framework
active
framework:rlfr-reinforcement-learning-from-feature-rewardsRLFR (Reinforcement Learning from Feature Rewards)
Core pipeline turning probe-read features into scalable reward functions for open-ended tasks.
Neighborhood — ranked by edge-count
Papers (1)
paper
Methods (10)
method
- ProbingusesStandard interpretability technique used to read feature beliefs.
- Attention probe predicting whether an Entity is hallucinated; AUC .94 on validation.
- Causal transformer probe predicting whether a token is in the same Entity as its predecessor.
- Test-time compute scaling guided by reward probes with majority-vote over action.
- CISPOusesPolicy optimization protocol used.
- Correction probeusesReward probe grading correction interventions (5-class).
- Value learning method inferring reward function from expert demonstrations; reviewed as insufficient for superintelligent alignment
- Learned multiplier maintaining 60:40 correction:retraction ratio.
- Overall reward is product of judge-based and probe-based terms.
- Retraction probeusesReward probe grading retraction interventions.
Concepts (1)
concept
- New affordance of features beyond monitoring/steering.
Frameworks (1)
framework
- ScaleRLusesRL recipe used for training.
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- Machine learning paradigm where agents learn to maximize cumulative reward through interaction.
- A competing alignment approach that fine-tunes models based on human evaluator feedback; discussed as complementary to SOO
- Alternative framework for agent behavior; based on reward maximization rather than free energy minimization.
- Method for fine-tuning LMs based on human preferences; mentioned as combining RL and LMs.
- AI training method inspired by behaviorism, used for autonomous cars and drones; cited as bioinspired success
- Operational definition of RL used throughout the paper, quoted from Sutton.
- The hypothesis that cellular collectives can be trained via rewards/punishments to produce specific morphological outcomes.
- Third post-training stage; contributes only marginal further reductions in persona expression