HyperAgents

简体中文

Core mechanism evidence. Paper and local evaluations are separate.

Paper information

Field	Value
Paper	arXiv
Affiliations	University of British Columbia, first author's affiliation
First public release	2026-03-19 (arXiv v1)
Authors' code	Author code.
Adapter / method	hyperagents
Local code	src/auto_research/agent_research/official_meta_backfill_20261004.py

Original paper

Background and contribution

Put the task agent and meta-agent in one editable program. Accumulate variants in an open archive; favor strong parents and those with fewer descendants, allowing the improvement mechanism itself to evolve.

flowchart LR
  A[Open archive] --> P[Performance and exploration weights] --> S[Parent selection] --> U[Bounded archive update]

Key figure from the paper

Original Figure 1: HyperAgents. From the paper; copyright belongs to the authors. Not a local result.

Core mechanism

On student trajectory \(y\), compute teacher–student differences \(d_t\). Select \(S_k=\operatorname{TopK}_t(d_t)\) and optimize \(\mathcal L=-|S_k|^{-1}\sum_{t\in S_k}\log p_T(y_t|y_{<t})/p_\theta(y_t|y_{<t})\).

Paper results

Paper evaluations cover code, review, robotic reward design, and IMO grading, with evolving evaluation analysis, cost planning, and memory. No attributable production A/B.

Local reproduction

Three-seed mechanism metrics compare disabled/default to enabled. Diagnostic only, no formal ranking.

Reproduction limits

Only parent sampling and bounded structured archive updates. Never run generated code or shell; not self-referential code-edit capability.