Mieaz/meituna-longcat-replications

MeituNa LongCat Open-Ended Paper Replication Portfolio

Generated: 2026-07-31T12:15:27.895345+00:00

Multi-instance Meituan LongCat-2.0 workers produce multi-file open-source replication scaffolds (method breakdown, code modules, experiment plans, scientific notes) grounded in real arXiv paper text — not single-file stubs.

Run metadata

• Parents: 8
• Papers: 16
• Model: LongCat-2.0
• Passes: ['replication_plan', 'code_modules', 'experiments_and_critique', 'scientific_notes']
• Token usage: {"prompt_tokens": 0, "completion_tokens": 0, "total_tokens": 0, "calls": 0}

Executive summary

This portfolio targets a critical inflection point in large language model research: the demonstration that data quality, algorithmic ingenuity, and inference-time strategies can substitute for raw parameter scale. The papers cluster around three pillars—data-centric small model training (Phi-3, phi-1.5, Qwen2.5-Coder), efficient adaptation and alignment (QLoRA, DPO), and inference-time reasoning and prompting (ICL, PAL, EvoPrompt, Lost in the Middle)—that collectively challenge the assumption that capability requires billion-dollar training runs. The central thesis is that carefully filtered web data, synthetic textbook data, and code-interpreter augmentation can enable models in the 1B–7B parameter range to rival or exceed systems 5x–10x their size, while quantization and preference optimization reduce the compute barrier for post-training by orders of magnitude.

The cross-paper algorithmic connections are dense and mutually reinforcing. DPO provides the alignment backbone for Phi-3's post-training; QLoRA enables single-GPU finetuning of the very models that Phi-3 and Qwen2.5-Coder demonstrate are viable; PAL's code-execution strategy offers a complementary reasoning path to phi-1.5's synthetic-data approach; and the ICL survey's demonstration-ordering findings directly inform how prompts should be structured for both PAL and EvoPrompt. Lost in the Middle serves as a cautionary boundary condition: the U-shaped retrieval curve constrains how much context can be exploited for repo-level code training (Qwen2.5-Coder) and in-context demonstration design.

The recommended replication strategy prioritizes algorithmic components with high modularity and low data dependency. DPO and QLoRA should be reimplemented first because they are self-contained, well-specified, and unlock downstream replications—QLoRA enables finetuning of small models on consumer hardware, while DPO simplifies the alignment pipeline that Phi-3 relies on. PAL and the ICL survey follow as inference-time methods that can be validated with open-weight models and require no training. Full pretraining replications (Phi-3, Qwen2.5-Coder) are long-horizon goals whose data-filtering heuristics remain partially proprietary, making them better suited for ablation studies on publicly available corpora.

Projects

✓ 2404.14219 — Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone

• Slug: 2404_14219_phi-3-technical-report-a-highly-capable-language
• Files: 6 (56028 bytes)
• Plan: Train a 3.8B model on 3.3T tokens of heavily filtered web and synthetic data using a two-phase pre-training and SFT+DPO post-training to rival GPT-3.5/Mixtral 8x7B.
• Reproducibility score: 7

✓ 2301.00234 — A Survey on In-context Learning

• Slug: 2301_00234_a-survey-on-in-context-learning
• Files: 8 (40723 bytes)
• Plan: Survey of in-context learning (ICL) covering formal definitions, model warmup, demonstration designing (selection, ordering, formatting), scoring functions, and analysis of influencing factors.
• Reproducibility score: 7

✓ 2305.18290 — Direct Preference Optimization: Your Language Model is Secretly a Reward Model

• Slug: 2305_18290_direct-preference-optimization-your-language-mod
• Files: 8 (29833 bytes)
• Plan: DPO eliminates RL from preference learning by reparameterizing the reward model so the optimal policy is recoverable in closed form, reducing RLHF to a simple binary cross-entropy classification loss on preference pairs.
• Errors: critique: empty extract

✓ 2307.03172 — Lost in the Middle: How Language Models Use Long Contexts

• Slug: 2307_03172_lost-in-the-middle-how-language-models-use-long-
• Files: 8 (43496 bytes)
• Plan: Replicate core method of Lost in the Middle: How Language Models Use Long Contexts
• Reproducibility score: 8
• Errors: plan: could not parse JSON object from model output: Unterminated string starting at: line 8 column 7 (char 418); notes: empty extract

✓ 2305.14314 — QLoRA: Efficient Finetuning of Quantized LLMs

• Slug: 2305_14314_qlora-efficient-finetuning-of-quantized-llms
• Files: 8 (54850 bytes)
• Plan: QLoRA finetunes 4-bit quantized LLMs via LoRA adapters, using NF4 quantization, Double Quantization, and Paged Optimizers to fit 65B models on a single 48GB GPU while matching 16-bit performance.
• Reproducibility score: 7

✓ 2309.05463 — Textbooks Are All You Need II: phi-1.5 technical report

• Slug: 2309_05463_textbooks-are-all-you-need-ii-phi-1-5-technical-
• Files: 8 (50954 bytes)
• Plan: phi-1.5 is a 1.3B parameter Transformer trained on 150B tokens of mostly synthetic 'textbook-quality' data, achieving reasoning performance comparable to models 5x-10x larger.
• Reproducibility score: 7

✓ 2211.10435 — PAL: Program-aided Language Models

• Slug: 2211_10435_pal-program-aided-language-models
• Files: 9 (61841 bytes)
• Plan: PaL uses an LLM to generate interleaved natural language and Python code reasoning steps, offloading the final solution to a Python interpreter to avoid arithmetic and logical calculation errors.
• Reproducibility score: 7

