Ngseo/dr-marl-papers

Distributionally-Robust RL & Cooperative MARL — top-tier paper index

A hand-reviewed index of 1,989 papers on distributionally-robust reinforcement learning and cooperative multi-agent RL, drawn from a complete harvest of 75,819 accepted papers at ICML, NeurIPS, ICLR, AAMAS and AISTATS (2010–2026).

Every candidate that survived the keyword filter was read and labelled one by one — 2,979 papers — rather than accepting an automatic classifier's output.

Buckets

bucket	meaning	n
A	Distributionally-robust / robust RL — robustness to model, dynamics or environment uncertainty	171
B	Cooperative multi-agent RL	1,083
C	The intersection: robust / distributionally-robust multi-agent RL	70
D	Adjacent, kept but separated — risk-sensitive & CVaR RL, state/observation-adversarial RL, safe RL, purely competitive / equilibrium-computation MARL, multi-agent bandits, MAPF, LLM multi-agent systems	665

Coverage

venue	A	B	C	D	total
AAMAS	4	474	23	173	674
NeurIPS	72	253	14	182	521
ICML	57	191	17	149	414
ICLR	22	144	12	131	309
AISTATS	16	21	4	30	71

Harvested venue-years: ICML 2013–2026, NeurIPS 2010–2025, ICLR 2013–2026, AAMAS 2013–2026, AISTATS 2010–2026. Per-year corpus counts are in reports/coverage.md.

Labelling criteria

Every bucket decision answers one question: what is the paper's contribution about? Not what it cites, not what it evaluates on. A paper that benchmarks on SMAC but contributes a single-agent exploration method is not bucket B.

A — distributionally-robust / robust RL

The contribution concerns uncertainty in the environment model itself:

• distributionally robust MDPs, robust MDPs / robust Markov games
• uncertainty or ambiguity sets, rectangularity, f-divergence / Wasserstein / KL balls
• robust Bellman operators, robust value / policy iteration, robust policy gradient
• transition-kernel or dynamics mismatch, model misspecification at the MDP level
• worst-case return over a set of environments; percentile / satisficing criteria
• sim-to-real gap and domain randomisation treated as a robustness objective
• unsupervised environment design with a minimax-regret objective

Not A → D: robustness to observation or state perturbations (the SA-MDP line), risk-sensitive / CVaR objectives, safe & constrained RL, data-corruption robustness, plain offline-RL pessimism, off-dynamics / cross-domain transfer without a worst-case formulation, reward-model robustness in RLHF.

Not A → not in this dataset: distributionally robust optimisation in supervised learning — no sequential decision component. (122 such papers were filtered out.)

B — cooperative multi-agent RL

Multi-agent RL, or multi-agent sequential decision learning, where the agents share or partly share an objective:

• Dec-POMDPs, CTDE, value decomposition, credit assignment
• multi-agent policy gradient / actor-critic, coordination graphs, roles & skills
• communication learning and emergent communication
• ad hoc teamwork, zero-shot coordination, human-AI collaboration with a learning agent
• team / common-reward games, Markov potential games, mean-field RL and control
• offline MARL, multi-agent exploration, scalability, MARL benchmarks and environments

Default rule: at these venues, a paper that calls itself multi-agent reinforcement learning counts as B unless it is framed as purely competitive. An explicit "cooperative" in the abstract is not required — many cooperative MARL papers never use the word.

Not B → D: purely competitive or zero-sum equilibrium computation, multi-agent bandits, multi-agent path finding and DCOP (planning, not cooperation learning), LLM multi-agent systems and agentic pipelines, opponent modelling in adversarial settings, mechanism design.

C — the intersection

A paper is C when it satisfies B and its contribution is about robustness or uncertainty — it is not a union, and C papers are not also counted in A or B:

• robust / distributionally robust Markov games and cooperative MARL
• adversarial attacks on, and defences for, multi-agent communication or agents
• Byzantine, corrupted, faulty or vulnerable agents
• partner / teammate robustness: unseen co-players, off-team play, diverse-partner training when the stated goal is robustness
• MARL under model uncertainty, multi-agent sim-to-real / domain calibration

D — adjacent

Everything the filter surfaced that borders the two topics without being either. D is a scoping decision, not a verdict on quality or relevance — if your definition of robust RL includes observation-adversarial work, or your definition of MARL includes competitive games, filter bucket == "D" and merge it in.

Worked examples

How these rules actually landed, counted over the released file:

paper family	where it went
ad hoc teamwork / zero-shot coordination	37 B, 3 C (C only when framed as robustness)
emergent communication	30 B
mean-field RL / games	48 B, 15 D (D when it is a pure equilibrium-computation result)
sim-to-real, domain randomisation	12 A, 4 D
off-dynamics / cross-domain transfer	4 A, 15 D (A only with a worst-case formulation)
risk-sensitive / CVaR	35 D, 6 A, 4 B
state / observation adversarial	4 D, 1 A
zero-sum / two-player	39 D
multi-agent bandits	53 D
multi-agent path finding, DCOP	9 D
LLM / agentic multi-agent systems	26 D, 12 B
"distributionally robust" anywhere in the title	37 A, 5 C, 1 D

Reproducibility of the labels

label_source records how each row was decided:

• manual — read and labelled by hand (2,979 candidates went through this)
• auto — an unambiguous title pattern, e.g. a title containing "cooperative multi-agent reinforcement learning"
• snowball — never surfaced by the keyword filter, recovered through the citation graph and then labelled by hand

abstract_chars == 0 marks rows judged on the title alone, because no abstract was reachable — almost all of them AAMAS demos, doctoral-consortium entries and some extended abstracts.

