If several AI checkers each dig up their own evidence, does their agreement really prove anything, and what does that cost?
Can validators gather evidence independently without raising disagreement costs?
This explores whether a panel of AI validators can each go find their own evidence, so that their agreement counts as real confirmation, without that independence causing more splits, stalemates, or extra expense.
This explores whether a panel of AI validators can each go find their own evidence, so that their agreement counts as real confirmation, without that independence causing more splits, stalemates, or extra expense. The corpus never measures "disagreement cost" directly. It does show why independence is worth wanting, and where the costs of getting it would show up.
A quorum's vote only counts as independent confirmation if the validators don't share a cause of error. The Honest Quorum work lists eight things validators can share, from model weights to provider infrastructure, so one fault can hit many validators at once Can a quorum of validators really provide independent judgment?. Swapping in different model families closes only two of those eight channels. Prompts, retrieval sources and infrastructure still tie the validators together Does model diversity actually reduce validator agreement failures?. So independent evidence gathering is the right lever, because retrieval sources are one of the channels model diversity leaves open.
The same framework says where the cost lands. Agreement is guaranteed deterministically. Semantic correctness and liveness, meaning the group actually reaching a decision, hold only statistically, and they depend on validator behavior the protocol can't enforce Can validator consensus guarantee both agreement and semantic correctness?. My inference, not the paper's claim, is that validators with truly separate evidence will disagree more often, and that pressure lands on the weaker liveness guarantee.
The evidence-gathering side looks promising, with one warning. An agent judge that collects evidence dynamically showed 0.27% judge shift, against 31% for an LLM judge on complex tasks. Its shared memory module cascaded errors, though, and that is a shared channel recreating a correlated fault in miniature, so independence needs error isolation Can agents evaluate AI outputs more reliably than language models?. Looking outside the validator's own draft matters too. Models over-trust answers they generated themselves, and comparing against broader alternatives breaks that loop Why do models trust their own generated answers?.
Extra checking can be kept cheap. Verifiers can run alongside generation with near-zero latency on correct runs, stepping in only when they detect a violation Can verifiers monitor reasoning without slowing generation down?. That is a design where disagreement costs something only when there is something to disagree about. Verification accuracy also scales at inference time through repeated evaluation and criteria decomposition, with no retraining Can verification accuracy scale without training models?. Confidence signals could decide when a validator needs more evidence at all, either by stopping early on step-level confidence Does step-level confidence outperform global averaging for trace filtering? or by retrieving only when uncertain Can simple uncertainty estimates beat complex adaptive retrieval?. Those results come from single-model settings, so treat them as design hints for a quorum, not proof.
The answer is plausibly yes, but untested. The corpus has the pieces: why independence matters, evidence-gathering that works, and ways to keep the overhead low. It has no experiment where quorum validators retrieve independently while disagreement rate and liveness are tracked. The closest proposal is to vary validators one shared channel at a time and measure how their errors correlate Does model diversity actually reduce validator agreement failures?.
Sources 9 notes
Honest Quorum's threshold theorems split into two kinds of guarantee: agreement rests on protocol assumptions alone, while semantic validity and liveness depend on statistical bounds over validator behavior that the protocol cannot enforce.
The Honest Quorum paper identifies eight things agentic validators may share—from model weights to provider infrastructure—allowing one fault to affect multiple validators simultaneously. Quorum voting then counts agreement as independent confirmation when it may simply reflect a common cause.
While different model families address two of eight shared fault channels, correlated epistemic errors likely persist through prompts, retrieval sources, and provider infrastructure. Measuring error correlation across validators differing on one channel at a time could quantify the effect.
Eight-module agentic evaluation achieved 0.27% judge shift versus 31% for LLM-as-a-Judge on complex tasks. However, the memory module cascaded errors, revealing that agentic systems need error isolation mechanisms to maintain gains.
LLMs exhibit structural bias toward validating their own outputs because high-probability generated answers feel more correct during evaluation. Comparing answers against broader alternatives breaks this self-agreement loop.
Show all 9 sources
Decoupling verification from generation lets verifiers run alongside a single trace, forking to extract verifiable state and intervening only on violations. On correct runs the latency penalty is near-zero; interwhen matches or beats CoT across benchmarks at similar token budgets.
Research shows verification accuracy improves independently via score granularity, repeated evaluation, and criteria decomposition—all deployable at inference without retraining. This reframes weak verifiers as under-scaled rather than fundamentally limited.
Local step-level confidence catches reasoning breakdowns that global averaging masks and enables early stopping before traces complete. This approach achieves comparable accuracy gains to naive majority voting with far fewer generated traces, proving trace quality matters more than quantity.
Calibrated token-probability uncertainty consistently beats multi-call adaptive retrieval on single-hop tasks and matches performance on multi-hop, using a fraction of the LM and retriever calls. The model's self-knowledge proves more reliable than external heuristics for deciding when to retrieve.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Local Coherence or Global Validity? Investigating RLVR Traces in Math Domains
- When AIs Judge AIs: The Rise of Agent-as-a-Judge Evaluation for LLMs
- What Characterizes Effective Reasoning? Revisiting Length, Review, and Structure of CoT
- A Survey on Test-Time Scaling in Large Language Models: What, How, Where, and How Well?
- The Honest Quorum Problem: Epistemic Byzantine Fault Tolerance for Agentic Infrastructure
- Deep Research: A Systematic Survey
- LLM-as-a-Verifier: A General-Purpose Verification Framework
- Deep Think with Confidence