✓ 2409.12186 — Qwen2.5-Coder Technical Report

• Slug: 2409_12186_qwen2-5-coder-technical-report
• Files: 5 (55056 bytes)
• Plan: Replicate core method of Qwen2.5-Coder Technical Report
• Reproducibility score: 7
• Errors: plan: could not parse JSON object from model output: Expecting ',' delimiter: line 23 column 4 (char 1717)

✓ 2309.08532 — EvoPrompt: Connecting LLMs with Evolutionary Algorithms Yields Powerful Prompt Optimizers

• Slug: 2309_08532_evoprompt-connecting-llms-with-evolutionary-algo
• Files: 6 (57223 bytes)
• Plan: Replicate core method of EvoPrompt: Connecting LLMs with Evolutionary Algorithms Yields Powerful Prompt Optimizers
• Reproducibility score: 7
• Errors: plan: could not parse JSON object from model output: Expecting ',' delimiter: line 19 column 4 (char 947)

✓ 2407.10671 — Qwen2 Technical Report

• Slug: 2407_10671_qwen2-technical-report
• Files: 7 (59587 bytes)
• Plan: The Qwen2 series introduces dense (0.5B-72B) and MoE (57B-A14B) LLMs pre-trained on 7T tokens, featuring GQA, Dual Chunk Attention with YARN for long context, and scalable post-training via SFT and DPO.
• Reproducibility score: 7

✓ 2402.03300 — DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models

• Slug: 2402_03300_deepseekmath-pushing-the-limits-of-mathematical-
• Files: 8 (55777 bytes)
• Plan: Continually pretrain a code base model on 120B math tokens mined from Common Crawl via an iterative fastText pipeline, then apply instruction tuning and GRPO reinforcement learning to achieve state-of-the-art open-source math reasoning.
• Reproducibility score: 7

✓ 2310.06825 — Mistral 7B

• Slug: 2310_06825_mistral-7b
• Files: 10 (51575 bytes)
• Plan: Mistral 7B is a 7B parameter transformer leveraging Sliding Window Attention (SWA) and Grouped-Query Attention (GQA) to outperform larger models like Llama 2 13B while improving inference efficiency.
• Reproducibility score: 7

✓ 2312.00752 — Mamba: Linear-Time Sequence Modeling with Selective State Spaces

• Slug: 2312_00752_mamba-linear-time-sequence-modeling-with-selecti
• Files: 8 (13798 bytes)
• Plan: Mamba replaces attention with input-dependent selective state spaces, achieving linear-time scaling and 5x inference throughput over Transformers while matching or exceeding Transformer quality across language, audio, and genomics.
• Reproducibility score: 7

✓ 2302.13971 — LLaMA: Open and Efficient Foundation Language Models

• Slug: 2302_13971_llama-open-and-efficient-foundation-language-mod
• Files: 9 (61203 bytes)
• Plan: LLaMA trains 7B-65B parameter transformers on 1.4T tokens from public datasets, using RMSNorm, SwiGLU, and RoPE to achieve state-of-the-art performance competitive with Chinchilla and PaLM.
• Reproducibility score: 7

✓ 2401.14196 — DeepSeek-Coder: When the Large Language Model Meets Programming -- The Rise of Code Intelligence

• Slug: 2401_14196_deepseek-coder-when-the-large-language-model-mee
• Files: 8 (46512 bytes)
• Plan: Open-source code LLMs (1.3B-33B) trained from scratch on 2T tokens with repo-level topological ordering and 50% PSM Fill-in-the-Middle training, achieving SOTA among open-source and surpassing GPT-3.5 Turbo.
• Reproducibility score: 7

✓ 2205.14135 — FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness

• Slug: 2205_14135_flashattention-fast-and-memory-efficient-exact-a
• Files: 6 (61220 bytes)
• Plan: FlashAttention computes exact attention with fewer HBM accesses via tiling and recomputation, enabling faster training and longer sequences.
• Reproducibility score: 7

Priority reimplementations

• 2305.18290 (M): DPO is the most self-contained algorithm in the portfolio: a single loss function, no reward model, no RL loop. It is used in Phi-3's post-training and can be validated on existing preference datasets (Anthropic HH, OpenAI WebGPT) with open SFT checkpoints. Success here unlocks alignment experiments for every other model in the portfolio.
• 2305.14314 (M): QLoRA's NF4 quantization and full-layer LoRA are well-documented and can be validated on LLaMA-2-7B with the released Guanaco checkpoints. It enables single-GPU finetuning of the small models that Phi-3 and Qwen2.5-Coder champion, making it infrastructure for downstream replications.
• 2211.10435 (L): PAL requires no training—only prompt engineering and a Python sandbox. It can be replicated immediately with GPT-3.5/4 or CodeLlama on GSM8K, providing a strong baseline for reasoning tasks and a testbed for EvoPrompt optimization.
• 2301.00234 (L): The ICL survey's scoring functions (Direct, PPL, Channel) and demonstration ordering experiments can be replicated on GPT-2 and open LLMs with SST-2 and GSM8K. This provides the theoretical foundation for understanding why PAL and EvoPrompt prompts work.

Layout

• projects/<slug>/ — multi-file replication tree per paper
• SYNTHESIS.md / PRIORITIES.md — corpus-level insight
• corpus_index.json — machine-readable index
• corpus_synthesis.json — structured synthesis