A structured library of 121,245 machine-learning paper analyses covering arXiv publications from 2023–2025. Each paper is distilled into a standardised markdown report so you can survey findings, compare mechanisms, and spot trends without reading thousands of PDFs.

Every file follows the same template. Here's a trimmed example (2511.21730_a-benchmark-for-procedural-memory-retrieval_…md):

# frontmatter
arxiv_id: '2511.21730'
core_contribution: >
  Introduces the first benchmark for evaluating procedural
  memory retrieval in language agents, isolating retrieval
  from execution …
tags: [procedural, retrieval, memory, …]   # ⚠ see caveat below

## Quick Facts          — arXiv link, authors, headline numbers
## Executive Summary    — what the paper does in one paragraph
## Method Summary       — experimental setup, models, data
## Key Results          — quantitative findings
## Mechanism Analysis   — *why* the approach works (multiple sub-sections)
## Reproduction Notes   — hyperparameters, compute, data details
## Limitations & Confidence

Tag caveat: Auto-generated tags are noisy (every record shares a long generic tail). Prefer searching core_contribution, titles, and body text. Tag regeneration is planned.

Pick a year folder and search with ripgrep:

# Find all papers mentioning mixture-of-experts
rg -l "mixture of experts|MoE" ml_research_analysis_2025/

# Full-text search across every year
rg -n "speculative decoding" ml_research_analysis_202*/

The search script works around noisy tags by matching across title, core_contribution, and filename:

python scripts/search_topic.py --topic "mixture of experts" --alias moe
python scripts/search_topic.py --topic "reinforcement learning" --alias rl --limit 25 --json

analysis_outputs/research_index.sqlite indexes the 2025 bucket (52,099 rows) with columns: title, arxiv_id, core_contribution, tags, filename, file_size.

# papers whose core contribution mentions "distillation"
sqlite3 analysis_outputs/research_index.sqlite \
  "SELECT arxiv_id, title FROM papers WHERE core_contribution LIKE '%distillation%' LIMIT 10"

# look up a specific paper
sqlite3 analysis_outputs/research_index.sqlite \
  "SELECT filename FROM papers WHERE arxiv_id = '1706.03762'"

The spot_analyses/ directory and the spot_analysis_paper_groups table contain deep-dive clusters across eight research themes:

The website/ directory contains a static site with full-text search. See website/README.md for build and deploy instructions.

ml_research_analysis_2023/   Per-paper markdown analyses
ml_research_analysis_2024/
ml_research_analysis_2025/
analysis_outputs/            SQLite index, digests, assessment outputs
scripts/                     index_frontmatter.py, search_topic.py
spot_analyses/               Curated topic deep-dives (8 groups, 1,824 papers)
website/                     Static browse/search UI
docs/                        Internal reference documents
archive/                     Superseded v1 analyses

A three-phase FlatAgents pipeline produces each report:

1. Prep — download arXiv PDF, extract text, match against ML terminology corpus
2. Expensive — parallel LLM calls for mechanism analysis, reproduction notes, and open questions
3. Wrap — limitations/confidence, tagging, report assembly, quality judge + auto-repair

The 2025 batch used GLM-5 (pony-alpha) for the expensive phase; 2023–2024 used Trinity Large throughout. Pipeline code, configs, and execution databases live in the pipeline repo — this repository is output only.

• ~190 permanent failures across all years: PDF 404s (~106), context overflow >256k (~60), provider errors (~9), PDF parse errors (~15). No pending retries.
• Tags are unreliable — the tail of every tag list contains generic terms. Use core_contribution and full-text search instead.
• Duplicate filenames exist where papers were rerun (1,158 in 2024, 582 in 2025). The SQLite index and filenames are deduplicated by (arxiv_id, timestamp).

After adding or removing analysis files, rebuild the SQLite index:

python scripts/index_frontmatter.py ml_research_analysis_2025
python scripts/index_frontmatter.py ml_research_analysis_2025 --prune  # also remove deleted files