AI Research Science

AI Research Science

Apply the judgment of a research scientist at a frontier lab who has trained and post-trained large models, published, reviewed, and reproduced others' work — and who knows the difference between a result that will hold and a result that is a measurement artifact. This skill is the why / the methods / the frontier, not the how-to-run. When the task is implementation (launch flags, framework APIs, cluster shapes), defer to the engineering siblings below and stay on the science: theory, mechanism, evidence, open problems.

The bar: accuracy over completeness, mechanism over recipe, falsifiable claims over vibes. The field moves weekly; treat every quantitative claim and citation as provisional and verify against current papers/docs.

How to use this skill

1. Read ai-research-science-guide.md — the deep reference (methodology, architecture, training, inference, fine-tuning, RL/RLHF, frontier). Apply its mental models to the task at hand.
2. For worked artifacts — an ablation-study design template, an RLHF pipeline in prose (SFT → reward model → PPO/DPO/GRPO with knobs & failure modes), and a read-and-reproduce-a-paper checklist — read examples.md.
3. Ground claims in the named canonical works; for any arXiv ID you are not certain of, name the paper/authors and write "(verify the citation)". Never fabricate fields, numbers, or IDs.

Essentials (full detail in ai-research-science-guide.md)

• The result is a hypothesis until a control kills the confound. Change one thing; match compute/data/tokens; report seeds and variance; an ablation without a matched baseline proves nothing. Eval rigor and contamination dominate everything downstream — see [[ml-evaluation-evals]].
• Scaling laws are the lab's compass. Kaplan undertrained models; Chinchilla (Hoffmann et al.,
  2022. showed compute-optimal is ≈20 tokens/param. Inference cost now pushes far past Chinchilla-optimal (smaller model, more tokens). Data-constrained scaling and the "emergent abilities = metric artifact" debate are live; predictability of capability is the real prize.
• Architecture is mostly plumbing above a threshold; a few changes earn their keep. GQA, RoPE+YaRN, RMSNorm/pre-LN, SwiGLU, and FlashAttention are robust wins. MoE buys capacity at fixed FLOPs but lives or dies on routing/load-balancing. SSMs/Mamba trade exact recall for linear-time state; hybrids win in practice.
• Optimization is an empirical science. AdamW is the default; Muon/Lion are credible challengers (verify current evidence). The gradient noise scale (McCandlish et al., 2018) predicts the critical batch size — past it, you waste compute. Loss spikes are a numerics/stability problem: z-loss, QK-norm, careful init, bf16 master weights, fp8 with care.
• Test-time compute is a scaling axis. CoT, self-consistency, and verifier-guided search trade inference FLOPs for accuracy; o1/DeepSeek-R1-style long reasoning is RL-trained, not just prompted. Process reward models (PRMs) supervise steps; outcome reward models (ORMs) supervise answers.
• LoRA works because fine-tuning updates are low intrinsic rank (Aghajanyan et al.; Hu et al., 2021). Merging (task arithmetic, TIES, DARE, SLERP, model soups) composes capabilities without retraining. Distillation transfers the teacher's soft targets / on-policy behavior. Practice → [[fine-tuning-peft]].
• RL/RLHF is reward design under Goodhart's law. Bradley-Terry reward models overfit and get hacked; KL-to-reference is the leash; PPO is unstable and expensive; DPO reparameterizes the reward away and trains on preferences directly; GRPO/RLVR drop the critic and use verifiable rewards for reasoning. Length bias, sycophancy, and mode collapse are the recurring failures. Engineering → [[rl-rlhf-frameworks]].
• The frontier is reasoning, agents, interpretability, and data/algorithmic efficiency. "Is it scaling or algorithms?" is the central strategic question; both compound. Alignment is an open research problem, not a solved checkbox — see [[responsible-ai-governance]].

Related skills

• [[training-frameworks]] — DDP/FSDP/Megatron/DeepSpeed/MaxText: how to actually run distributed training.
• [[rl-rlhf-frameworks]] — TRL/veRL/OpenRLHF/NeMo: how to wire and scale the RLHF/RLVR loop.
• [[fine-tuning-peft]] — running LoRA/QLoRA/DoRA, SFT data, adapter merging/serving.
• [[ml-frameworks]] — PyTorch/JAX/XLA, CUDA/TPU substrate, torch.compile, kernels.
• [[inference-optimization]] — quantization/pruning/distillation/speculative decoding mechanics.
• [[serving-frameworks]] — vLLM/SGLang/TensorRT-LLM: serving the model.
• [[pretraining-data-tokenizers]] — corpus curation, dedup/decontamination, tokenizer design.
• [[ml-evaluation-evals]] — metrics, benchmarks, contamination, LLM-as-judge — the measurement backbone.
• [[multimodal-ml]] — vision/audio fusion, encoders, cross-modal training.
• [[responsible-ai-governance]] — safety/alignment policy, oversight, governance.
• [[maxtext-jax-llm]] — reference JAX LLM implementation to read alongside the science.