Why LLM Judges Flip Their Verdicts When You Change the Question Format episode artwork

EPISODE · May 19, 2026 · 25 MIN

Why LLM Judges Flip Their Verdicts When You Change the Question Format

from AI Papers: A Deep Dive

Why LLM Judges Flip Their Verdicts When You Change the Question Format Source: https://arxiv.org/abs/2605.16023 Paper was published on May 15, 2026 This episode was AI-generated on May 19, 2026. The script was written by an AI language model and the host voices were synthesized by Eleven Labs. The producer is not affiliated with Anthropic or Eleven Labs. Ask a language model to rate text from 1 to 5 and it says four. Ask it yes-or-no on the same text and it says no. A new paper opens the hood and finds the judgment itself is stable — what wobbles is a tiny piece of machinery near the output that translates an abstract verdict into whichever token the prompt demanded. If they're right, a lot of what we call evaluator unreliability is actually a formatting artifact. Key Takeaways: - Why LLM judges produce inconsistent scores across prompt formats — and why the inconsistency lives in output routing, not evaluation quality - The 'transplant' experiment: copying activations from a rating prompt into a yes-no prompt flips the model's answer over 99% of the time on some models - Evidence that judgment is encoded along a single direction in activation space — a 'compass needle' that transfers across grammar, entailment, similarity, and preference tasks - A practical alternative to prompted scoring: read the judgment axis directly from mid-layer activations and bypass the noisy formatter - Where the clean modularity story breaks down — Gemma-3 at 12B entangles judgment with world knowledge in a way no other tested model does - Honest limits of the result: small per-cell sample sizes, probe design that partly presupposes a 1D encoding, and a universality claim the authors deliberately don't make 00:00 - The format-inconsistency puzzle: Why the same model gives different verdicts under different output formats, and why prior work stopped at behavioral observation. 03:12 - Latent Evaluator and Task Formatters: The paper's central decomposition: a shared judgment core in the middle layers, and format-specific translators at the output. 04:07 - Format Transfer Injection: Transplanting activations from a rating run into a yes-no run flips the output — judgment is portable, formatting is local. 09:36 - Judgment as a single direction: Evidence that the model's verdict reduces to a scalar along one axis in activation space, transferable across tasks and steerable. 12:48 - How they found the circuit: Position-aware Edge Attribution Patching, contrastive circuit tracing, and independent SAE corroboration of the shared core. 16:00 - Where the modularity story breaks: Gemma-3 at 12B entangles judgment with world knowledge while its 27B sibling doesn't — modularity is architecture- and scale-dependent. 19:12 - Pressure points and limitations: Small sample sizes behind the headline numbers, probe-design assumptions, and the absence of head-to-head comparison with an independent attribution method. 22:24 - What this means for LLM-as-a-judge: Reading the internal judgment axis directly can beat the model's prompted output, and benchmark format differences may measure formatter geometry rather than evaluation quality. Recommended Reading: - Towards Automated Circuit Discovery for Mechanistic Interpretability: Introduces the circuit-discovery methodology that PEAP builds on, giving listeners the foundation for how researchers identify causally important sub-networks like the Latent Evaluator. (https://arxiv.org/abs/2304.14997) - Attribution Patching: Activation Patching At Industrial Scale: The gradient-based attribution method that PEAP extends to edges; useful for understanding the one-forward-one-backward shortcut the episode describes for scoring millions of connections. (https://www.neelnanda.io/mechanistic-interpretability/attribution-patching) - Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena: The canonical reference for the LLM-as-a-judge paradigm whose format-inconsistency problems this episode's paper goes inside the model to explain. (https://arxiv.org/abs/2306.05685) - Do Llamas Work in English? On the Latent Language of Multilingual Transformers: A parallel 'compute abstractly, then translate' finding for multilingual models, supporting the episode's broader claim that judgment-vs-formatting is part of a recurring pattern in interpretability. (https://arxiv.org/abs/2402.10588)

