Related work and reuse decisions

Primary sources checked on 15 October 2026. This is a build-versus-reuse map, not a claim that a single framework or model is currently best. Most commit links below are research snapshots; a link alone is neither a dependency lock nor installation evidence. Explicitly integrated runtime pins live in pyproject.toml, requirements/ and the corresponding setup instructions. Qualification is stated per integration, and a recipe needs a successful hardware run before promotion to a validated cluster recipe.

The OpenEnv/NeMo/FineEnvs comparison and German RAG/workplace benchmark designs add primary-source checks from 20 October 2026. These references do not change the frozen runtime pins.

The smallest useful stack

Layer	Academic / community and industrial work	What we build on	Status here
Reasoning adaptation	TRL, Hugging Face	Existing SFT and GRPO trainers; our contribution is the recipe, data provenance, and evaluation evidence.	CPU SFT/GRPO checks and real German Qwen A100 SFT/agent-RL pilots completed; measured results and limits.
Offline distillation	OpenThoughts, DataComp / Bespoke Labs collaboration	Transparent teacher-data curation and ablations. We implement a small generation/filtering path, not the complete OpenThoughts pipeline.	Implemented local preparation and optional generation; reference methodology.
Recipe collection	NeMo Skills, NVIDIA	Synthetic data → training → evaluation and local/Slurm conventions. Upstream is moving rollout and benchmark implementations into Gym.	Architectural precedent; no dependency.
Existing agent harness RL	Agent Lightning, Microsoft Research	Gateway capture with the deployment harness and a verl-backed trainer.	Optional real launcher and external calculator harness; source-checked and mock-tested, GPU execution unqualified.
Reusable environments	NeMo Gym, NVIDIA	Environment lifecycle and authoritative tool verification.	Actual CPU service/client checks and an eight-step A100 TRL bridge qualification completed; setup, evidence and limits.
Reusable environments	OpenEnv 0.9.0, Hugging Face	Typed stateful environments, persistent SDK WebSocket sessions and trainer adapters. The official reasoning walkthrough explains Qwen plus GRPO.	Native service and SDK/TRL bridge CPU checks pass; matched German arm retains the frozen TRL 0.29 API and protocol. GPU qualification and quality tests pending.
Reusable environments	Verifiers 0.4.0, Prime Intellect	Native tasksets, harnesses, MCP tools and reward lifecycle.	Optional German environment in an isolated stack; scripted native execution is distinct from model training.
Distributed RL	verl / HybridFlow, ByteDance / academic-community collaboration	Distributed optimizer and inference scheduling rather than a new trainer.	Extension backend through a future Lightning pilot.
System and custom evaluation	Inspect, UK AI Security Institute	Existing evaluation SDK and task implementations.	Optional adapter; actual pinned SDK mock execution passed.
Conventional model evaluation	LM Evaluation Harness, EleutherAI	Existing task catalog and model interfaces.	Optional dependency for native upstream CLI use; no report importer implemented.
Evaluation methodology	HELM, Stanford CRFM	Transparent scenarios, multiple metrics, and reproducible reporting.	Methodological reference, no dependency.

OpenThoughts: Data Recipes for Reasoning Models is the principal academic data-recipe reference. The Agent Lightning v1.0 paper is the closest systems reference for preserving an existing harness. These works motivate controlled baselines and data/trajectory contracts; their reported results do not predict gains for participating teams.

Agent Lightning versus NeMo Gym

Start with Lightning + verl when a team must retain an existing orchestration loop. Its current upstream design separates the model gateway, rollout execution, and GPU training. See the architecture documentation. Use Gym when reusable task environments and verifiers are the main shared asset. Its current repository documents environment servers, external harnesses, token capture, and Slurm-oriented execution. These are upstream capabilities. This repository now has a concrete native Gym resource server and external TRL bridge; its narrower checks and limits are documented in RL agent gyms. Token capture in that bridge remains the existing TRL audit rather than an asserted upstream Gym capture feature.

Do not install both distributed stacks in the default environment. Freeze one representative task, policy, reward, and compute budget; compare trace fidelity, recovery after failures, environment isolation, and time to a first reproducible update. The concrete pilot and its acceptance criteria are in agent-harness.md.

Benchmarks: reuse before inventing

The supplied programme proposal now targets German RAG with cited evidence and business-communication agent sequences. The included reasoning and tool-workflow tracks remain synthetic development fixtures. See the benchmark designs for existing German RAG meta-evaluation, citation metrics, workplace simulations and tool-injection suites to reuse; neither new programme benchmark is implemented.

For interactive workflow design, study Sierra's τ²-bench: evaluate task state and user/tool interaction rather than answer text alone. For physical materials predictions, use the native Matbench Discovery protocol rather than converting the problem into LLM question answering.

Each proposed public benchmark must justify a decision existing evaluations cannot support, include expert review and contamination controls, separate public development data from held-out evaluation, and measure reliability across repeated trials. An upstream track is preferable when the gap fits an existing benchmark.

Domain-native references

These are sketches only; there are no working domain adapters in this release.

Domain	Reference snapshot	Reuse boundary
Robotics	LeRobot, Hugging Face	Robot datasets, policy adaptation, and native episode evaluation.
Biomedical agents	Biomni, Stanford	Tool-rich biomedical harness reference; not an RL optimizer.
Medical text / image adaptation	MedGemma, Google Health	Model-specific adaptation examples and intended-use terms.
Biological prediction models	BioNeMo Recipes, NVIDIA	Native biological model workflows and recipe packaging.
Materials / atomistic models	MACE, academic-community project	Equivariant interatomic potentials, native training and simulation evaluation.

The proposed experiments, artifact contracts, and domain-specific evaluation limits are in domains.md.

What remains research-only

These additional references broaden the experimental choices without becoming dependencies:

Reference	Useful experiment or methodology	Status here
AllenAI open-instruct	Transparent SFT, preference optimization and RL with verifiable rewards. Its maintainers direct evaluation users to OLMES rather than the unmaintained internal evaluations.	Related work only.
OpenThoughts-Agent	Data recipes and infrastructure for agent trajectories; inspect trajectory curation and controlled comparisons. This is evolving research code.	Candidate for agent-distillation research; not integrated.
s1	Small curated reasoning-SFT baseline and test-time budget control.	Baseline reference; no current best-model claim.
Open-R1	Open reasoning adaptation/data recipes around the R1 approach; not a guarantee of reproducing the original private training process.	Baseline reference, not a dependency.
SDPO, paper	Learn from rich environment feedback through feedback-conditioned self-distillation; compare against scalar-reward RL at matched compute.	Research-only extension.

On-policy distillation, rich-feedback self-distillation, asynchronous distributed RL, and multi-agent credit assignment are extension topics. They are not synonyms for the implemented offline sequence-distillation path. Require a matched SFT baseline, frozen test protocol, and measured cost before promoting any additional algorithm. New upstream claims and API changes must be checked at the pinned revision rather than inferred from an evolving README.