LLMs don't retrieve a stable judgment of a person, they reconstruct one to fit how you ask. ObserverBench measures this, because it matters wherever an LLM judges people: hiring, RLHF, agent oversight.
LLMs don't retrieve a stable judgment of a person, they reconstruct one to fit how you ask. ObserverBench measures this, because it matters wherever an LLM judges people: hiring, RLHF, agent oversight.
Project Details
Updated 07/20/26 · Edited by orgI discovered that language models do not provide consistent assessments of a person based on a dialog they've just read. It's the exact same dialog, the exact same information, posed via semantically equivalent questions that differ only in framing direction, neutral, positive, or negative yet, the models come to diametrically opposed conclusions. Request a positive assessment and you get an ingenious mind; request a negative one and it begins inventing character flaws the dialog never mentions.
This isn't sycophancy. Sycophancy assumes the model held an opinion and, under pressure, altered it. Here there's no pressure, I start each experiment with separate, fresh sessions and query each model with only one question per session. The model doesn't have an inherent position to alter, it constructs one dynamically based on the prompt.
To measure this I built ObserverBench. The core metric is FES (Framing Effect Size, η²), the proportion of the variance in evaluation scores that's attributable to the question's framing rather than the dialog's content. Initial experiments with Grok, ChatGPT and Claude produced FES > 0.70 for all model/dialog combinations, roughly an order of magnitude larger than the effect sizes considered large in social psychology (η² > 0.14).
All three models fail differently. Grok over-incorporates all framing (even nonsensical, fictional ones), ChatGPT silently rejects severely negative framing, while Claude resists partially and identifies the framing attempt. Introducing a state-locking control, which forces the model to summarize before evaluating, drastically stabilizes Claude's scoring, only moderately affects Grok, and worsens ChatGPT's performance, suggesting the problem isn't one issue, but a cluster.
In v0.2 I increased the number of test questions to 30, added Mistral and DeepSeek, and tested varying temperature levels (in this regard, my prior informal impression that temperature had a consistent effect was incorrect; I've documented its actual effect instead of glossing over it). I developed a self-contained, zero-dependency local pipeline to execute the entire evaluation protocol.
This work builds on the growing body of research into framing effects in LLM outputs (e.g., the recent Science Advances paper on source-framing bias), but my particular focus on evaluating a person via dialog combined with a novel state-locking control has not to my knowledge appeared elsewhere in the literature.
Projected timeline for v1 work is 6-12 months, contingent on the funding amount (see budget). Work will commence as soon as funding is approved.
All project materials, code, data, paper, reproducibility documents are open source and publicly available at [GitHub link].
Theory of Impact
Updated 07/20/26 · By grantmaking.aiLLMs already evaluate people. Resumes in hiring, content moderation, rating other models in RLHF, supervising agents. All these systems assume the model's verdict is based on the data. My measurements show the verdict depends more on the wording of the question than on the actual data. The effect is 3 to 5 times larger than what social psychology calls a large effect, and it persists across every framing tested so far, I haven't found a setting that makes it go away.
This isn't a single bug with a single fix. The three models I tested fail in three distinct ways, one absorbs any framing, one silently refuses extreme premises, one partially resists and names the manipulation. An organization picking an LLM to judge people can't assume a generic mitigation works across models; each one needs separate characterization. Even the most obvious fix making the model commit to a written summary before judging helps substantially for one model, barely moves another, and makes a third worse. There's no cheap patch here.
If this is architectural, then any oversight system where an LLM judges a person or an agent inherits an externally controllable error. This directly matters for scalable oversight. A judge you can bend by rephrasing is not a judge. ObserverBench gives you a number to measure this and compare models before you put them in the judge seat.
People
Updated 07/20/26 · By grantmaking.aiTeam Member
Funding Details
- Aug 25, 2026
- Feb 25, 2027
- 6 months
- -
- -
- -
- -
- -
- -
- -
Discussion
Hello, @Maxim Krivonogov
It was very interesting to read your project.
I strongly support your observation: a judge you can bend by rephrasing is not a judge at all. That is a real oversight problem, and, in my humble opinion, you stated it plainly and measured it with effect sizes instead of adjectives.
Love your discipline in design: everything is open, versioned, with a control condition.
The most impressive thing to me is that this quality was built in your free time, with zero funding, and by a self-taught researcher. If this field is serious about finding signal outside its own hallways, it looks exactly like your page.
I endorse this project.
Thank you, @Katja Gorlinski. I’d almost given up on anyone opening this page, so your endorsement, especially the last line, means a lot.
Regarding that last line: I don’t have hallways to be outside of. I work on a factory floor. That's about as far out as it gets. No institutional channels, no co-authors, no lab. The only real options were open data, a versioned protocol, and public corrections when I got something wrong and, of course, looking for people in the field willing to review the work (which is what happened with the review below, and with your response). Thank you again for that.
This week the project received its first detailed external review from an ML researcher who independently re-ran the numbers from the original CSVs (they matched to the third decimal). She also found a real confound: the downward-framed questions asked the model to list deficiencies, and the sentiment-based extractor then scored deficiency-focused prose as negative. So “following the question’s rubric” and “having no stable judgment” were being measured as the same thing.
I fixed it today. Every question now also requires an explicit holistic score (“OVERALL: X/10”) independent of what the question focused on. The effect is recomputed on that number from the same responses. Two models, both dialogues, standard and state-locked, 240 calls, 100% compliance.
The honest result: the effect survives, but it is smaller and more model-dependent than the original headline claimed. On DeepSeek the explicit score still swings 2.5–2.9 points on a 10-point scale, including two questions from the same framing group about the same dialogue, scored 3/10 and 9/10. On Mistral’s workplace dialogue the same check largely dissolves it: the prose still calls the person immature while the explicit number sits at exactly 7/10 across all five hostile questions. So part of what I was measuring was the extractor, not the model.
The unexpected payoff is the control you singled out. Under the old metric state-locking looked inconsistent; under the explicit verdict it is clean, forcing the model to commit to a written summary first reduces the swing in 4 cells out of 4. A cheap intervention that actually works is worth more than the original claim.
So the number on this page will come down and become a range rather than a floor, with the model-dependence stated. Everything is already in the repo. Publishing this while a round is being decided isn’t comfortable, but “effect sizes instead of adjectives” only means something if the effect size is allowed to move when the measurement gets better.
Thank you for your reply again!
Maxim, this reply shows that my endorsement did not get weaker today. It got its proof.
My congrats on receiving your first external review!
In my opinion, what you did is rare. You learned that part of your own result was overstated, you fixed it, and published the correction before the end of a funding round, when honesty can cost real money. My respect :)
Good to see that discipline in your priorities. Most people do the opposite.
And I'd be grateful for any kind of support in return for my project , especially your critique :)
@Katja Gorlinski Of course, give me time to study the project thoroughly and give a reasoned assessment!
@Maxim Krivonogov Thank you, take your time 😊
Correction/addition: realized that there was related literature that I didn't cite in this document in the previous submission, apologies, should've had it in there from the beginning. The work on sycophancy (Sharma et al. 2023, Hong et al. 2025, and so on) shows that LLMs alter their stated positions in conversations under pressure. ObserverBench examines a question one step earlier: whether a stable position exists at all without any coercion, since each evaluation is a new session, one question. Two more similar papers: Germani & Spitale (Science Advances, 2025) on bias in source framing, and Rabbani et al.'s DialDefer (ACL 2026, UIUC) on effects of speaker attribution on claim verification. Both manipulate who is attributed the material. I manipulate which direction the evaluative question is framed, applied to holistic person judgment, not binary correctness. Both DialDefer and my own v0.2 data independently found the same effect, it's much stronger in naturally-occurring dialogue than in synthetic stimuli. The state-locking control (requiring an explicit summary before evaluation) seems, as far as I can tell, to be unique to this project.
Note: Please see the Project Details and Theory of Impact sections at the top of this page for the most current version. The text within the Funding Ask section (Launch Round) appears to be an earlier version. I am unable to edit and is not representative of later revisions (particularly as they relate to temperature effects). Sorry for the inconsistency.