ObserverBench: measuring the Observer Problem in LLMs

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

I 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

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

Maxim Krivonogov

Team Member

Funding Details

Start Date: Aug 25, 2026
End Date: Feb 25, 2027
Expected Duration: 6 months
Funding Raised to Date: -
Annual Budget: -
Monthly Burn Rate: -
Current Runway: -
Funding Goal: -
Funding Stage: -
Fiscal Sponsor: -

Grants Received– no grants recorded

Add grant

Funding Asks

grantmaking.ai Launch Round

Applied

Build v1 of an open benchmark, dataset, and tool to measure LLM instability in evaluating a person from dialogue under different question framings, across models, temps, and human baselines.

Minimum

$16,500

Ideal

$44,000

What this ask is

I found that language models cannot give a stable evaluation of a person based on a dialogue they just read. Same dialogue, semantically equivalent questions with different framing, and the model gives you opposite conclusions. Ask for positives and it sees a brilliant thinker. Ask for negatives and it starts inventing flaws that are not even in the text. This is not sycophancy. The model just does not have a stable representation of the person. It reconstructs them from scratch every time based on how you ask.

I built a benchmark that measures this. Started with a pilot on Grok, ChatGPT, and Claude with 10 questions. Then expanded to 30 questions, ran Mistral and DeepSeek across three temperatures, added a state-locking control where the model answers based on its own fixed summary of the person. Wrote a pure Python tool with zero dependencies that runs the whole protocol through a local web interface. Everything is open: code, raw data, paper.

For v1, I need to close the main gaps from the pilot. Repeated runs to estimate variance, 8 models instead of 5, manual verification of all automated scores, a human baseline with at least 10 raters and inter rater reliability, factor analysis of the construct. The concrete output of v1 is a public dataset, the updated benchmark tool, and a paper.

About me: I am a self-taught researcher. I work as a CNC machine operator at a factory and built all of this in my free time, with zero funding.

Theory of impact

LLMs 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 does not go away with temperature or objectivity instructions. 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.

How the money will be spent

Minimum 16500 dollars. My stipend for 6 months part time, 12000. API credits, 2000. Human rater compensation, 1000. Buffer, 1500.

Ideal 44000 dollars. 12 months full time, 30000. API for roughly 480 runs across 8 models, 5000. 20 human raters, 2500. New dialogues, 1500. Infrastructure, 1000. Buffer, 4000. Full budget table in the additional information section.

Discussion

Maxim Krivonogov2d

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.

Maxim Krivonogov2d

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.