Check out the results from the AI Safety X Physics Grand Challenge! 👉

Jan 20, 2025

View Related Sprint

Neural Seal

Kailash Balasubramaniyam, Mohammed Areeb Aliku, Eden Simkins

Neural Seal is an AI transparency solution that creates a standardized labeling framework—akin to “nutrition facts” or “energy efficiency ratings”—to inform users how AI is deployed in products or services.

Download

Review Project

See Code

Reviewer's Comments

Shivam Raval

The proposal introduces Neural Seal, an AI transparency solution that creates a standardized labeling framework of “nutrition facts” that inform users on the level of AI involvement in products or services. Visibility for AI involvement is limited but necessary in many domains like finance, healthcare, and social media, and Neural Seal aims to create a universal labeling standard for AI transparency. The solution involves a standardized structural evaluation with a multi-step questionnaire on the level of AI usage and generating a color-coded rating (A/B/C/D) that represents AI involvement in a digestible manner. The team showcases a demo build of the questionnaire and mockups of the labeling system, along with a yearly plan for building a prototype, integrating interpretability techniques, and widespread adoption of the standardization metrics.

I think this is a great idea! The proposal highlights an important need and the team understands that the consumer-facing aspect of their product requires simple and intuitive metrics and ratings. Some things to consider that can greatly strengthen the proposal:

1. Different industries and use-cases can be classified as low/medium/high risk and might require different rubrics to assess the impact of AI involvement

2. Some guidelines or information on how the level of AI involvement is categorized can be helpful for companies filling out the proposal

3. Details on the breakdown of different areas and how the AI usage in these areas is categorized, along with how the impact score can be calculated would add a strong algorithmic component to the proposal

4. The vision of creating a singular standardized metric across all industries might require extended research. I would suggest starting with a few use cases and showing some proof of concept for those areas (which might require different metrics, language, and jargon that the area-specific target users are familiar with) and using the insights to inform what a universal standardized metric might look like.

5. Some visual indication of the AI involvement (like chemical/bio/fire hazard symbols on instruments and chemicals) can be an additional way to showcase the rating in an accessible manner.

6. On the technical side, using a react-based framework can be relatively easy to implement and flexible to modify and build upon, maybe using Claude can be helpful since it can natively render react components.

7. For explainable AI methods, it might be important to consider other interpretability approaches that have shown great promise for large language models (like patching, probing, SAEs, etc) since future systems will most definitely include an LLM/multimodal AI-based agent, and explainable AI methods like LIME/SHAP that are compliance supported might not be sufficient to explain the inner working of these highly capable systems.

I might have more comments and suggestions, so feel free to reach if you need any further feedback. Good luck to the team!

Finn Metz

I like the idea of pushing for this scale for end-consumer use. A few points: What are the incentives from relevant stakeholder? It seems like the current solution is only a questionnaire. Right now the only reason why someone at t company fills out a questionnaire is compliance or PR. Both are currently not given, so time might be best spend to look at where legislation is on this and who might be closest to adopting such a standard and what is in the way. I really like the mention of LIME and SHAP, I think it would have been very cool to have them included in the demo/product. That would seems like the most interesting part of this project, unless you are very keen and involved on the policy side of things. With this project there is definitely a risk of just becoming unnessecary overhead for the adoption of normal AI, while not addressing actual existential risk (like the cookie banner), so that would need to be kept in mind. Would spend most of my energy in finding someone who this is useful for, maybe in policy or maybe for some internal AI governance/risk teams at large enterprises.

Cite this work

@misc {

title={

Neural Seal

author={

Kailash Balasubramaniyam, Mohammed Areeb Aliku, Eden Simkins

date={

1/20/25

organization={Apart Research},

note={Research submission to the research sprint hosted by Apart.},

howpublished={https://apartresearch.com}

}

Recent Projects

View All

Jul 28, 2025

Local Learning Coefficients Predict Developmental Milestones During Group Relative Policy Optimization

In this work, we investigate the emergence of capabilities in reinforcement learning (RL) by framing them as developmental phase transitions. We propose that the individual components of the reward function can serve as direct observables for these transitions, avoiding the need for complex, derived metrics. To test this, we trained a language model on an arithmetic task using Group Relative Policy Optimization (GRPO) and analyzed its learning trajectory with the Local Learning Coefficient (LLC) from Singular Learning Theory. Our findings show a strong qualitative correlation between spikes in the LLC—indicating a phase transition—and significant shifts in the model's behavior, as reflected by changes in specific reward components for correctness and conciseness. This demonstrates a more direct and scalable method for monitoring capability acquisition, offering a valuable proof-of-concept for developmental interpretability and AI safety. To facilitate reproducibility, we make our code available at \url{github.com/ilijalichkovski/apart-physics}.

Jul 28, 2025

AI agentic system epidemiology

As AI systems scale into decentralized, multi-agent deployments, emergent vulnerabilities challenge our ability to evaluate and manage systemic risks.

In this work, we adapt classical epidemiological modeling (specifically SEIR compartment models) to model adversarial behavior propagation in AI agents.

