Jul 29, 2024

AI Alignment Toolkit Research Assistant

Luciano Hanyon Wu

The AI Alignment Toolkit Research Assistant is designed to augment AI alignment researchers by addressing two key challenges: proactive insight extraction from new research and automating alignment research using AI agents. This project establishes an end-to-end pipeline where AI agents autonomously complete tasks critical to AI alignment research

Reviewer's Comments

Reviewer's Comments

Arrow
Arrow
Arrow
Arrow
Arrow

I really like the idea of this project. I like how this can be used for brainstorming and exploring relevant scientific literature. I would be excited to see more features being added that let the user guide the model in the right direction if necessary.

I think this tool is great - given the notoriously fast pace of the field, I think any tool that can help researchers use their reading time more effectively would be extremely valuable. I like that you started with the Alignment Data Set too, but could easily switch to a different data source in the future. If you were going to spend more time on this, I’d feel excited about having it do one thing excellently, before spending time improving the output of the agents - simply having an accurate relevancy score for a paper would be useful to researchers. For example, it could work as a personalised weekly digest for researchers - they enter a summary of the research direction, and then the tool sends them the new papers it thinks are most relevant every Monday morning. Researchers could then give feedback on the accuracy of the relevancy score.

I think this project is a great start! Pulling from the research database and seeing how different documents relate to your project makes sense. I think the things that need to be improved are the following: 1) making it more consumable, researchers are not gonna read that much text for each paper so there needs to be a way to show only the things that would interest the individual researchers and then allowing them to dive deeper if they would like to 2) using some cross-paper brainstorming that don’t just look at one paper, but multiple papers to come up with ideas 3) Allow the researcher to provide direction at different parts of the pipeline in order to improve the overall quality of the output (”I want more papers like this and less like that”, “only tell me about how x relates to my work, ignore the rest”, “once you come across a high-relevance paper, do a deep dive looking at other things to improve the output even more”). I think this kind of project will be useful infrastructure that we can later throw more capable AI agents at to help us generate better research ideas. It could also be done periodically as you are working through projects so that it pro-actively suggests new things for you based on what you are working on and your open questions. Lastly, improving the prompts from the model so that it is providing more alignment-relevant outputs would be valuable, instead of leaving the prompt to be more general (e.g. if the project is related to mech interp, then you pull up mech interp specific prompt that makes the model better at gathering insights).

Cite this work

@misc {

title={

AI Alignment Toolkit Research Assistant

},

author={

Luciano Hanyon Wu

},

date={

7/29/24

},

organization={Apart Research},

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

howpublished={https://apartresearch.com}

}

Recent Projects

View All

View All

Feb 2, 2026

Markov Chain Lock Watermarking: Provably Secure Authentication for LLM Outputs

We present Markov Chain Lock (MCL) watermarking, a cryptographically secure framework for authenticating LLM outputs. MCL constrains token generation to follow a secret Markov chain over SHA-256 vocabulary partitions. Using doubly stochastic transition matrices, we prove four theoretical guarantees: (1) exponentially decaying false positive rates via Hoeffding bounds, (2) graceful degradation under adversarial modification with closed-form expected scores, (3) information-theoretic security without key access, and (4) bounded quality loss via KL divergence. Experiments on 173 Wikipedia prompts using Llama-3.2-3B demonstrate that the optimal 7-state soft cycle configuration achieves 100\% detection, 0\% FPR, and perplexity 4.20. Robustness testing confirms detection above 96\% even with 30\% word replacement. The framework enables $O(n)$ model-free detection, addressing EU AI Act Article 50 requirements. Code available at \url{https://github.com/ChenghengLi/MCLW}

Read More

Feb 2, 2026

Prototyping an Embedded Off-Switch for AI Compute

This project prototypes an embedded off-switch for AI accelerators. The security block requires periodic cryptographic authorization to operate: the chip generates a nonce, an external authority signs it, and the chip verifies the signature before granting time-limited permission. Without valid authorization, outputs are gated to zero. The design was implemented in HardCaml and validated in simulation.

Read More

Feb 2, 2026

Fingerprinting All AI Cluster I/O Without Mutually Trusted Processors

We design and simulate a "border patrol" device for generating cryptographic evidence of data traffic entering and leaving an AI cluster, while eliminating the specific analog and steganographic side-channels that post-hoc verification can not close. The device eliminates the need for any mutually trusted logic, while still meeting the security needs of the prover and verifier.

