When there's no answer key to grade an AI against, what do people use instead — and can you trust it?
Why are ground truth labels missing from unlabeled domain evaluations?
This explores why some domains have no answer key to grade against, and what people use in place of one. The corpus has no note that asks this directly, so this is stitched together from neighboring work.
This explores why some domains have no answer key to grade against, and what people substitute when it's missing. The corpus has no note that asks this head-on, so what follows is stitched together from neighbors. The short version is that a ground truth label exists only when someone can produce it cheaply and defensibly. A math answer or a medical fact with a known path in a knowledge graph has a right answer. Open-ended judgment calls in an unlabeled domain don't, and every workaround in the corpus is a way of manufacturing a stand-in.
The first stand-in is human judgment, but the corpus suggests a human label is not the same thing as ground truth. Do all annotation responses measure the same underlying thing? finds that annotations mix three things. Some are genuine preferences. Some are non-attitudes, where the person answers without holding a view. Some are constructed preferences, invented on the spot from how the question was framed. You can only tell them apart by checking whether answers stay consistent across measurement conditions. So even where labels exist, some of what looks like truth is noise. What rescues human judgment is scale and question quality: Can crowdsourced votes reliably rank language models? shows that 240K+ votes on diverse, discriminating questions agree with expert raters. Agreement across many judges can work as a stand-in key, though it is still a proxy.
The second stand-in is a machine-made label. Can self-supervised process rewards replace human annotation? reaches o3-mini-level results by dynamically weighting pseudo-labels instead of paying humans to annotate each step. The caveat is the one that matters for your question: it says generalization to fuzzy-outcome domains remains unproven. Those are the domains where labels are missing in the first place. Can smaller models outperform their LLM teachers with enough data? points the same way. Walmart's small student models beat the LLM teacher whose labels they were trained on. That only makes sense if the teacher's labels were a smoothed approximation rather than truth.
The third route is to build the labels from structure. Can knowledge graphs teach models deep domain expertise? turns medical knowledge graph paths into 24,000 reasoning tasks. Each task has a checkable answer built in, and a 32B model trained on them reaches state-of-the-art results across 15 medical domains. Domains that already have structured knowledge get labels almost for free, and domains without it don't.
Why this matters: How do domain training techniques actually reshape model behavior? reports that visible gains from domain adaptation often come with hidden losses in reasoning faithfulness, capability transfer and format flexibility. A single accuracy label can't see those losses. So missing ground truth is more than a cost problem. Without it, the damage from adapting a model to a domain can go unnoticed.
Sources 6 notes
Behavioral science reveals that annotations contain genuine preferences, non-attitudes, and constructed preferences—distinguishable by consistency across measurement conditions. Treating them uniformly contaminates reward model training and downstream alignment.
Chatbot Arena's 240K+ crowdsourced preference votes produce credible model rankings because the underlying questions are diverse and discriminating, and crowd judgments correlate with expert raters—validating human preference as a scalable evaluation signal.
MetaStone-S1's SPRM achieves o3-mini-level results using dynamic weighting of pseudo-labels instead of human-annotated steps. This eliminates the annotation bottleneck for process supervision, though generalization to fuzzy-outcome domains remains unproven.
Walmart's student cross-encoders outperformed their LLM teachers when trained on sufficiently large augmented datasets of teacher-labeled queries. The student's broader input distribution exposure, smoothed by teacher predictions, enabled better generalization than the teacher achieved.
Fine-tuning a 32B model on 24,000 reasoning tasks derived from medical knowledge graph paths produces state-of-the-art performance across 15 medical domains, demonstrating that structured knowledge composition matters more than scale.
Show all 6 sources
Research shows every adaptation method—from parameter-efficient tuning to knowledge graph curricula—has optimal conditions tied to specific domains. The key finding: visible benefits like performance gains often come with hidden degradation in reasoning faithfulness, capability transfer, and format flexibility.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Bottom-up Domain-specific Superintelligence: A Reliable Knowledge Graph is What We Need
- Measuring Human Preferences in RLHF is a Social Science Problem
- Replacing Judges with Juries: Evaluating LLM Generations with a Panel of Diverse Models
- Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond)
- Faith and Fate: Limits of Transformers on Compositionality
- Chatbot Arena: An Open Platform for Evaluating LLMs by Human Preference
- Test-Time Scaling with Reflective Generative Model
- Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena