Research map / Large language models / Agents, code and math
Long-horizon LLM agents
Train and structure LLM agents for long-horizon planning, exploration, and in-context learning.
155 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2019–2026) · an approach in Agents, code and math, keyword cluster “evaluating · capabilities · bench”
State History Tried Checklist Preprints Who Export
Where it stands
Not enough benchmark data Fewer than four papers report the same benchmark, so there is no record to follow.
26 papers in 2023–24 and 125 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 7.2% to 9.9%.
How this is decided
A benchmark is tracked when at least four papers of this approach report their own method on the same dataset and metric in at least two different years. Its record is the best number reported up to 2024 compared with the best reported in 2025–26: still improving means at least 1% better, flat means at most 0.2% better with at least two papers from 2025–26 trying. Numbers on a different scale from the rest (a fraction among percentages) and far outliers are left out. Results come from different papers with different setups, so they are signals, not head-to-head comparisons. The share of papers is counted at ICML, NeurIPS, ICLR, CVPR and AAAI.
How it got here
Most cited paper of each year
- 2019 · Hierarchical Decision Making by Generating and Following Natural Language Instructions (NeurIPS 2019 · 33 citations)
- 2020 · MALA: Cross-Domain Dialogue Generation with Action Learning (AAAI 2020 · 19 citations)
- 2022 · PlanVerb: Domain-Independent Verbalization and Summary of Task Plans (AAAI 2022 · 5 citations)
- 2023 · On the Planning Abilities of Large Language Models - A Critical Investigation (NeurIPS 2023 · 53 citations)
- 2024 · ChatScene: Knowledge-Enabled Safety-Critical Scenario Generation for Autonomous Vehicles (CVPR 2024 · 79 citations)
- 2025 · Mixture-of-Agents Enhances Large Language Model Capabilities (ICLR 2025 · 26 citations)
- 2026 · Multi-turn Evaluation of Anthropomorphic Behaviours in Large Language Models (ICLR 2026 · 8 citations)
Earliest papers here
- Hierarchical Decision Making by Generating and Following Natural Language Instructions (NeurIPS 2019 · 33 citations)
- Plan-and-Write: Towards Better Automatic Storytelling (AAAI 2019)
- MALA: Cross-Domain Dialogue Generation with Action Learning (AAAI 2020 · 19 citations)
What has been tried
62 papers here report 608 measured changes (a component added, removed or swapped, compared with a baseline): helped 66%, no clear effect 7%, hurt 24%, mixed 3%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| increase proposer count from 1 to 6 using different models | AlpacaEval 2.0 · LC win rate | 47.8\% -> 61.3\% | Helped | Mixture-of-Agents Enhances Large Language Model Capabilities (ICLR 2025 · 26 citations) |
| Use T5-Large instead of GPT2-Large | HumanML3D motion generation · R Top-1 | 0.454 -> 0.510 | Helped | AvatarGPT: All-in-One Framework for Motion Understanding Planning Generation and Beyond (CVPR 2024 · 24 citations) |
| Use beam Search Tree without Analysis and Repair or Selection | Tidy kitchen · task completion (%) | 77.5 -> 50.0 | Hurt | Improving Knowledge Extraction from LLMs for Task Learning through Agent Analysis (AAAI 2024 · 19 citations) |
| Add P1 single-perception tuning | Touchdown Dev Set · TC | 40.16 -> 41.15 | Helped | FLAME: Learning to Navigate with Multimodal LLM in Urban Environments (AAAI 2025 · 9 citations) |
| add reflection and iterative refinement to DLM | 16 public-health prompts · MNR | 0.87 ± 0.008 -> 0.93 ± 0.006 | Helped | A Decision-Language Model (DLM) for Dynamic Restless Multi-Armed Bandit Tasks in Public Health (NeurIPS 2024 · 9 citations) |