By solving systems of ODEs describing the systems with physics-informed neural networks (PINNs), we analyze stable and unstable equilibria, bifurcation points, and the effectiveness of interventions.

We estimate parameters from real-world data (e.g., adversarial success rates, detection latency, patching delays) and simulate attack propagation scenarios across 8 sectors (enterprise, retail, trading, development, customer service, academia, medical, and critical infrastructure AI tools).

Our results demonstrate how agent population dynamics interact with architectural and policy design interventions to stabilize the system.

This framework bridges concepts from dynamical systems and cybersecurity to offer a proactive, quantitative toolbox on AI safety.

We argue that epidemic-style monitoring and tools grounded in interpretable, physics-aligned dynamics can serve as early warning systems for cascading AI agentic failures.

Jul 28, 2025

Momentum–Point-Perplexity Mechanics in Large Language Models

This work analyzes the hidden states of twenty different open-source transformer language models, ranging from small to medium size and covering five major architectures. The key discovery is that these models show signs of "energy conservation" during inference—meaning a certain measure combining changes in hidden states and token unpredictability stays almost constant as the model processes text.

The authors developed a new framework inspired by physics to jointly analyze how hidden states and prediction confidence evolve over time. They propose that transformers' behavior can be understood as following certain mechanical principles, much like how physical systems follow rules like conservation of energy.

Their experiments show that this conserved quantity varies very little between tokens, especially in untrained (random-weight) models, where it's extremely stable. In pre-trained models, the average energy drops more due to training, but there are larger relative fluctuations from token to token.

They also introduce a new method, based on this framework, for controlling transformer outputs by "steering" the hidden states. This method achieves good results—producing completions rated as higher in semantic quality, while still maintaining the same kind of energy stability.

Overall, the findings suggest that viewing transformer models through the lens of physical mechanics gives new, principled ways to interpret and control their behavior. It also highlights a key difference: random models behave more like balanced systems, while trained models make quicker, more decisive state changes at the cost of less precise energy conservation.

Jul 28, 2025

Local Learning Coefficients Predict Developmental Milestones During Group Relative Policy Optimization

Jul 28, 2025

AI agentic system epidemiology

As AI systems scale into decentralized, multi-agent deployments, emergent vulnerabilities challenge our ability to evaluate and manage systemic risks.

In this work, we adapt classical epidemiological modeling (specifically SEIR compartment models) to model adversarial behavior propagation in AI agents.

By solving systems of ODEs describing the systems with physics-informed neural networks (PINNs), we analyze stable and unstable equilibria, bifurcation points, and the effectiveness of interventions.

Our results demonstrate how agent population dynamics interact with architectural and policy design interventions to stabilize the system.

This framework bridges concepts from dynamical systems and cybersecurity to offer a proactive, quantitative toolbox on AI safety.

We argue that epidemic-style monitoring and tools grounded in interpretable, physics-aligned dynamics can serve as early warning systems for cascading AI agentic failures.

Jul 28, 2025

Local Learning Coefficients Predict Developmental Milestones During Group Relative Policy Optimization

Jul 28, 2025

AI agentic system epidemiology

As AI systems scale into decentralized, multi-agent deployments, emergent vulnerabilities challenge our ability to evaluate and manage systemic risks.

In this work, we adapt classical epidemiological modeling (specifically SEIR compartment models) to model adversarial behavior propagation in AI agents.

By solving systems of ODEs describing the systems with physics-informed neural networks (PINNs), we analyze stable and unstable equilibria, bifurcation points, and the effectiveness of interventions.

Our results demonstrate how agent population dynamics interact with architectural and policy design interventions to stabilize the system.

This framework bridges concepts from dynamical systems and cybersecurity to offer a proactive, quantitative toolbox on AI safety.

We argue that epidemic-style monitoring and tools grounded in interpretable, physics-aligned dynamics can serve as early warning systems for cascading AI agentic failures.

Jul 28, 2025

Local Learning Coefficients Predict Developmental Milestones During Group Relative Policy Optimization

Jul 28, 2025

AI agentic system epidemiology

As AI systems scale into decentralized, multi-agent deployments, emergent vulnerabilities challenge our ability to evaluate and manage systemic risks.

In this work, we adapt classical epidemiological modeling (specifically SEIR compartment models) to model adversarial behavior propagation in AI agents.

By solving systems of ODEs describing the systems with physics-informed neural networks (PINNs), we analyze stable and unstable equilibria, bifurcation points, and the effectiveness of interventions.

Our results demonstrate how agent population dynamics interact with architectural and policy design interventions to stabilize the system.

This framework bridges concepts from dynamical systems and cybersecurity to offer a proactive, quantitative toolbox on AI safety.

We argue that epidemic-style monitoring and tools grounded in interpretable, physics-aligned dynamics can serve as early warning systems for cascading AI agentic failures.

This work was done during one weekend by research workshop participants and does not represent the work of Apart Research.

Careers

AI Safety Ideas

Code of Conduct

Responsible disclosure policy

hello@apartresearch.com