What does the strongest evidence show?
The most direct evidence comes from a 2025 study on multi-agent RL for collaborative LLMs (AT-GRPO). In long-horizon planning tasks, the multi-agent RL approach increased accuracy from a 14–47% single-agent RL baseline to 96–99.5% [5]. That means the multi-agent system went from failing most of the time to succeeding almost every time—a clear sign it learned to avoid the mistakes that plagued the single-agent baseline.
The same study also reported average gains of 3.87–7.62% on coding tasks and 9–17.93% on math tasks, showing the benefit extends beyond planning [5]. These are large, consistent improvements across multiple task types, making this the strongest evidence in the set that cohort-based multi-agent RL can prevent repeated errors in long, multi-step workflows.
How does RL actually prevent mistake repetition?
A 2025 study on interactive digital agents (IDAs) trained with RL in a stateful, multi-domain environment found that the agent learned to consult API documentation, avoid unwarranted assumptions, minimize confabulation, and recover from setbacks [2]. These are exactly the behaviors that prevent repeated mistakes in long tasks: checking facts, not guessing, and correcting course after an error.
The RL-trained 32-billion-parameter agent outperformed a much larger OpenAI o1 agent by 9 percentage points (15% relative) on the AppWorld benchmark [2]. This shows that RL, even in a single-agent setting, can teach an agent to avoid the kind of errors that typically accumulate over many steps—and the multi-agent studies build on that by adding collaboration.
Another 2025 study on multi-agent post-co-training (MAPoRL) found that co-training multiple LLMs together with RL significantly enhanced collaboration performance across multiple datasets, with generalization to unseen domains [3]. The authors explicitly note that training single LLMs alone was insufficient to encourage collaboration, which supports the idea that the cohort aspect is key to avoiding mistakes that stem from poor coordination.
Does this also work in non-LLM multi-agent systems?
Yes, the principle extends beyond language models. A 2023 study on model-based multi-agent RL (MAG) addressed the problem of local prediction errors propagating across agents during multi-step rollouts [4]. By treating local models as decision-making agents, MAG reduced the accumulation of global errors, and experiments on the StarCraft II benchmark showed effectiveness [4]. This is a different domain but the same core issue: in multi-step tasks, small mistakes can snowball, and multi-agent coordination helps contain them.
The 2023 study on distributed multi-agent RL for edge caching (DeepDMRE) also shows that multi-agent coordination improves long-term performance, though it focuses on caching rather than mistake avoidance [1]. It improved edge hit rate by roughly 5% over a simpler multi-agent method and 19–40% over single-agent baselines [1]. While not about mistake repetition per se, it reinforces that multi-agent RL can outperform single-agent approaches in sequential decision-making.
About These Sources
This answer is built on 5 peer-reviewed studies — published from 2023 to 2025, 3 from 2024 or later, collectively cited 244 times — selected as the most relevant from 5 studies that passed quality screening, drawn from 49 papers retrieved from a database of over 500 million.
Sources used in this answer
Distributed Deep Multi-Agent Reinforcement Learning for Cooperative Edge Caching in Internet-of-Vehicles
In a study on edge caching in Internet-of-Vehicles, a deep multi-agent RL method (DeepDMRE) improved edge hit rate by roughly 5% over a simpler multi-agent method and 19–40% over single-agent baselines (Q-learning, LFU, LRU) at a cache capacity of 1,000 MB.
Reinforcement Learning for Long-Horizon Interactive LLM Agents
In a study on interactive digital agents, RL training in a stateful environment (LOOP) enabled a 32-billion-parameter agent to outperform OpenAI o1 by 9 percentage points (15% relative) on AppWorld, and analysis showed the agent learned to consult documentation, avoid assumptions, minimize confabulation, and recover from setbacks.
MAPoRL: Multi-Agent Post-Co-Training for Collaborative Large Language Models with Reinforcement Learning
In a study on multi-agent post-co-training for LLMs (MAPoRL), co-training multiple LLMs with RL significantly enhanced collaboration performance across multiple datasets and generalized to unseen domains, whereas training single LLMs alone was insufficient.
Models as Agents: Optimizing Multi-Step Predictions of Interactive Local Models in Model-Based Multi-Agent Reinforcement Learning
In a study on model-based multi-agent RL (MAG), treating local models as decision-making agents reduced error propagation during multi-step rollouts, and experiments on StarCraft II demonstrated effectiveness.
Stronger-MAS: Multi-Agent Reinforcement Learning for Collaborative LLMs
In a study on multi-agent RL for collaborative LLMs (AT-GRPO), long-horizon planning accuracy improved from a 14–47% single-agent RL baseline to 96–99.5%, with average gains of 3.87–7.62% on coding and 9–17.93% on math tasks.