Reinforcement learning

Reinforcement learning methods, RL for language models and agents, reward modelling, preference optimisation

207 articles · page 1 of 3

All Agents Language models Reinforcement learning Vision Generative media Speech & audio Robotics & embodied AI Safety & security Evaluation & benchmarks Efficiency & systems Code & software AI for science & health Theory

Viewgrid list compact

World models learn true dynamics with fewer parameters

Teacher corrections with dense feedback double AI success on hard logic problems

AI models can be trained end to end with one algorithm

Language models solve more problems when training accounts for confidence

Targeting uncertain local decisions cuts GUI agent training cost by 56-61%

Online reinforcement learning on live iOS apps yields UI agents with 7x better cross-platform task success

AI reasoning models learn from reference examples alone, no reward needed

Self-teaching language models improve when the teacher gets feedback too

Feedback criteria that track a model's growing ability halve its training steps

Coding agents solve more tasks when tool interfaces adapt during training

Language models can learn from expert guidance when every attempt fails

Aligning auxiliary rewards with task success improves multi-task language model training

Training on visual errors improves image understanding and tool use

Rewarding precision reduces unnecessary and incorrect content in rubric-trained models

AI agents’ performance can be predicted using 100-fold smaller training runs

Guided self-exploration lets language models teach themselves better reasoning

AI agents improve at search and automation when trained around complete replies

AI driving assistants reduce collisions by anticipating what drivers intend

Language models can match math accuracy with eight times fewer answers

Language models find more valid solutions without supplied correct answers

Aligning rounding locations between inference and training stabilizes BF16 LLM reinforcement learning

Regularised reinforcement learning boosts policy robustness in sim-to-real transfer

Detached shorter-horizon features boost long-horizon reinforcement learning

Better static judges do not guarantee better training signals for generative reward models.

Rewarding moral arguments makes a language model more consistent, not sounder

Language models reason better when conflicting teacher guidance is filtered

Reconstructing app behavior before coding improves simulated GUI agent training

Training search agents to assess missing evidence improves research performance

Training world models for policy improvement beats optimizing pixel accuracy

AI agents answer questions better using 68% fewer extra training tokens

Chatbots using personal memories do not always produce users’ preferred answers

Polynomial-time algorithm achieves optimal regret for adversarial linear MDPs

Retry-aware credit assignment boosts LLM reasoning efficiency and accuracy

A simulator predicts when detectable reward hacking will dominate training

Training models to recover regions from descriptions improves visual grounding

Shared hidden-state drift can collapse language generation while preference rankings improve

Self-Credit rescales self-assigned credit to improve long-horizon agentic RL

DPO hyperparameters do not transfer across model scales or datasets; validation loss is unreliable

Trajectory refinement improves reinforcement learning for long-horizon language-model agents

Critics that predict task success, not just returns, improve long-horizon agent credit assignment

Teacher rollouts repair myopic token scoring in agent distillation

Sparse probes improve credit assignment for language model reasoning

Adaptive replay preserves reasoning diversity in RL fine-tuning

Offline experience can reveal where agents have greater future control

Adaptive rubrics improve reinforcement learning for open-ended text generation

Interventional step credit prunes redundant chain-of-thought while preserving accuracy

Varying sampled video frames strengthens reinforcement learning for temporal reasoning

Simulation priors speed adaptation of model-based autonomous racing controllers

Value Disaggregation trains policies without inference-time goal conditioning

Balancing hidden quality dimensions reduces reward hacking in rubric-trained models

Adaptive training harness keeps rubric rewards useful as policies improve

Conditioning actions on returns improves value learning in reinforcement learning

Language models learn to propose improvements for other language models via RL with a learned world model

Bayesian belief framework jointly handles two uncertainties in process reward models

Sequence-level credit assignment improves on-policy distillation of reasoning models

Separating search and answer credit improves accuracy and evidence grounding

JAX-based Atari environments enable GPU-accelerated training and easy modification

Stochastic attention routing enables efficient RL alignment for diffusion models

Decoupling skills from expert policies improves multi-capability vision-language model training

A few clean labels can fix sample-wise noise in reward model training

Adversarial rubric repair improves self-training on open-ended expert tasks

Calibrated rewards curb answer degradation during verifiable instruction training

Imagined goals from geometric horizon models boost behavior cloning generalization

Backpropagating verifier rewards improves reasoning in continuous flow language models

One-step generative policy matches Boltzmann target for online MaxEnt RL

Experience replay boosts RL training efficiency for large language models by reusing trajectories

Gradient norm ranks attention heads for efficient sparse and continual RL fine-tuning

Turning failed reasoning attempts into training signal improves GRPO on hard problems

Unified platform enables GUI agent training across mobile and computer environments

Equal reward weighting can steer language models toward easier shortcuts

RL gains in no-think mode stem from reactivating existing reasoning

Muon optimizer boosts training success for long-horizon repository agents

Hindsight assigns token credit from one verified language-model rollout

Hyperbolic geometry beats extreme network depth in goal-conditioned maze navigation

Context-aware credit assignment improves language agents on long-horizon tasks

Symplectic latent dynamics prevent rollout contraction in visual world models

Advantage guidance makes supervised policy training competitive with offline RL

Frozen LLM critics sharpen credit assignment for long-horizon agents

Specialized GFlowNet mixtures find more modes without extra oracle calls

Routing gradients can isolate reward hacking for removal after training

Criterion-aware credit assignment improves rubric-based reinforcement learning for language models

Generative models as proposal distributions improve long-horizon decision making

Progress-aware advantage shaping resolves uniform credit assignment in agentic RL

Reward models for vision-language AI can be trained without human preference labels

Self-judged trajectories improve SWE agent training beyond binary verifier rewards

Recursive synthesis turns unsolved problems into useful reinforcement learning data

Agents keep more options open by seeking states rich in skills

Adaptive curriculum training lets vision-language models solve hard visual Sudoku puzzles

Divergence geometry and estimation jointly improve language-model policy optimization

Trajectory-aware updates stabilize reinforcement learning for longer LLM reasoning

Intervening During Rollouts Helps Agents Learn Beyond Their Initial Limits

Rewards Barely Drive Learning in Direct In-Context Reinforcement Learning

Separating visual and language reasoning phases improves vision-language model accuracy

Residual advantage uses teacher-student disagreement as bounded step reward

Safety filter separates task and safety learning in robot control

Calibrating group rewards can avoid failures from independent KL penalties