APIUp to 28% cheaper than official pricesTry the API →
Hermes
Back to News

DeepReinforce's GrandCode Beats Every Human in Three Live Coding Contests

GrandCode won Codeforces Rounds 1087, 1088, and 1089 , beating every human including grandmasters. Multi-agent system: hypothesis proposer, solver, summarizer, and adversarial test generator, built on Qwen 3.5 MoE.

DeepReinforce's GrandCode Beats Every Human in Three Live Coding Contests
Source
AlphaSignal
Published
Author
AlphaSignal
Read
2 min read

Takeaways

  • GrandCode won Codeforces Rounds 1087, 1088, and 1089 , beating every human including grandmasters.
  • Multi-agent system: hypothesis proposer, solver, summarizer, and adversarial test generator, built on Qwen 3.5 MoE.
  • New Agentic GRPO algorithm handles delayed rewards and off-policy drift in long agent rollouts.
  • Prior best AI results: Gemini 3.1 Pro placed 8th, OpenAI o3 placed 175th a year earlier.
  • Training uses seed problems from TACO, USACO, CodeContests, IOI, expanded with Gemini and Claude-generated traces.
  • Code released on GitHub ; approach generalizes best to verifier-rich coding tasks.

DeepReinforce reports that GrandCode, its multi-agent reinforcement learning system, placed first in three consecutive live Codeforces contests: Round 1087 on March 21, Round 1088 on March 28, and Round 1089 on March 29, 2026. The results and system design appear in the project’s technical report , and the code is available in a GitHub repository .

Competitive programming has remained difficult for coding models because success requires algorithm design, exact implementation, and rapid verification against hidden tests. GrandCode’s reported wins suggest that specialized agents, automated testing, and online policy updates can outperform strong human contestants within that tightly defined environment.

A steep rise under different conditions

The paper compares GrandCode with earlier Codeforces results from OpenAI and Google systems. Those placements indicate rapid progress, although differences in model families, compute budgets, and evaluation conditions prevent a controlled comparison.

SystemDatePlacementEvaluation context
OpenAI o3April 2025175thEarlier Codeforces evaluation
Gemini 3.1 ProFebruary 20268thNon-live evaluation
GrandCodeMarch 20261st in three roundsLive contests, according to the paper

Why Codeforces resists coding models

Codeforces problems combine original algorithm design with strict time, memory, and correctness constraints. Partial credit is uncommon, and a single unhandled edge case can invalidate an otherwise sound submission. Contestants typically have two or three hours to derive an approach, prove its complexity, implement it, and survive adversarial hidden tests.

Codeforces ratings also reflect deep specialization. Ratings from 2400 to 2599 correspond to Grandmaster, 2600 to 2999 to International Grandmaster, and 3000 or higher to Legendary Grandmaster. Finishing first means outperforming contestants who have spent years practicing this specific form of problem solving.

Conventional coding-model pipelines often fail before implementation because many contest problems depend on one structural or mathematical insight. Exploring every plausible approach creates an enormous search tree, and a candidate may appear correct until an unseen input exposes a flaw.

Four specialists close the loop

GrandCode coordinates a hypothesis proposer, main solver, test generator, and summarizer. Its main solver uses a Qwen 3.5 mixture-of-experts backbone, which activates a subset of model parameters for each token. DeepReinforce improves the components through post-training and online test-time reinforcement learning.

Reporting is indexed from AlphaSignal. Rights remain with the original publisher and cited sources.

Read original report

Sources

  1. 01arxiv.org
  2. 02hyper.ai
  3. 03x.com
  4. 04github.com
  5. 05en.wikipedia.org