arXiv subjects

Peter Evans

Publications and source records attributed to Peter Evans.

3 recordsLinked to original sources

arXiv2026

Learning by Doing: Autonomously Compiling and Curating Agent Skills from Verified Traces

Memory systems for production LLM agents mostly store semantic memory (facts about users and domains) or episodic memory (earlier interactions). Research systems such as Trace2Skill, WikiSkill, and SkillRevise learn procedures from agent runs, package them as Skills, and evaluate them on benchmarks. We describe Zep Agent Skills, a production service that gives deployed agents this procedural memory. The service uses only verified attempts as evidence, cites the supporting events for each field of a Skill, and governs the admission and distribution of each Skill version. The service records each task attempt as a Trace with ordered events, an outcome, and a verification strength. A compiler with separate Reflector and Curator roles turns verified Traces into versioned Skills. Scans and admission gates control which candidates become active, and verified use outcomes change the search rank of each Skill. The Agent Skills format is an open standard that Anthropic originated and donated to the Agentic AI Foundation, so any host of the format can load a compiled Skill. Agents retrieve Skills through lexical search with progressive disclosure, over an API or MCP. We evaluate Zep Agent Skills on $τ^3$-bench banking, AppWorld, and EnterpriseOps-Gym. Skills compiled from the Traces of a teacher model increase the pass rate of three weaker executor models on the same tasks, at every reasoning effort, by 4.5 to 18.8 points. Most of the gain is on the tasks that the teacher solved. On ALFWorld, Skills compiled from the runs of a small model give a gain similar to or larger than the gain that prior work reports for continuous fine-tuning, with no training.

cs.AI↗

arXiv2017

Do Reichenbachian Common Cause Systems of Arbitrary Finite Size Exist?

The principle of common cause asserts that positive correlations between causally unrelated events ought to be explained through the action of some shared causal factors. Reichenbachian common cause systems are probabilistic structures aimed at accounting for cases where correlations of the aforesaid sort cannot be explained through the action of a single common cause. The existence of Reichenbachian common cause systems of arbitrary finite size for each pair of non-causally correlated events was allegedly demonstrated by Hofer-Szabó and Rédei in 2006. This paper shows that their proof is logically deficient, and we propose an improved proof.

stat.OT↗

arXiv2010

New Slant on the EPR-Bell Experiment

The best case for thinking that quantum mechanics is nonlocal rests on Bell's Theorem, and later results of the same kind. However, the correlations characteristic of EPR-Bell (EPRB) experiments also arise in familiar cases elsewhere in QM, where the two measurements involved are timelike rather than spacelike separated; and in which the correlations are usually assumed to have a local causal explanation, requiring no action-at-a-distance. It is interesting to ask how this is possible, in the light of Bell's Theorem. We investigate this question, and present two options. Either (i) the new cases are nonlocal, too, in which case action-at-a-distance is more widespread in QM than has previously been appreciated (and does not depend on entanglement, as usually construed); or (ii) the means of avoiding action-at-a-distance in the new cases extends in a natural way to EPRB, removing action-at-a-distance in these cases, too. There is a third option, viz., that the new cases are strongly disanalogous to EPRB. But this option requires an argument, so far missing, that the physical world breaks the symmetries which otherwise support the analogy. In the absence of such an argument, the orthodox combination of views -- action-at-a-distance in EPRB, but local causality in its timelike analogue -- is less well established than it is usually assumed to be.

quant-ph↗