Nous
A statistical learning theory paper on agentic memory decision-making that proposes a framework for learning and certifying memory decisions without prior calibration of information sources. The work argues, from a sample-complexity perspective, that learning decisions and calibrating sources are separable problems, enabling effective memory-based decision-making independently of source calibration.
Timeline (last 90 days)
The arXiv paper "Nous: Learning and Certifying Memory Decisions Before Source Calibration" argues that memory decisions can be learned and their validity certified without recovering source reliability.
The Nous framework explicitly separates learning, calibration, and revision certification.
Certifying the improvement of a decision relative to an informative incumbent policy requires O(l⁻²) new records drawn from the same observation law.
The number of records required to learn and certify useful decisions may be a quadratic order of magnitude fewer than those required to calibrate the source.
A broader model class still preserves the same decision rate and the lower bound for source estimation.
The paper characterizes the precise identifiable interval for policy improvement under an identity-plus-background reporting channel and derives finite-sample certificates that do not depend on pure class anchors.
The method introduces robustness extensions capable of tolerating bounded, history-dependent misspecification and conditional duplication.
Empirical testing used 45,000 held-out variable-state histories and 9,000 episodes, covering three external MiniGrid memory environments with a noisy reporting interface introduced.
In MiniGrid, the new certificates accepted 9/9 improvements over the constant incumbent policy and 4/9 improvements over the last-write-wins policy, whereas the previous certificates accepted none.
The paper acknowledges that strong established inference baselines remain competitive or even superior in many cases.
1 more timeline events
All Facts (11)
The arXiv paper "Nous: Learning and Certifying Memory Decisions Before Source Calibration" argues that memory decisions can be learned and their validity certified without recovering source reliability.
50%UnverifiedThe Nous framework explicitly separates learning, calibration, and revision certification.
50%UnverifiedCertifying the improvement of a decision relative to an informative incumbent policy requires O(l⁻²) new records drawn from the same observation law.
50%UnverifiedThe number of records required to learn and certify useful decisions may be a quadratic order of magnitude fewer than those required to calibrate the source.
50%UnverifiedA broader model class still preserves the same decision rate and the lower bound for source estimation.
50%UnverifiedThe paper characterizes the precise identifiable interval for policy improvement under an identity-plus-background reporting channel and derives finite-sample certificates that do not depend on pure class anchors.
50%UnverifiedThe method introduces robustness extensions capable of tolerating bounded, history-dependent misspecification and conditional duplication.
50%UnverifiedEmpirical testing used 45,000 held-out variable-state histories and 9,000 episodes, covering three external MiniGrid memory environments with a noisy reporting interface introduced.
50%UnverifiedIn MiniGrid, the new certificates accepted 9/9 improvements over the constant incumbent policy and 4/9 improvements over the last-write-wins policy, whereas the previous certificates accepted none.
50%UnverifiedThe paper acknowledges that strong established inference baselines remain competitive or even superior in many cases.
50%UnverifiedThe researchers position this work as a statistical account of when memory decisions can be learned and certified without recovering source reliability, rather than as a memory algorithm that outperforms all others in every scenario.
50%