LLM Post-Training Paper Digests

A topic-organized synthesis of 99 post-training research papers and technical reports on LLM post-training — covering data curation, training curriculum, post-training methods (preference optimization, PPO/GRPO/DAPO, process reward models, distillation, self-play, agent RL), and evaluation & metrics — spanning arXiv-ID years 2022 → 2026.

Each paper is summarized across four lenses and the results are grouped into four topic files, sorted by the paper's arXiv-ID year (2022→2026). A chronological evolution narrative ties the eras together.

Repository layout

.
├── pdfs/                 # 99 source PDFs (code_ / data_ / method_ / report_ / rl_ prefixes)
├── texts/                # 99 extracted full-text files — the source material for the digests
├── summaries/            # the synthesis
│   ├── 01-data-curation.md
│   ├── 02-training-curriculum.md
│   ├── 03-post-training-methods.md
│   ├── 04-evaluation-metrics.md
│   └── EVOLUTION_OVERVIEW.md
├── digests/              # 99 structured per-paper JSON digests (machine-readable supplement)
└── README.md             # you are here

The summaries

File	Lens	Papers with substantive content
summaries/01-data-curation.md	Data curation — sourcing, generation, filtering, de-dup, quality control, scaling; datasets, sizes, selection criteria	88 / 99
summaries/02-training-curriculum.md	Training curriculum — stage ordering (pretrain→SFT→preference→RL→distillation), multi-stage RL, self-play loops, warmup/anneal, cross-stage data mixing	67 / 99
summaries/03-post-training-methods.md	Post-training methods — preference optimization, RL algorithms (PPO/GRPO/DAPO/…), process reward models, distillation, self-play, agent loops, infra	99 / 99
summaries/04-evaluation-metrics.md	Evaluation & metrics — benchmarks and metrics (AIME, MATH, HumanEval, SWE-bench, Arena, win rate, pass@k, avg@k, …) and headline results	99 / 99
summaries/EVOLUTION_OVERVIEW.md	Chronological evolution narrative — older trends, how the field evolved, why newer methods were proposed, and the most-recent (2025–2026) technology shift	all 99

The digests/ folder holds the structured per-paper digests (one JSON per paper, with all four lenses, key results, and an evolution note) that back these summaries — kept as a machine-readable supplement.

How to read each topic file

• Papers appear under ## <Year> headings in ascending year order (2022 → 2026), preserving corpus order within a year. The "2026" entries are dated by their arXiv identifier in this corpus; that is a corpus date, not a claim about real-world publication status.
• Each paper entry is a synthesized summary (prose + compact bullets) of one lens only, followed by a one-line Key results: and a one-line italic Evolution: note that places the paper in the timeline.
• Frontier technical reports (DeepSeek-R1/V3/V3.2/V4, Kimi K1.5/K2/K3, Qwen2.5/3/3.5-Omni/3-Coder-Next, GLM-4.5/5/5V, Llama3, Gemma2/3, Phi-4, Seed1.5, …) span all four lenses, so they appear in every topic file where they contribute; method/data/RL papers appear under their primary lens.

How these summaries were generated

• Source: built solely from the full-text files in texts/.
• Style: synthesized prose + bullets in the model's own words (concise, concrete, skimmable).
• Organization: aggregate-by-topic layout (four lens files, each sorted by year) plus a dedicated summaries/EVOLUTION_OVERVIEW.md that traces older trends → newer methods → rationale, with explicit emphasis on the most-recent (2025–2026) technology evolution.
• Evolution framing: every entry carries an Evolution: line, and the overview ties the eras together (surface alignment 2022–2023 → verifiable reasoning RL 2024–2025 → environment-grounded, self-evolving agents + efficient reasoning 2026).

Corpus at a glance

• 99 papers, years from arXiv IDs: 2022 (3), 2023 (14), 2024 (30), 2025 (31), 2026 (21).
• Filename prefixes: code_ (17), data_ (21), method_ (22), report_ (25), rl_ (14). Each filename embeds the arXiv ID (e.g. rl_DAPO_2503.14476.txt → arXiv 2503.14476, year 2025).