Can verification accuracy scale without training models?
Does correctness-checking improve as a separate capability when you invest more inference compute, using techniques like repeated evaluation and criteria decomposition instead of training new verifiers?
Generation has three well-worn scaling axes — pre-training, post-training, test-time compute — and LLM-as-a-Verifier argues verification is a fourth that has not been scaled the same way. The claim is that determining whether a solution is correct is not a byproduct of a good generator but a separable capability with its own scaling dimensions: score granularity, repeated evaluation, and criteria decomposition. Each can be dialed up at inference, without additional training, to buy more verification accuracy.
This reframes a bottleneck. Since What is the actual reusable unit of reasoning data?, the verifier is the load-bearing component of reasoning training — yet current verifiers are weak: LM judges collapse into coarse discrete scores that tie, and learned reward models are stuck inside their training distribution and fail to generalize across domains. If verification is a scaling axis, then those failures are not fixed properties but under-scaled ones, and the way to a better verifier is to spend more inference compute on it rather than to train a bespoke reward model.
The three dimensions decompose the gains cleanly: granularity gives better positive/negative separation, repeated evaluation reduces variance, and criteria decomposition reduces complexity per judgment. This connects to the panel-of-judges result — Can a panel of smaller judges outperform one large judge? is a special case of scaling repeated evaluation across evaluators — but LLM-as-a-Verifier generalizes it into a training-free framework that provides dense agentic feedback where discrete judges provide none.
Inquiring lines that read this note 10
This note is a source for these research framings, grouped by the broader line of inquiry each explores. Scan the bold lines of inquiry; follow any specific question forward.
How does evaluation scope and dimensionality affect what we measure? How effectively can language models perform reasoning, especially combined with symbolic methods? What makes imperfect LLM judges safe for optimization? Can harness architecture and protocols provide agent reliability without model scaling? Can validator consensus certify semantic correctness beyond agreement? How effective are honeytokens and decoys against different security threats? What capability trade-offs arise from domain specialization through fine-tuning? Is reasoning capability latent in base models or created by post-training? How does the generation-verification gap limit what we can measure about AI reasoning? Why do locally safe actions create system-level safety gaps?Related concepts in this collection 4
This note in its neighbourhood — explore the map, then jump to a related concept in the list below.
Click a node to walk · click center to open · click Open in graph to see this note in the full knowledge graph
-
What is the actual reusable unit of reasoning data?
Does post-training reasoning transfer as prompt-response pairs, or as something more complex? Understanding what artifact actually drives gains matters for reproducibility and attribution.
grounds: the verifier is the reusable core, so scaling it scales reasoning
-
Can a panel of smaller judges outperform one large judge?
Does aggregating votes from multiple smaller language models across different families produce better evaluations than relying on a single large model like GPT-4? This matters because evaluation cost and bias directly affect the reliability of AI-generated content assessment.
extends: panel judging is one instance of scaling repeated evaluation
-
Can generative reasoning beat discriminative models with less training data?
Do process reward models that generate reasoning before judging achieve better performance than traditional discriminative approaches when trained on dramatically smaller datasets? This tests whether generative verification can scale more efficiently.
parallel: another route to denser, more general verification signal
-
How much do deterministic guardrails actually cost to run?
The paper claims mechanical checks around LLM judges are orders of magnitude cheaper than running the judge itself. But what specific costs were measured, and does this account for false positives?
contrasts: this axis buys verification accuracy with more inference, while checks that call no model sit off it and are cited as far cheaper than the judge they contain; no figures in that excerpt
Related papers in this collection 8
Papers most semantically related to this note, ranked by cosine similarity in the embedding space.
- LLM-as-a-Verifier: A General-Purpose Verification Framework
- Mind the Gap: Examining the Self-Improvement Capabilities of Large Language Models
- A Primer in Post-Training Reasoning Data: What We Know About How It Works
- SFT Memorizes, RL Generalizes: A Comparative Study of Foundation Model Post-training
- Local Coherence or Global Validity? Investigating RLVR Traces in Math Domains
- Reinforcing General Reasoning without Verifiers
- A Survey on Test-Time Scaling in Large Language Models: What, How, Where, and How Well?
- Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters
Original note title
verification is a distinct scaling axis alongside pre-training, post-training, and test-time compute