reinforcement learning

Agent Lightning v1.0 connects any harness to RL through an endpoint proxy" — Article demonstrates practical RL application for agent training using Agent Lightning v1.0, showing concrete implementatio

When an agent acts in an environment, the environment's response to that action is always true." — Article demonstrates RL principle: environment responses provide supervision signals for learning age

During each step of a task, the agent is in a specific state, generates an action (typically an LLM output), and receives a reward depending on whether the action helps achieve the task goal. These re

OpenClaw-RL is a reinforcement learning framework that turns everyday conversations into training signals for personalized AI agents." — Article demonstrates a concrete RL implementation for LLM agent

Specializes in Reinforcement Learning and Multi-Agent Systems" — Professor Gini explicitly specializes in reinforcement learning as a primary research focus

SAPO targets long-horizon agents. It shares one backbone between the policy and value function, learns from a single rollout, and assigns credit across turns. It reports stronger results than PPO and

Reasoning effort is treated as a trainable capability. Token budgets are estimated from the SFT model, and a stage-wise curriculum anneals from long to short rollouts to produce low-, high- and max-ef

There is still limited evidence that today's models can learn reusable abstractions from experience over the long term, which I believe is a crucial capability for agents that continuously improve." —

They then use RL to train the policy model to leverage those structural blueprints under outcome rewards." — RA-RFT demonstrates practical RL application for training policy models with outcome-based

In traditional agentic RL, the training engine owns the environment interaction loop. In harnessed agentic RL, the harness owns this loop, while the training engine observes only a sequence of LLM req

First, supervised fine-tuning teaches the harness action space. Then cost-aware GRPO learns when to read the state, update it and consolidate accumulated information during a long run" — Demonstrates

agent dreaming can select and normalize sessions from every harness on the machine, so lessons learned in a Claude Code or Codex session inform the agent's memory the same way its own sessions do" — A

for tasks with shared structure that look different, the root model naturally learns the same trajectory, meaning it views the two task trajectories as the same!" — This paper introduces a novel findi

Every improved workflow generates better training signal, which accelerates the accumulation of tacit knowledge unique to the firm." — Article demonstrates organizational learning as a compound mechan

the actual skill is a stack of smaller skills, and almost every one of them can be deliberately trained" — Explicitly claims research skills are learnable and trainable through deliberate practice

Trains the Solver, Verifier, and Corrector agents together with separate rewards for each and a pipeline-style RL setup" — MarsRL demonstrates RL applied to multi-agent training with agent-specific re

Agent 可以根据环境学习，环境也可以根据 Agent 暴露出的弱点继续变化。以后持续进化的系统会同时维护两个循环，Agent 在变强，训练它的世界也跟着一起变。" — Articulates a co-evolutionary model where both agent and environment adapt bidirectionally, extending traditional

In this project I am trying to train a small model (Qwen3-1.7B) on 15 puzzle using grpo." — Article demonstrates RL application using GRPO algorithm to train language model on combinatorial task

Memory as infrastructure enables continual learning for agents - they can learn from experience, adapt to feedback, and improve over time without you having to manually manage any of it." — Article ar

agent training can be reframed as a cross-episode Meta-RL problem, explicitly incentivizing the model to gather diverse experiences in early attempts to maximize long-term rewards" — Article extends R

self-directed learning from real + simulated experience" — The article proposes a novel distinction between learning from real vs simulated experience as a component of true agent autonomy.