| Add grounding-based feasibility feedback to the PDDL description | Five long-horizon tasks · executable-action success rate | Helped | Instruction-Augmented Long-Horizon Planning: Embedding Grounding Mechanisms in Embodied Mobile Manipulation (AAAI 2025 · 5 citations) | |
| Increase sampled plans from N=25 to N=50 | VirtualHome w/o correction · SR | 28.33±1.18 -> 28.14±2.45 | No clear effect | Tree-Planner: Efficient Close-loop Task Planning with Large Language Models (ICLR 2024 · 5 citations) |
| combine LLM reward with the VLM reward | Open-vocabulary humanoid skills · task completion (C) | 6.108 -> 6.954 | Helped | GROVE: A Generalized Reward for Learning Open-Vocabulary Physical Skill (CVPR 2025 · 4 citations) |
| Use transcript-style rather than dialogue-style negotiations | Negotiation games · agreement rate | 0.49 to 0.44 | Hurt | Evaluating Language Model Agency Through Negotiations (ICLR 2024 · 4 citations) |
| Replace summarized history with raw history | Hard 5-armed bandit, T=200 · suffix failure frequency | Hurt | Can large language models explore in-context? (NeurIPS 2024 · 4 citations) | |
| replace GPT-4o with GPT-4o-mini | FEA projects test set · success rate (%) | 90.2 -> 81.6 | Hurt | AutoFEA: Enhancing AI Copilot by Integrating Finite Element Analysis Using Large Language Models with Graph Neural Networks (AAAI 2025 · 3 citations) |
| add a checklist to reward assignment | WebRewardBench Mind2Web Cross-Task · Acc. (traj) | 0.0 -> 5.0 | Helped | Web-Shepherd: Advancing PRMs for Reinforcing Web Agents (NeurIPS 2025 · 3 citations) |
What reviewers will expect
Datasets and baselines named in the 59 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- ALFWorld · 4 of 59 papers
- Mind2Web · 3 of 59 papers
- AITW · 2 of 59 papers
- MiniWob · 2 of 59 papers
- VirtualHome · 2 of 59 papers
- ToolBench · 2 of 59 papers
- API-Bank · 2 of 59 papers
Baselines
- GPT-4o · 7 of 59 papers
- ReAct · 6 of 59 papers
- Claude-3.7-Sonnet · 4 of 59 papers
- Gemini-2.5-Pro · 4 of 59 papers
- DeepSeek-V3 · 3 of 59 papers
- Claude-3.5-Haiku · 3 of 59 papers
- GPT-4o-mini · 3 of 59 papers
- Reflexion · 3 of 59 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 151 papers; from OpenAlex and the papers' affiliations).
- Shanghai Jiao Tong University · 6 papers
- Massachusetts Institute of Technology · 6 papers
- Microsoft · 5 papers
- Fudan University · 4 papers
- University of Washington · 4 papers
- Mohamed bin Zayed University of Artificial Intelligence · 3 papers
- University of Chinese Academy of Sciences · 3 papers
- Stanford University · 3 papers
- Princeton University · 3 papers
- University of California, Berkeley · 3 papers
Use it
Download BibTeX of 28 key papersOpen the topic in the interactive map
Get alerts. Describe your idea in scime atlas and save it (free scime account): you get a weekly email when new preprints land close to it.
Related approaches
Same keyword cluster
- Agent instruction-following benchmarks (176 papers)
- Software engineering agents (140 papers)
- Reasoning benchmarks and platforms (47 papers)
- LLM-driven embodied and driving agents (54 papers)
Elsewhere in the topic
- Mathematical and structural reasoning benchmarks (304 papers)
- Inductive and abstract reasoning evaluation (158 papers)
- LLMs for optimization and heuristics (147 papers)
- LLMs for scientific discovery (100 papers)
- Code generation benchmarks (96 papers)
- LLM-guided program synthesis (93 papers)
- Execution-based code verification (90 papers)
- Retrieval-augmented formal theorem proving (62 papers)
Benchmark numbers, ablations and baseline numbers are extracted by AI from each paper's full text and can be wrong; the paper is the authority. Data: results index of 2026-11-02, citation counts from OpenAlex (2026-10-09).