- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
This paper studies multi‑LLM evaluation: high consensus among models does not guarantee correct judgements, as models can share common errors. It analyzes the capability boundary of relying only on internal model information, and derives theoretical findings including consensus reversal, Bayesian error lower bound and unidentifiability of semantic orientation. It proposes GAVEL: train and freeze an aggregation policy using ground‑truth labels, then validate risk via independent sequential auditing, and provide the optimal sampling scheme for stratified auditing. Experiments are conducted on synthetic, semi‑synthetic data and four real‑world benchmarks. Performance gains are mixed on real‑world tasks, and labeling‑cost savings are mainly observed under heterogeneous‑risk scenarios.
Critical strengths
Overall, this paper is clear and rigorous, and presents a closed‑loop investigation. It points out that LLMs can make the same mistakes even when they agree with each other. The paper explains this problem with theory and experiments, and suggests using independent ground-truth labels to check how reliable the system is. It covers problem finding, cause analysis and solution design, and also states when this method works, so it is helpful for us to understand and improve multi‑model evaluation.
Critical weaknesses
While the theoretical analysis in this paper is thorough and rigorous, its incremental contributions compared to prior work are not sufficiently established. The early results on information limits and semantic orientation draw heavily on existing statistical ideas. The more novel part lies in the specific solution for optimal audit allocation. However, the paper’s own control experiments show that the simpler Local allocation achieves most of the gains and performs closely to GAVEL.
As a result, the paper does not fully demonstrate how substantial an advance these theoretical results and methods represent over existing approaches. This shapes my assessment of its novelty and overall significance.
Questions for authors
-
Could the authors clarify what the semantic‑orientation result adds beyond the sign ambiguity already discussed in unlabeled ensemble learning, such as Jaffe et al. (2015: Estimating the Accuracies of Multiple Classifiers Without Labeled Data, AISTATS)?
-
Could the authors clarify which parts of the auditing theory are new contributions and which follow from existing anytime-valid risk-control methods?
-
How do the theoretical results on the limitations of peer‑only evaluation inform the specific design choices in GAVEL?
-
Given the competitive performance of Local allocation, under what conditions does exact Chernoff allocation provide meaningful additional benefits? A small extension of the existing synthetic experiments, varying the number of strata or their prevalences, would help clarify this.
7 comments
Response to yAf1 — Part 2/6: Q1–Q3: Auditing contributions and theory-guided design
What the auditing theory adds. Ville's inequality, predictable betting, and sequential change-of-measure lower bounds are established. The problem-specific step is solving the allocation game for a composite aggregate-risk null:
\sum_k w_k\\,\mathrm{kl}(r_k,q_k).$$ The closest **unsafe profile** depends on the allocation, and it can move risk between strata. A variance rule sees only the local geometry. `Proposition 2 — One-dimensional Chernoff characterization` solves the full game for interior safe risks through a single equal-KL equation. It gives a unique positive optimal allocation and the characteristic rate $C^{\ast}$. The decomposition $C^{\ast}\geq C_\pi\geq\mathrm{kl}(\pi^\top r,\alpha)$ distinguishes the gain from allocating labels from the gain from retaining stratum identities. The fixed-allocation likelihood-ratio calculation connects the same unsafe profile to the log-optimal evidence factor. `Theorem 4` constructs anytime-valid evidence and, under the stated convergent-design assumptions, attains the characteristic almost-sure evidence and stopping-time rates. `Theorem 5` gives the expected-label lower bound with the same constant. **The adaptive result is an almost-sure rate matched to an expected-time lower bound. Expected-time attainment is established only for the scalar benchmark.** The derivation combines convex saddle-point/KKT geometry with predictable evidence and change of measure. The new `Proposition 3` and `Theorem 6`, detailed in Q4, address why this exact solution can matter despite Local's strong performance. The former supplies an exact loss identity and computable bounds. The latter identifies when Local's rate collapses and gives a sharp separation rate. A nearly error-free stratum can still require substantial auditing because an unsafe alternative can place enough errors there. This provides a concrete direction for future work: estimate whether such alternatives remain plausible, and choose audit initialization or allocation accordingly. **Revision.