Skip to content
Apart Research
Apart Research
Contact us
Sprint projectAug 17, 2026Saarbrücken

Be More Introspective

Ali Haider Khan, Harsh Puri, Fatehbir Singh Gill, Nidhish Pajni, Tanveer Singh · Team Udta Punjab

Submitted to Digital Minds Research Sprint. Sprint projects are early-stage work by participants, not Apart Research publications.

Read the report

Report: Be More Introspective

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

This project is an extension of the previous work done by Lindsey (2026) [1]. Large Language Models can notice the presence of injected concepts and can be aware of its happening. They also demonstrate the ability to recall prior representations and compare them with potential changes at a later stage. It is found that some models can use their ability to recall prior intentions in order to distinguish their own outputs from artificial prefills. Here, in our project, we investigate the introspective nature of LLMs. We have tried to reproduce the results and extend them to a wider range of models. The pipeline we adhered to begins with injecting representations of known concepts in a model’s activations and measuring the influence of those modifications on the model’s self-awareness abilities. Our experiments included Qwen2.5-0.5B, Qwen2.5-32B and Qwen2.5-1.5B, which are small to medium-sized models and demonstrate moderate introspective awareness. Overall, our results indicate that the self-awareness nature is most of the time directly proportional to the size and complexity of the models, as could be seen in the graphs later in the report.

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 the field 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 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. This paper is a partial replication of research on the capacity of LLMs to detect injected concepts. It would be improved by a clearer explanation of why the z-score is a good measure of performance in this task, and figures that more clearly bring out the most important aspects of the results.

  2. I'm a bit hung up on the unsteered baseline. You report z of -0.23 / +1.43 / +2.18 across 0.5B / 1.5B / 32B with no injection at all, so the 32B baseline already clears your own z>2 separation threshold before you do anything to it.

    Fix is cheap though: report steered-minus-unsteered deltas instead of raw z, and show the null distribution. I'd also push back on the vector itself e.g. ("Hi! How are you?" vs a shouted version) isn't really a concept, so I'm not sure this is the best example for testing introspection over semantic content. Lindsey uses semantic concepts, and porting that properly would give you a much stronger replication than what you have here.

    The presentation makes it harder to check any of this. Section 4.1 says Qwen2.5-32B-Instruct but both figures under it are 1.5B plots, so the model carrying your headline has no figure anywhere. Figures 1/3 and 2/4 are duplicates with near-identical captions. Your LLM usage statement describes an LLM judge validated against human coders, but I don't (think I see it). That said, replicating this with open weights is genuinely useful, since the original is stuck behind closed models, and adding a norm-matched random-direction control also a nice add.

    So it's generally good as an experiment, just the baseline is sort of iffy.

    Read full reviewShow less

Cite this project

@misc{khan2026be,
  title = {{Be More Introspective}},
  author = {Ali Haider Khan and Harsh Puri and Fatehbir Singh Gill and Nidhish Pajni and Tanveer Singh},
  year = {2026},
  month = aug,
  note = {Submitted to Digital Minds Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/be-more-introspective-9gxz}},
  url = {https://apartresearch.com/sprints/projects/be-more-introspective-9gxz}
}

More from Digital Minds Research Sprint

All 233 entriesBrowse all projects

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026