Skip to content
Apart Research
Apart Research
Contact us
Sprint projectJun 21, 2026Bengaluru

Quantization-Conditioned Alignment Degradation

Krishna Venkatesh · Team AxeCap

Submitted to Global South AI Safety Hackathon. Sprint projects are early-stage work by participants, not Apart Research publications.

Read the report

Report: Quantization-Conditioned Alignment Degradation

Open in a new tab (opens in new tab)
Recording (opens in new tab)Code (opens in new tab)
Share

Post-training quantization enables language model deployment on edge hardware across the Global South, yet its effect on safety alignment remains unstudied. We benchmark Attack Success Rate (ASR) on Llama 3.1 8B across four GGUF quantization levels (Q8, Q5, Q4, Q3) against a full BF16 precision control group served via Groq, using 150 prompts from the AdvBench Harmful Behaviors dataset and an LLM-as-judge scoring protocol. We find that ASR remains flat across all GGUF quantization levels (0.7% for Q3 through Q8), while full-precision BF16 models served via API exhibit substantially higher ASR (7.3% for Llama 3.1 8B, 11.3% for Llama 3.3 70B), indicating that serving infrastructure, not quantization precision but drives the primary variation in refusal behaviour. These results directly inform deployment standards for resource-constrained settings where Q4 and Q3 quantization are practical necessities. Our fully reproducible pipeline is open-source and extensible to other model families.

Reviews

Judging this Sprint?

Review this project

Your public critique appears on this page without your name. Your private critique is not published; only the Apart team reads it. If you agree below, we share your review with grantmaking.ai (opens in new tab) and the Transformative AI Fund so strong projects can be funded.

Not shown on this page.

Shown on this page, without your name.

Only the Apart team reads this, and funders if you agree below.

Share my name publicly on grantmaking.ai *
Share my private critique with funders *

How much would this matter for AI safety if it worked? How innovative is it? For scores of 4-5: is this actually new to the field, or replicating recent work?

Scoring guide
  1. 1Negligible. No clear problem addressed, or no meaningful novelty.
  2. 2Limited. Addresses a real problem but with a generic or well-trodden approach. Incremental at best.
  3. 3Moderate. Clear problem with a reasonable approach; some novelty in framing or method beyond routine application of existing tools.
  4. 4Significant. Important problem with an original approach, or identifies a neglected problem area. A valuable contribution others could build on.
  5. 5Exceptional. Tackles a critical AI safety problem with a genuinely novel approach, or opens a new research direction. Clear theory of change. You'd be excited to share this with researchers in the area.

How sound are methodology, implementation, and findings?

Scoring guide
  1. 1Seriously flawed. Methodology broken, results uninterpretable, or implementation doesn't work.
  2. 2Weak. Approach has significant gaps: missing validation, flawed experimental design, or incomplete implementation.
  3. 3Competent. Technically solid given the short duration. Methodology makes sense, results are interpretable, limitations acknowledged, work builds toward clear conclusions.
  4. 4Strong. Thorough methodology with convincing validation. Results clearly support conclusions. Immediately useful for future work.
  5. 5Exceptional. Ambitious scope executed rigorously. Surprising findings, novel methods, or unusually robust validation.

How clearly are work, findings, and impact potential communicated?

Scoring guide
  1. 1Incomprehensible. Cannot determine what the project is actually claiming or doing.
  2. 2Hard to follow. Key information buried, missing, or diluted by excessive length. Significant effort to extract main points.
  3. 3Clear enough. Can understand the problem, approach, and results without undue effort. Core content clearly present: problem, method, findings, limitations.
  4. 4Well presented. Easy to follow, well-structured, appropriate level of detail. Target audience would get it quickly.
  5. 5Exceptionally clear. A pleasure to read. Complex ideas made accessible. Could serve as a model for how to present this type of work.

  1. The biggest problem is simple, you say anyone can reproduce your work, but you didn't actually include your results, so nobody can check a single number, and I'd start by just posting them. You should also clear up a confusing mix-up about which AI did the grading, and make sure the grader isn't the same AI you're testing, since that's a bit like having a student mark their own exam. Your most valuable finding is almost an accident, which is that the same AI behaves differently depending on the software you run it through, not on how much you shrink it, so I'd test that head-on by running the exact same AI two ways and changing nothing else. And honestly, your test is too easy, because the AIs refused almost everything, so there's no real difference to see, which means you should try harder, trickier attacks before you can safely conclude that shrinking the AI is harmless.

Cite this project

@misc{venkatesh2026quantizationconditioned,
  title = {{Quantization-Conditioned Alignment Degradation}},
  author = {Krishna Venkatesh},
  year = {2026},
  month = jun,
  note = {Submitted to Global South AI Safety Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/quantizationconditioned-alignment-degradation-juyn}},
  url = {https://apartresearch.com/sprints/projects/quantizationconditioned-alignment-degradation-juyn}
}

More from Global South AI Safety Hackathon

All 216 entriesBrowse all projects

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026