Why LLM Judges Flip Their Verdicts When You Change the Question Format Source: https://arxiv.org/abs/2605.16023 Paper was published on May 15, 2026 This episode was AI-generated on May 19, 2026. The script was written by an AI language model and the host voices were synthesized by Eleven Labs. The producer is not affiliated with Anthropic or Eleven Labs. Ask a language model to rate text from 1 to 5 and it says four. Ask it yes-or-no on the same text and it says no. A new paper opens the hood and finds the judgment itself is stable — what wobbles is a tiny piece of machinery near the output that translates an abstract verdict into whichever token the prompt demanded. If they're right, a lot of what we call evaluator unreliability is actually a formatting artifact. Key Takeaways: - Why LLM judges produce inconsistent scores across prompt formats — and why the inconsistency lives in output routing, not evaluation quality - The 'transplant' experiment: copying activations from a rating prompt into a yes-no prompt flips the model's answer over 99% of the time on some models - Evidence that judgment is encoded along a single direction in activation space — a 'compass needle' that transfers across grammar, entailment, similarity, and preference tasks - A practical alternative to prompted scoring: read the judgment axis directly from mid-layer activations and bypass the noisy formatter - Where the clean modularity story breaks down — Gemma-3 at 12B entangles judgment with world knowledge in a way no other tested model does - Honest limits of the result: small per-cell sample sizes, probe design that partly presupposes a 1D encoding, and a universality claim the authors deliberately don't make 00:00 - The format-inconsistency puzzle: Why the same model gives different verdicts under different output formats, and why prior work stopped at behavioral observation. 03:12 - Latent Evaluator and Task Formatters: The paper's central decomposition: a shared judgment core in the middle layers, and format-specific translators at the output. 04:07 - Format Transfer Injection: Transplanting activations from a rating run into a yes-no run flips the output — judgment is portable, formatting is local. 09:36 - Judgment as a single direction: Evidence that the model's verdict reduces to a scalar along one axis in activation space, transferable across tasks and steerable. 12:48 - How they found the circuit: Position-aware Edge Attribution Patching, contrastive circuit tracing, and independent SAE corroboration of the shared core. 16:00 - Where the modularity story breaks: Gemma-3 at 12B entangles judgment with world knowledge while its 27B sibling doesn't — modularity is architecture- and scale-dependent. 19:12 - Pressure points and limitations: Small sample sizes behind the headline numbers, probe-design assumptions, and the absence of head-to-head comparison with an independent attribution method. 22:24 - What this means for LLM-as-a-judge: Reading the internal judgment axis directly can beat the model's prompted output, and benchmark format differences may measure formatter geometry rather than evaluation quality. Recommended Reading: - Towards Automated Circuit Discovery for Mechanistic Interpretability: Introduces the circuit-discovery methodology that PEAP builds on, giving listeners the foundation for how researchers identify causally important sub-networks like the Latent Evaluator. (https://arxiv.org/abs/2304.14997) - Attribution Patching: Activation Patching At Industrial Scale: The gradient-based attribution method that PEAP extends to edges; useful for understanding the one-forward-one-backward shortcut the episode describes for scoring millions of connections…

NOW PLAYING

Why LLM Judges Flip Their Verdicts When You Change the Question Format

0:00 25:38

No transcript for this episode yet

We transcribe on demand. Request one and we'll notify you when it's ready — usually under 10 minutes.

MG Show MG Show The MG Show, hosted by Jeffrey Pedersen and Shannon Townsend, is a leading alternative media platform dedicated to uncovering the truth behind today’s most pressing political issues. Launched in 2019, the show has grown exponentially, offering unfiltered insights, comprehensive research, and real-time analysis. With a commitment to independent journalism and factual integrity, the MG Show empowers its audience with knowledge and encourages active participation in the political discourse. Ask A Spaceman Archives - 365 Days of Astronomy Ask A Spaceman Archives - 365 Days of Astronomy Podcasting Astronomy Every Day of the Year French Your Way Jessica: Native French teacher founder of French Your Way Boost your French listening skills and test your comprehension with this one of a kind series of podcasts. Get the chance to listen to a real conversation between native speakers talking at normal speed AND customise your learning experience through carefully designed sets of questions (2 levels of difficulty) available for download at www.frenchvoicespodcast.com. All interviews also come with the transcript. French teacher Jessica interviews native speakers of French from around the world who share a bit of their life and passion. Where else would you meet in one same place a French yoga teacher based in Melbourne, a soap manufacturer from Provence, or a couple cycling around the world? The Small Business Startup School – Business Notes | Financial Literacy | Retail Psychology – For Professionals & Entrepreneurs The Small Business Startup School Inc. Starting or buying a small business? While personal circumstances may vary, business patterns remain timeless. On The Small Business Startup School, we explore strategies, insights, and practical solutions to help entrepreneurs confidently navigate their journey.Hosted by Ola Williams—a retail entrepreneur, fintech founder, and financial coach with over two decades of experience—this podcast marries financial awareness and retail psychology with optimism to deliver actionable takeaways.Join us to learn, grow, and connect as we uncover the keys to business success.Let’s continue to learn together and be encouraged to keep on connecting!

Frequently Asked Questions

How long is this episode of AI Papers: A Deep Dive?

This episode is 25 minutes long.

When was this AI Papers: A Deep Dive episode published?

This episode was published on May 19, 2026.

What is this episode about?

Why LLM Judges Flip Their Verdicts When You Change the Question Format Source: https://arxiv.org/abs/2605.16023 Paper was published on May 15, 2026 This episode was AI-generated on May 19, 2026. The script was written by an AI language model and...

Can I download this AI Papers: A Deep Dive episode?

Yes, you can download this episode by clicking the download button on the episode player, or subscribe to the podcast in your preferred podcast app for automatic downloads.
URL copied to clipboard!