Technical Summary: Comprehensive AI Governance Requires Addressing Non-Model Capability Gains

1. Problem Statement

The prevailing paradigm in Frontier AI governance is model-level governance, which operates on the assumption that a model's capability profile is primarily a function of the compute and data used during its pre-training and standard post-training phases. This approach relies heavily on pre-deployment evaluation and mitigation by the model developer.

However, this paper argues that model-level governance is becoming structurally inefficient as capability progress is increasingly driven by "non-model gains." These are improvements in capability that are independent of advances in the base model itself. As the "floor" of accessible capabilities rises through these non-model vectors, the space of actors capable of achieving dangerous outcomes expands, even if the absolute capability ceiling remains with frontier developers. Consequently, risk management strategies hinging solely on pre-deployment evaluation of the base model are insufficient to manage the evolving risk landscape.

2. Methodology and Framework

The paper employs a conceptual analysis and taxonomy development approach to formalize the concept of non-model gains. It does not present new empirical experiments but rather synthesizes emerging trends in AI development (e.g., reasoning models, multi-agent systems, and physical embodiment) to categorize capability gains that fall outside the traditional training loop.

The authors construct a framework distinguishing between:

1. Model-level gains: Derived from scaling pre-training compute and data.
2. Non-model gains: Derived from post-training enhancements, inference-time scaling, or access to restricted assets.

The analysis proceeds by identifying three primary vectors of non-model gains and three emerging future vectors, then mapping these against the limitations of current governance strategies to propose a broader portfolio of governance mechanisms.

3. Key Contributions: Taxonomy of Non-Model Gains

The paper formalizes a taxonomy of capability gains that undermine model-level governance:

A. Primary Vectors

1. Inference Gain: Capability improvements resulting from scaling computational resources at test-time (inference).

   • Mechanism: Techniques like sequential and parallel inference scaling (e.g., "thinking" models) allow smaller or open-weight models to emulate the performance of larger frontier models.
   • Governance Impact: Reduces the performance delta between proprietary and open-weight models, potentially allowing malicious actors to bypass the need for jailbreaking proprietary models by simply scaling inference on accessible open models.

2. Systems Gain: Capability improvements from bespoke post-training enhancements applied after the base model is released.

   • Mechanism: Includes scaffolding, multi-agent orchestration, tool use, and iterative self-improvement schemes.
   • Governance Impact: These gains often come at a low cost relative to pre-training. Once a "recipe" for a high-performance scaffold is discovered, it proliferates freely. Developers cannot anticipate all downstream modifications or the specific "best scaffold" a third party might discover.

3. Asset Gain: Capability increases resulting from combining an AI model with restricted assets unavailable to the original developers.

   • Mechanism: Integration with government-held expertise (e.g., nuclear physics, stealth), specialized hardware (physical embodiments), or classified data (e.g., undisclosed vulnerabilities, pathogen data).
   • Governance Impact: Creates a "capability overhang" that is difficult to estimate or evaluate pre-deployment due to the restricted nature of the assets.

B. Future Vectors

The paper also identifies three emerging vectors that may further challenge governance:

• Embodiment Gain: Integration with physical sensors and actuators, shifting risk from informational to physical safety (e.g., robotics).
• Continual Learning: The ability of models to evolve over time, potentially leading to safety drift or catastrophic forgetting of safety training post-deployment.
• Diffusion Effects: Aggregate effects of widespread deployment, such as monoculture risks or cascading failures in interdependent systems (e.g., financial markets).

4. Results and Governance Analysis

The analysis demonstrates that reliance on pre-deployment elicitation is insufficient because:

• Elicitation Failure: It is increasingly difficult for developers to predict how their models will be modified by downstream actors (scaffolding, inference scaling).
• Mitigation Failure: Even if dangerous capabilities are elicited, the low cost of replicating these gains via non-model vectors makes mitigation difficult.
• Overhang Cost: The potential harm of failing to elicit a capability increases as non-model gains allow more actors to access dangerous capabilities.

Proposed Governance Portfolio

The paper argues that model-level governance must be complemented, not replaced, by a diversified portfolio of governance strategies:

1. Enhanced Model-Level Governance:

   • Evolutive Elicitation Standards: Shifting from "what is the model capable of" to "what can an actor with given resources elicit over a given timespan?"
   • Partnerships: Deepening cooperation with third-party testers, national security authorities, and hardware manufacturers to improve evaluation stacks.
   • Post-Deployment Monitoring: Continuous re-assessment of deployed models via leaderboards and threat intelligence sharing.
   • Forecasting: Investing in methodologies to predict capability overhangs, particularly for inference and system gains.

2. Beyond Model-Level Governance:

   • System Governance: Holding system providers (e.g., those building complex scaffolds or agents) accountable for risk management, particularly when their systems significantly uplift base model capabilities.
   • Entity Governance: Focusing on the organizational structure, incentives, and risk culture of the entities developing or deploying AI, rather than just the model weights.
   • Agent Governance: Addressing risks arising from multi-agent dynamics, delegation, and autonomous interaction (e.g., miscoordination, collusion).
   • Cloud Governance: Exploring privacy-safe monitoring of inference compute usage and content at the cloud provider level to detect misuse patterns.
   • Societal Resilience: Interventions to help communities cope with and recover from AI-enabled disturbances (e.g., securing biological supply chains, economic infrastructure) when technical governance fails to prevent risk materialization.

5. Significance and Claims

The paper claims that the relative returns to model-level intervention may diminish as non-model gains grow in significance. Its primary significance lies in:

• Formalizing the Concept: Providing a rigorous taxonomy of non-model gains that have previously been discussed anecdotally.
• Policy Preparation: Arguing that the AI ecosystem must prepare for a transforming risk landscape where the "base model" is no longer the sole locus of control or risk.
• Strategic Diversification: Demonstrating that a comprehensive governance strategy must move beyond the model to include system, entity, agent, and cloud nodes, as well as societal resilience.

The authors maintain a modest stance, acknowledging that alternative governance strategies face prohibitive practical, legal, and technical barriers. They do not claim these alternatives are immediately implementable or that model-level governance will become redundant. Instead, they argue that a comprehensive analysis and exploration of these broader mechanisms must happen "sooner rather than later" to prevent the governance ecosystem from being left unprepared as non-model gains become the dominant driver of capability progress.