Brain2Qwerty v2 collects 90 hours and 2,724 unique sentences of MEG from nine people, then chains a CTC encoder, a word aligner, and a LoRA-tuned LLM to decode natural sentences from non-invasive brain signals at 39% WER

  43 min read   ·   August 24, 2026

  2026   ·   brain-computer-interface   meg   brain-to-text   ctc   llm   lora   neuroscience   ·   paper-review

  

• [Paper Review] Agentic Method for Deterministic Validation of Legacy Code Migration

  American Express's Locksmith Loop: when input search stalls during COBOL-to-Java migration validation, mutate the harness itself to open new execution regions, apply the mutation symmetrically to both languages, and let a deterministic parity oracle decide what survives

  40 min read   ·   August 21, 2026

  2026   ·   software-testing   agentic-systems   legacy-migration   cobol   mutation-testing   differential-testing   search-based-testing   ·   paper-review

  

• [Paper Review] RAGU: A Multi-Step GraphRAG Engine with a Compact Domain-Adapted LLM

  A GraphRAG engine that separates extraction from consolidation, and a scaling hypothesis that justifies running the whole indexing pipeline on a 7B model

  34 min read   ·   August 20, 2026

  2026   ·   graphrag   retrieval-augmented-generation   knowledge-graph   information-extraction   small-language-models   open-source   ·   paper-review

  

• [Paper Review] Hierarchical Self-Improvement: A Framework for Task-Specific Evolvable Agent Harnesses

  Freeze the model, evolve only the harness across three nested scopes — +39.3 on BabyAI and +33.0 on Crafter, plus an honest account of exactly where the gains stop

  38 min read   ·   August 14, 2026

  2026   ·   llm-agents   self-improvement   agent-harness   meta-evolution   balrog   frozen-backbone   ·   paper-review

  

• [Paper Review] NVIDIA-labs OO Agents: Native Python Object-Oriented Agents

  An agent is not a bundle of prompt templates, tool schemas, and workflow graphs. It is a Python object: methods are actions, fields are state, docstrings are prompts, type annotations are contracts.

  47 min read   ·   August 12, 2026

  2026   ·   llm-agents   agent-framework   code-as-action   tool-use   python   memory   benchmarks   ·   paper-review

  

• [Paper Review] Skaling: Chinchilla's Exponents Meet Kaplan's Coupling

  Chinchilla's additive scaling law hard-codes the assumption that model size and data act independently. Measured mixed derivatives say otherwise, and one extra exponent cuts boundary extrapolation error by more than 10x.

  38 min read   ·   August 11, 2026

  2026   ·   scaling-laws   llm   pretraining   compute-optimal   chinchilla   extrapolation   ·   paper-review

  

• [Paper Review] ScientistOne: Towards Human-Level Autonomous Research via Chain-of-Evidence

  Papers from autonomous research agents read well but their evidence chains are broken. An audit of 75 papers exposes a systematic failure in every baseline, and a system that keeps every claim tied to its evidence while writing.

  55 min read   ·   August 04, 2026

  2026   ·   autonomous-research   llm-agents   verifiability   hallucination   evaluation   ai-scientist   ·   paper-review

  

• [Paper Review] Frontis-MA1: Training an AI4AI Model towards Recursive Self-Improvement in Machine Learning Engineering

  Four atomic operators — Draft, Improve, Debug, Crossover — shared between post-training and evolutionary search. A 35B model hits 71.21% on MLE-Bench Lite on a single RTX 4090 with a 12-hour per-task budget.

  46 min read   ·   August 03, 2026

  2026   ·   ai4ai   recursive-self-improvement   mle-agent   evolutionary-search   reinforcement-learning   post-training   agent   ·   paper-review

  

• [Paper Review] Kimi K3: Open Frontier Intelligence

  A 2.8T-parameter, 104B-active, 1M-context open-weight MoE. Kimi Delta Attention, Attention Residuals, and Stable LatentMoE push scaling efficiency 2.5x over Kimi K2 to reach the frontier.

  34 min read   ·   July 31, 2026

  2026   ·   mixture-of-experts   linear-attention   long-context   reinforcement-learning   multimodal   open-weights   llm-infrastructure   ·   paper-review

  

• [Paper Review] Autodata: An agentic data scientist to create high quality synthetic data

  An LLM agent that acts as a data scientist — creating synthetic data, evaluating it, and revising its recipe in a loop — plus how to meta-optimize the agent itself.

  29 min read   ·   June 29, 2026

  2026   ·   synthetic-data   llm-agents   self-instruct   reinforcement-learning   grpo   data-generation   ·   paper-review