Research map / Multi-agent systems and games / Games and equilibria
Path-Specific Objectives for Safer Agent Incentives
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Optimize a path-specific objective to remove the incentive to manipulate user preferences | Content recommendation systems · Content recommendation simulation | user-preference drift | Helped |
| Use a policy baseline rather than the state-baseline or fixed-state estimate (ablation) | PSO content recommendation · Content recommendation simulation | user-preference drift | Helped |
| Use a state baseline to estimate the natural preference distribution (ablation) | PSO content recommendation · Content recommendation simulation | user-preference drift | Helped |
| Use a fixed initial preference state to estimate the natural preference distribution (ablation) | PSO content recommendation · Content recommendation simulation | user-preference drift | Helped |
| Optimize a path-specific objective rather than the ordinary objective | Content recommendation systems · Content recommendation simulation | accuracy | No clear effect |
| Optimize a path-specific objective rather than the ordinary objective | Content recommendation systems · Content recommendation simulation | loyalty drift | No clear effect |
| Change environments every step using context-swapping (ablation) | Content recommendation systems · Content recommendation simulation | accuracy | Helped |
| Change environments every step using context-swapping (ablation) | Content recommendation systems · Content recommendation simulation | user-preference drift | Mixed |
| Use a fixed-intervention PSO instead of ordinary reward maximization | Barging agent · Barging | oracle return | Helped |
| Make the agent epsilon-greedy with ε = 0.1 (ablation) | Fixed-intervention PSO agent · Barging | oracle return | Hurt |
| Let preference dynamics operate even without a deliberate incentive to manipulate preferences (ablation) | Content recommendation systems · Content recommendation simulation | user-preference drift | Hurt |
Closest papers by meaning
- Agent Incentives: A Causal Perspective (AAAI 2021 · similarity 0.88)
- Embedding Safety into RL: A New Take on Trust Region Methods (ICML 2025 · similarity 0.85)
- Safe Exploitative Play with Untrusted Type Beliefs (NeurIPS 2024 · similarity 0.85)
- Incentivized Learning in Principal-Agent Bandit Games (ICML 2024 · similarity 0.84)
- Building a Foundational Guardrail for General Agentic Systems via Synthetic Data (ICLR 2026 · similarity 0.84)
- Emergent Risk Awareness in Rational Agents under Resource Constraints (NeurIPS 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.