** Sections 2–4 and Appendix A.1 now distinguish the contributions from the results of Kim, Goel, Li, Jaffe, and Shi, and clarify the complete-state and orientation assumptions. Section 6 separates the explicit aggregate-risk solution from the classical validity tools and states the scope of the rate guarantees. It also contains the new `Proposition 3` and `Theorem 6`; their full proofs are in Appendices D.10 and D.11. ### Q3 — How the limits determine GAVEL's design > **Q3. “How do the theoretical results on the limitations of peer‑only evaluation inform the specific design choices in GAVEL?”** Each stage addresses a specific obstruction. The results justify verified orientation, the separation of policy fitting from auditing, and the allocation of fresh evidence. **`Theorem 1 — Exact agreement reversal`** In the stated shared-error family model, crossing an explicit dependence threshold makes the less accurate family score higher under agreement, determinant mutual information, and Shannon mutual information. GAVEL therefore fits against verified truth rather than peer agreement. The synthetic study checks the phase boundary; the semi-synthetic MMLU intervention reverses the ranking at 0.85 on its grid. In the **new real-panel experiment** added for this revision in response to Reviewer `Lrwg`'s Q1, the natural same-family panel has `3.92 percentage points more agreement but 3.84 points less accuracy` than the cross-family panel, without injected errors. **`Theorem 2 — Complete-state Bayes barrier` and `Proposition 1 — Exact shared-error floor`** A peer-only transcript cannot outperform the Bayes decision based on the complete private state. In the shared-flip model, the exact ensemble risk is $R_n^{\ast}=\beta+(1-2\beta)\delta_n$: increasing the number of judges removes the idiosyncratic term but leaves the common-error floor. These results allow interaction to improve access to existing information and computation. They identify when adding judges or discussion still leaves factual uncertainty. GAVEL acquires verified evidence rather than treating a larger panel as a certificate. The shared-flip experiment checks the exact residual floor and the decay of the idiosyncratic term.Response to yAf1 — Part 3/6: Q3: Verified fitting and the fit–freeze–audit pipeline
Theorem 3 — Exact semantic-orientation minimax limit
Unlabeled peer observations cannot select between truth-swapped worlds, even with dependence, interaction, or peer-observable selective acceptance. A truth-swap-closed class has minimax risk .
Verified fitting labels orient the builder before it is frozen. Unlabeled data can still estimate useful report structure; they do not generally determine its alignment with truth. The anchor study in Appendix E.2, based on Appendix B.7, covers four signal strengths, nine label budgets, and 100,000 repetitions per direction. It checks how external labels resolve the ambiguity, while preserving the distinction between orientation and residual deployment-time uncertainty.
Proposition 5 — Oracle correlated-error certificate
The population certificate for a weighted report rule depends on both its alignment with truth and the covariance of its truth-aligned errors. Unlabeled reports identify ; they do not generally identify .
The regularized builder combines an unlabeled estimate of with verified fitting estimates of , then freezes its coefficients and threshold. The moment identities and Cantelli bound motivate this choice. Experiments compare nine builders at three synthetic fitting budgets and ten on the real panel. The four cached tasks retain mixed outcomes: +1.07, +1.14, 0.00, and -1.85 percentage points versus matched fitted single evaluators on MMLU, HellaSwag, ARC-Challenge, and WinoGrande. Ridge is one eligible builder; the auditor can also assess a frozen EBCC predictor or single judge.
Proposition 2 — One-dimensional Chernoff characterization
For interior safe risks, an equal-KL unsafe profile determines the unique positive allocation that maximizes the worst-case information rate.
Fresh audit labels are directed toward the unsafe alternatives that are hardest to exclude. This determines the exact target used by GAVEL's adaptive allocation. The 192-setting known-risk grid and the matched-evidence allocation experiments test that consequence while holding the evidence construction fixed. The new Proposition 3 quantifies the loss from using Local, and Theorem 6 identifies its boundary failure; Q4 gives their conditions and the complete new mechanism experiment.
Theorem 4 — Anytime validity and characteristic evidence rate and Theorem 5 — Information-theoretic lower bound
After the policy and audit population are frozen, predictable sampling and valid evidence factors permit adaptive allocation and optional stopping while controlling false certification. Under the stated convergence assumptions, the evidence attains the characteristic rate; every qualifying valid procedure faces the corresponding expected-label lower bound.