Traditional pipelines need sandboxed test runs = 💰💰💰 🎯 Add hypothesis confidence tracking (20-30% error reduction)" — Article extends RL training approaches by proposing hypothesis confidence trac

Without feedback, you have automation — efficient execution of predetermined steps. With feedback, you have intelligence — a system that learns what works and adapts accordingly." — Article directly a

Reinforcement learning (RL) is expensive, sample-inefficient, brittle, and fails to generalize" — Article directly contradicts the assumption that RL can scale effectively due to fundamental sample in

The kind of thing that has made apprenticeship like models so important throughout history. Getting a PhD is like this. There are obviously best practices for doing science, but it's just too hard to

Yet our most sophisticated neural networks suffer catastrophic forgetting when asked to learn sequentially." — Article identifies catastrophic forgetting as a core problem in neural networks and propo

reduced mastery by 17%" — Presents evidence that AI-assisted coding reduces developer mastery and understanding, challenging assumptions that AI tools uniformly enhance developer capability

The ability of the Claude team to learn from things like OpenClaw and implement features like this on a daily basis" — Demonstrates rapid learning from external implementations and accelerated feature

We call this Memory Scaling, and it's related but different from continual learning." — Introduces memory scaling as a distinct concept from continual learning, clarifying that agent improvement throu

Each agent employs AI algorithms—like reinforcement learning or game theory—to make decisions. Over time, agents can learn from interactions, improving their strategies." — Article provides explicit e

We know standard RL on pass/fail test cases gets models to generate correct code. The moment you try rewarding execution speed, though, things usually fall apart." — Article demonstrates a frontier ch

those evals double as training data if you choose to RL" — Article proposes using trace-based evaluation environments as RL training data, leveraging behavioral data for agent improvement

Scaling agentic RL environments: today we're publishing 365,000+ tasks" — Article explicitly announces large-scale RL environment dataset (365,000+ tasks), providing infrastructure and data for traini

I think we'd greatly benefit from more RL on memory extraction and/or discovery." — Article proposes using reinforcement learning specifically for memory extraction/discovery, suggesting RL can optimi

For juniors, you want someone who is willing to slow down and be intentional... We should take on the friction! Read your code, try to understand it." — Article argues that junior engineers in AI-assi

Private reinforcement learning environments should let models grow stronger on real traces from inside the organization." — Concrete example of using private RL on organizational data to build proprie

Use the AI to explain a complex block, then try to explain it back to the AI in your own words. If the AI corrects you, stay on that block until you truly own the logic." — Article advocates using AI

engineers must learn new skills to manage and guide AI-generated work effectively" — Article emphasizes that AI integration creates new skill requirements: managing, validating, and directing AI-gener

Reinforcement learning trains them on rewards, not on soft judgments" — Directly explains the mechanistic reason why RL-trained agents fail on subjective tasks — their optimization target is binary/qu

slow learning is LLM weights trained with RL while fast learning is context / prompt (fast weights)" — Article extends RL paradigm by decomposing it into slow (weight-level) and fast (context-level) c

[INFERRED] "the policy never optimize the same set of rubrics twice" — Addresses vulnerability in RL optimization where policies exploit static evaluation criteria; solution targets training stability

[INFERRED] "passively learn through interaction and feedback" — Article describes agents learning from user interactions without explicit retraining, supporting autonomous capability development parad

as firms build out more sophisticated eval / RL env (increasingly the same thing) infra, it starts to become viable to post-train an custom model on top of an OSS base." — Article extends RL concept b

Coding with AI led to a decrease in mastery—but this depended on how people used it." — The article presents empirical evidence from an experiment with software engineers showing that AI assistance ca

[INFERRED] "devs who are already super jacked and have years of experience building complex systems can crush juniors with ai. THE GAP IS REAL." — Article directly argues that AI amplifies existing ex

post-training a small CNN policy outperforms LLMs, but only with legal action masks" — Demonstrates PPO/RLHF post-training effectiveness on policy models with constraints, validating the training meth

Memory subagents can rapidly ingest and generate Git-backed context trees." — Extends learning capabilities by enabling agents to rapidly process and generate structured context representations, creat

This method improves reinforcement learning by making rewards more reliable, especially for complex or subjective tasks." — Rubric-based rewards directly improve RL training quality by providing more

[INFERRED] "RL makes sparse, near-orthogonal updates, letting capabilities accumulate with far less interference." — Article demonstrates key advantage of RL over SFT: sparse parameter updates reduce

[INFERRED] "self-modifiable harness state" — Article describes self-improving RLM with self-modifiable state, indicating capability for learning and adaptation