Columns

column	notes
paper_id	{venue}:{year}:{hash of normalised title}
bucket	A / B / C / D
subtopic	e.g. A2 uncertainty-set design, B1 value decomposition, C3 partner / teammate robustness
venue, year, track	track separates AAMAS full papers from extended abstracts / demos / doctoral consortium, and NeurIPS main from Datasets & Benchmarks
title, authors	authors are ; -separated
url_abs, url_pdf	publisher links
decision	accept type where the venue publishes it (oral / spotlight / poster)
source	which harvester produced the record — provenance for every row
label_source	manual (read and labelled by hand), auto (unambiguous title pattern), snowball (recovered through the citation graph)
abstract_chars	length of the abstract we worked from, so you can tell title-only judgements apart

Abstracts are not included — fetch them yourself

python fetch_abstracts.py     # -> papers_with_abstracts.csv

One request per venue-year where the conference publishes a whole-year JSON, one request per paper otherwise; everything is cached under .cache/. It recovers 1,311 / 1,989. The remaining 678 are AAMAS rows: IFAAMAS publishes PDFs only, and its proceedings notice forbids republishing, so there is no source we can point you at for those.

How labels were assigned

1. Harvest every accepted paper from the official proceedings (PMLR, papers.nips.cc, the conference virtual sites, IFAAMAS). DBLP was unusable — every dblp.org endpoint, including the XML dumps and the uni-trier / dagstuhl mirrors, returns an anti-bot challenge; the OpenReview /notes API likewise. We did not try to defeat either.
2. High-recall keyword filter (config/lexicon.py), two tiers per topic. Deliberately over-inclusive: 5,285 candidates out of 75,819. distributional RL (C51/QR-DQN) is explicitly separated from distributionally robust.
3. Precision rule pass (src/classify_rules.py). Multi-agent RL papers at these venues are treated as cooperative by default; explicit zero-sum / competitive-only framing demotes to D.
4. Manual review of 2,979 candidates, including every rule-reject that still carried a strong multi-agent or robustness keyword, and every record that had no abstract.
5. Citation-graph snowballing from the A+C core (15,446 reference/citation edges) to find papers the keyword filter never surfaced.

Recall evidence

check	result
Anchor test — 62 known papers listed before any results were inspected (config/anchors.json)	52/53 in-corpus anchors survive the filter (98%)
Offline audit — every corpus paper whose title states a topic unambiguously (reports/audit.md)	98.6% (DR-RL) / 98.9% (coop-MARL) land in the expected bucket
Citation-graph snowball	0 unrecovered papers whose title states either topic outright — it returned 2 before the lexicon was widened, which is what it is there for

The audit lists every paper that fell outside its expected bucket, so the scoping calls stay auditable rather than implicit.

Reproducing

./run_all.sh          # harvest -> merge -> filter -> classify -> triage -> export

Harvests are resumable and every HTTP response is cached, so re-runs are cheap. Manual labels live in data/labels_manual.json, keyed by a stable hash of paper_id, so the pipeline can be re-run without redoing the review.

Licensing

• This dataset — the bucket/subtopic labels, the lexicon, the code and the compilation — is released under CC BY 4.0.

• The underlying papers are not. Titles, authors, venues and years are bibliographic facts and are redistributed here on that basis. Abstracts and full texts are not redistributed, because no source grants us the right to:

  • AAMAS / IFAAMAS: "To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee."
  • PMLR (ICML, AISTATS): "Copyright © The authors and PMLR" — per-paper licences vary, there is no blanket grant.
  • NeurIPS: the proceedings pages carry no licence notice; authors retain copyright.
  • OpenReview / ICLR: authors grant OpenReview a non-exclusive licence and keep copyright.
  • Semantic Scholar: its API licence is a bilateral agreement with the API user, not a redistribution grant.

  If you rehydrate abstracts with fetch_abstracts.py, they remain under their publishers' terms — use them locally, do not redistribute them.

Limitations

• ICLR 2013–2017 has no machine-readable official proceedings we could reach. Those records come from a Semantic Scholar venue dump, carry track: "unverified" and source: "s2:venue-dump", and may include workshop-track papers. Semantic Scholar also reports the arXiv year for some of them, so ICLR records are deduplicated on title across years.
• ICML 2010–2012 is not on PMLR and is not covered.
• NeurIPS 2026 had not happened at collection time. NeurIPS 2025 comes from the conference virtual site because the main-track proceedings were not yet on papers.nips.cc.
• ~1,045 AAMAS corpus records (mostly demos, doctoral consortium and some extended abstracts) had no abstract available and were reviewed on title alone; abstract_chars == 0 marks them.
• Labels reflect one reviewer's reading of a title and abstract, not of the full paper.

Citation

@misc{drrl_coopmarl_index_2026,
  title  = {Distributionally-Robust RL and Cooperative MARL: a hand-reviewed index of
            top-tier papers (ICML, NeurIPS, ICLR, AAMAS, AISTATS, 2010-2026)},
  author = {Ngseo},
  year   = {2026},
  note   = {Hugging Face dataset}
}

Files

file	what
papers.csv / papers.jsonl	the index — 1,989 labelled papers
fetch_abstracts.py	rehydrates abstracts locally from the publishers
src/, config/, run_all.sh	the full pipeline, including the keyword lexicon and the 62-paper anchor list
labels/labels_manual.json	the 2,979 hand-assigned labels, keyed by a stable hash of paper_id
labels/snowball_manual.json	citation-graph recoveries, keyed by normalised title
reports/	corpus coverage, final counts, and the offline recall audit
counts.json	machine-readable summary