Verifiable Social Reasoning for LLM Assistants

简体中文

L1 core-mechanism diagnostic. Runs only the observation-safe evaluation state machine; does not call 12 external LLMs.

Paper information

Field	Details
Paper	Verifiable Social Reasoning for LLM Assistants
Organization	First-author affiliations: Google Research / Hebrew University of Jerusalem / University of Cambridge
First public release	2026-09-15 (arXiv v1)
Authors' code	No public author repository was found as of 2026-09-16.
Adapter / method	fuse-evaluator
Local reproduction code	src/auto_research/agent_research/latest_20260916.py

Original paper

Motivation and method

Multi-agent simulations with verifiable hidden motivations evaluate user retelling, framing bias, and social reasoning.

The local reference kernel retains decisive gates, weights, or state transitions and exposes auditable intermediates so related methods can be compared at a common budget.

flowchart LR
  I[Public input and current state] --> M[fuse-evaluator mechanism]
  M --> A[Auditable intermediates]
  A --> O[Output or state update]

Key figure from the paper

Figure 1: Overview of the paper's method and its key components. From the original paper; copyright remains with the authors. Click the image for the source.

Core computation

The implementation preserves the paper's decisive state transitions and reports weights, gates, depth, stages, or cache statistics. Fixed budgets and three seeds check mechanism invariants only, not paper-scale training.

Paper results

The paper's capability, speed, or cost findings are original evidence only. Local metrics are not directly compared with them or extrapolated to online gains.

Local reproduction

See the three-seed results. The common receipt marks diagnostic_only=true, so the public dashboard does not treat it as a full capability improvement.

Reproduction boundaries

The local run executes only an observation-safe evaluation state machine and does not call 12 external LLMs.