Read More

Feb 2, 2026

Markov Chain Lock Watermarking: Provably Secure Authentication for LLM Outputs

We present Markov Chain Lock (MCL) watermarking, a cryptographically secure framework for authenticating LLM outputs. MCL constrains token generation to follow a secret Markov chain over SHA-256 vocabulary partitions. Using doubly stochastic transition matrices, we prove four theoretical guarantees: (1) exponentially decaying false positive rates via Hoeffding bounds, (2) graceful degradation under adversarial modification with closed-form expected scores, (3) information-theoretic security without key access, and (4) bounded quality loss via KL divergence. Experiments on 173 Wikipedia prompts using Llama-3.2-3B demonstrate that the optimal 7-state soft cycle configuration achieves 100\% detection, 0\% FPR, and perplexity 4.20. Robustness testing confirms detection above 96\% even with 30\% word replacement. The framework enables $O(n)$ model-free detection, addressing EU AI Act Article 50 requirements. Code available at \url{https://github.com/ChenghengLi/MCLW}

Read More

Feb 2, 2026

Prototyping an Embedded Off-Switch for AI Compute

This project prototypes an embedded off-switch for AI accelerators. The security block requires periodic cryptographic authorization to operate: the chip generates a nonce, an external authority signs it, and the chip verifies the signature before granting time-limited permission. Without valid authorization, outputs are gated to zero. The design was implemented in HardCaml and validated in simulation.

Read More

Feb 2, 2026

Markov Chain Lock Watermarking: Provably Secure Authentication for LLM Outputs

We present Markov Chain Lock (MCL) watermarking, a cryptographically secure framework for authenticating LLM outputs. MCL constrains token generation to follow a secret Markov chain over SHA-256 vocabulary partitions. Using doubly stochastic transition matrices, we prove four theoretical guarantees: (1) exponentially decaying false positive rates via Hoeffding bounds, (2) graceful degradation under adversarial modification with closed-form expected scores, (3) information-theoretic security without key access, and (4) bounded quality loss via KL divergence. Experiments on 173 Wikipedia prompts using Llama-3.2-3B demonstrate that the optimal 7-state soft cycle configuration achieves 100\% detection, 0\% FPR, and perplexity 4.20. Robustness testing confirms detection above 96\% even with 30\% word replacement. The framework enables $O(n)$ model-free detection, addressing EU AI Act Article 50 requirements. Code available at \url{https://github.com/ChenghengLi/MCLW}

Read More

Feb 2, 2026

Prototyping an Embedded Off-Switch for AI Compute

This project prototypes an embedded off-switch for AI accelerators. The security block requires periodic cryptographic authorization to operate: the chip generates a nonce, an external authority signs it, and the chip verifies the signature before granting time-limited permission. Without valid authorization, outputs are gated to zero. The design was implemented in HardCaml and validated in simulation.

Read More

Feb 2, 2026

Markov Chain Lock Watermarking: Provably Secure Authentication for LLM Outputs

We present Markov Chain Lock (MCL) watermarking, a cryptographically secure framework for authenticating LLM outputs. MCL constrains token generation to follow a secret Markov chain over SHA-256 vocabulary partitions. Using doubly stochastic transition matrices, we prove four theoretical guarantees: (1) exponentially decaying false positive rates via Hoeffding bounds, (2) graceful degradation under adversarial modification with closed-form expected scores, (3) information-theoretic security without key access, and (4) bounded quality loss via KL divergence. Experiments on 173 Wikipedia prompts using Llama-3.2-3B demonstrate that the optimal 7-state soft cycle configuration achieves 100\% detection, 0\% FPR, and perplexity 4.20. Robustness testing confirms detection above 96\% even with 30\% word replacement. The framework enables $O(n)$ model-free detection, addressing EU AI Act Article 50 requirements. Code available at \url{https://github.com/ChenghengLi/MCLW}

Read More

Feb 2, 2026

Prototyping an Embedded Off-Switch for AI Compute

This project prototypes an embedded off-switch for AI accelerators. The security block requires periodic cryptographic authorization to operate: the chip generates a nonce, an external authority signs it, and the chip verifies the signature before granting time-limited permission. Without valid authorization, outputs are gated to zero. The design was implemented in HardCaml and validated in simulation.

Read More

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