Choose what to learn
Capability structure and curriculum selection direct attention toward skills and observed task types.
MULTI-AGENT GRAPH-GUIDED EVOLUTION
The model stays frozen.
The memory evolves.
MAGE stores what an agent learns in a knowledge graph. Explore how a recorded mistake becomes a correction—and later, a reusable success.
Explore the replayTHE EVOLUTION LOOPSame weights · growing memory
Append experience. Adapt search & routing. Reuse next iteration.
While building EvoKG: the guidance model writes corrections and principles. The learner’s weights stay fixed.
After evolution: the frozen learner answers using saved EvoKG memories. No guidance-model calls are needed.
01 / EXPERIENCE REPLAYRecorded evidence, step by step
FROM A MISTAKE TO A MEMORY
Watch the learning move
outside the model.
Nine benchmarks, one example each. Explore recorded recoveries, held-out answers, and saved action sequences—alongside the evolving graph.
Replay data could not be loaded. Please reload the page.
SELECTEVALUATEWRITE MEMORYREUSEHARVEST
Recorded evidence & illustrative memory selection
Each tab names its evidence type: evolution-pool recovery, held-out answer, or saved action recipe. Answers, corrections, curriculum targets, memories, and graph snapshots come from saved runs. The dual-memory view pairs real, task-matched records available before the recovery. It illustrates the success-first / correction-second mechanism; it is not a reconstruction of the exact historical prompt. No per-query retrieval log was saved.
Evolution-pool recoveries and held-out records are kept distinct. A later correct answer does not by itself establish which memory caused the improvement. Action recipes preserve action order; intermediate observations and per-action rewards are not reconstructed.
Download the curated replay data02 / THE MEMORY SUBSTRATEEvery dot is a saved node
EVOLUTION YOU CAN INSPECT
Watch experience enter the graph.
Scrub through the run. New nodes appear with an outline; older nodes keep their positions. Highlight the case to find its correction, success, and illustrative memory bundle.
CapabilitiesPrinciplesFailure correctionsSuccess memories / recipesTasks & environmentNew this snapshot
Solid: recorded edgeDashed: shared task type with the selected case
Each dot represents one node in this snapshot. Layout groups nodes by type for readability; group positions do not encode semantic distance. Dashed case links are derived from task-type metadata, not stored graph edges.
03 / EMPIRICAL RESULTSFrom the working manuscript
BEYOND THE SELECTED CASES
Five reasoning benchmarks.
Qwen3-8B remains frozen for all methods. The table compares MAGE with the strongest reported prompt-based baseline on each benchmark.
| Benchmark | Strongest baseline | MAGE / Sonnet | MAGE / Opus | Best MAGE gap |
|---|
MAGE values are mean ± standard deviation; n denotes the number of seeds. All rows use the manuscript’s Claude Sonnet 4.6 semantic-judge protocol. The gap selects the better of the two MAGE teacher settings. Source: main reasoning-results table in the working manuscript.
Scope matters.
These five benchmarks are one part of the nine-benchmark study. The draft’s broader results are mixed: MedQA is below its zero-shot baseline, while Crafter remains within the standalone baseline’s seed-variance range.
04 / HOW MAGE WORKSLearning in the external state
A PERSISTENT LEARNING LOOP
Keep the weights.
Build the experience.
During evolution, guidance writes corrections and principles. The learner’s own successful answers become reusable examples. At inference, the frozen learner consults task-conditioned memories.
Record both outcomes
Harvest correct traces from the learner and add teacher-written corrections for observed failures.
Retrieve and reuse
Task-filtered success and failure memories guide the frozen learner. Search and routing bandits adapt during evolution.
05 / PAPER & CITATIONNeurIPS 2026 · Sydney
RESEARCH PAPER
MAGE: Multi-Agent Self-Evolution with Co-Evolutionary Knowledge Graphs
Cite MAGE
Ruiyi Yang, Zechen Li, Hao Xue, Imran Razzak, and Flora Salim. MAGE: Multi-Agent Self-Evolution with Co-Evolutionary Knowledge Graphs. NeurIPS, 2026.
@inproceedings{yang2026mage,
title = {{MAGE}: Multi-Agent Self-Evolution with Co-Evolutionary Knowledge Graphs},
author = {Yang, Ruiyi and Li, Zechen and Xue, Hao and Razzak, Imran and Salim, Flora},
booktitle = {Advances in Neural Information Processing Systems},
year = {2026},
url = {https://openreview.net/forum?id=dhUiKcjH2m}
}