This yields fit–freeze–audit separation: fitting labels choose the policy, fresh labels assess it, and stopping requires the evidence threshold. The real-panel replay tests this pipeline at four preset targets. At the primary target, GAVEL gains 2.71 accuracy points over the development-selected native single judge and uses 34.5% fewer audit labels; both certify on all replay paths. Independent Bernoulli null checks assess the implementation. These savings concern the full fitted policy and audit, not Exact allocation alone.
The counterintuitive implication is that agreement, confidence, and low observed error can each be insufficient for a different reason. Agreement may reward shared mistakes; confidence may not resolve semantic orientation; low variance may hide an unsafe alternative from Local. The design uses external labels at the stages where they resolve the relevant uncertainty. Future extensions can study richer sources of external evidence, noisy anchors, and estimation-aware allocation while retaining a separate audit of the final policy.
Revision. Sections 4–6 now state these links between the limits, builder, and auditor explicitly. Section 7 reports the corresponding mechanism and empirical checks. Natural reversal, real-panel aggregation, and the full audit replay appear in Appendices E.12, E.13, and E.14, respectively. Appendix E.10 retains the complete initialization and audit-length comparison; Appendix E.15 adds the Exact/Local mechanism checks.
Response to yAf1 — Part 1/6: Q1–Q2: Scientific question and information limits
Response to Reviewer
yAf1Thank you for reviewing our paper and for your comments. Below, we respond to each of your concerns point by point.
Q1–Q2 — Scope of the information limits and the auditing contribution
Our scientific question is which evidence can establish factual reliability under shared evaluator errors, and how many verified labels are needed to certify it. A panel can become more consistent without becoming more correct. This matters when models are used to judge or supervise other models: more peer observations can estimate a misleading agreement ranking more precisely, while verified labels remain costly. The paper separates three questions that require different answers: whether the complete peer state contains the fact, whether unlabeled observations orient that state toward truth, and how fresh labels certify the risk of a chosen policy.
The related work already establishes important parts of this motivation. Kim et al. document correlated LLM errors across a large model collection and examine downstream effects. Goel et al. show how model similarity affects judging and weak-to-strong supervision. Li et al. explicitly model worker dependence through EBCC for label aggregation. Our additional target is the relationship between observable peer evidence, factual identifiability, and the verified-label cost of a risk certificate. The exact reversal boundary, decision-level orientation limit, and aggregate-risk allocation solution address these different parts of that target.
The comparison with Shi et al., Fundamental Limits of Game-Theoretic LLM Alignment: Smith Consistency and Preference Matching, also clarifies the scientific gap. They characterize payoff conditions for desirable preference-based equilibrium properties and establish a preference-matching impossibility within their specified smooth ratio-based payoff class. Our question concerns factual correctness when peer observations themselves may share an error. Preference consistency does not supply a factual orientation assumption. In particular, if a game uses only peer observables that remain unchanged under a truth swap, its output is still subject to our orientation limit, even if it satisfies a preference-consistency criterion. This applies our information model to game outputs built from peer observations. Together, the two perspectives motivate methods that state both what their interaction objective preserves and what external evidence grounds its factual interpretation.
What the orientation result adds. Jaffe et al. explicitly identify sign ambiguity. Their estimation results use conditional independence, and their setup fixes orientation by assuming that a majority of classifiers have balanced accuracy above one half.
Theorem 3 — Exact semantic-orientation minimax limitstates a decision-level identity for the same learner in two observationally identical worlds:RiskP(h)+RiskP↔(h)=1.
It allows dependent peer states, arbitrary peer-only interaction, and arbitrarily many unlabeled training samples. The same identity holds conditional on a peer-observable acceptance event with common positive coverage. Consequently, the minimax risk on any nonempty truth-swap-closed class is exactly 1/2. Confidence-based abstention cannot orient the accepted predictions. An assumption that a designated evaluator is reliably better than chance can break the symmetry; verified labels provide another source of orientation. The contribution is the scope and decision consequence of the ambiguity, including selective prediction.
Theorem 2compares a transcript with the evaluators' complete private state, not only their initial answers. Interaction may improve computation or reveal information already present in that state. The bound identifies the residual uncertainty that such interaction cannot remove without additional evidence. Its proof builds on classical comparison-of-experiments and data-processing arguments.