acceptodds
Under review as a conference paper at ICLR 2027

Forming Pairwise Coherent Event Probabilities from Language Models

Abstract

Large language models (LLMs) are increasingly asked for numerical probabilities of uncertain events. Individually, these estimates are often reasonable; collectively, they are frequently incoherent: no joint distribution realizes all of the reported marginal and pairwise-intersection probabilities at once. Existing remedies (prompting, calibration, fine-tuning, learned representations) reduce incoherence but do not eliminate it, while trivially coherent substitutes such as assuming independence discard the information in the judgments. We present, to our knowledge, the first model-agnostic post-processing method for LLM probabilities that guarantees exact pairwise coherence while certifiably staying closest to the original judgments. Starting from the classical fact that a co-occurrence matrix is pairwise coherent if and only if it is a probability mixture of binary outer products (i.e., lies in the full correlation polytope), we cast coherence enforcement as a projection onto this polytope up to a global rescaling, and solve it by column generation: an exactly solved 0-1 quadratic program selects new “possible worlds” and an active-set nonnegative least squares solver refits their weights. For every input and every stopping tolerance, the output is coherent by construction and comes with an explicit joint distribution certifying it; the stopping rule also certifies global optimality of the projection up to that tolerance. Deciding coherence is NP-complete, but hardness affects only worst-case running time, never correctness. On event families with known ground truth, four LLMs are substantially incoherent. Local Fr\'echet repairs and calibration-style rescaling (even when fit to the ground truth) reduce but never eliminate this incoherence, whereas our projection eliminates it, lowers root mean squared error in 35 of 40 model–premise settings, and attains the lowest error of all compared methods in 29.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
coherence, probability, calibrarion, NP-complete, LLM, projection

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_MXXeOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how to post-process LLM-generated marginal and pairwise-intersection probabilities so that they are jointly realizable by a single probability distribution. The method represents coherent probabilities as mixtures of binary outer products and solves a globally rescaled projection problem through column generation, combining exact binary quadratic optimization with nonnegative least squares. The output includes an explicit joint-distribution certificate, together with a tolerance-dependent optimality guarantee for the stated objective. Experiments evaluate four LLMs on coin and dice problems with known ground truth, reporting reduced RMSE in 35 of 40 model–premise settings.

Critical strengths
  1. The method provides global realizability of the supplied marginal and pairwise probabilities, an explicit joint-distribution certificate, and a certified optimality gap for the stated projection objective with global rescaling. These guarantees provide a principled advantage over local consistency repairs and require neither model retraining nor ground-truth labels.

  2. The controlled experiments provide evidence that coherence enforcement can improve probability accuracy. RMSE decreases in 35 of 40 settings, with the average falling from 0.307 to 0.185. Using problems with exact ground truth allows the evaluation to distinguish improvements in coherence from improvements in accuracy.

Critical weaknesses
  1. The practical benefit for recent frontier models remains unclear. The evaluation covers Gemma 2, Qwen 2.5, GLM-4.5, and GPT-OSS-120B, but does not establish how much incoherence remains in more recent frontier models or whether the proposed correction improves their probability estimates. This matters because projection increases RMSE in all five GPT-OSS coin-flip settings, where the raw estimates are already comparatively accurate. The current evidence therefore leaves the method’s practical benefit with stronger initial estimates insufficiently established.

  2. The empirical comparison lacks recent LLM-specific probability-repair methods. Relevant alternatives include Joint-Coherent Decoding (JCD) [1], which projects LLM probabilities onto coherent polytopes defined by small logical cliques, and ArbitrageForecaster [2], which repairs forecasts through arbitrage-based consistency checks. The paper discusses ArbitrageForecaster but does not evaluate it, and does not discuss or compare against JCD. These methods enforce different consistency constraints from the full pairwise coherence considered here. A direct empirical comparison with these methods would be helpful, or clarification of why such a comparison is not applicable.

References

[1] Arbitrage-Free Forecasts from Language Models via Coherence Projection. Forecast@ICML 2026. https://openreview.net/pdf?id=Tqos7VqQhH

[2] Paleka et al. Consistency Checks for Language Model Forecasters. ICLR 2025. https://arxiv.org/abs/2412.18544

Questions for authors
  1. How do the degree of incoherence and the accuracy gains from projection change when using more recent frontier models?

  2. How does the proposed method compare empirically with JCD and ArbitrageForecaster under comparable inputs and inference budgets, or what differences prevent a meaningful comparison?

Reviewer_SG38OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper presents a post-processing method for LLM-predicted probabilities that guarantees exact pairwise coherence while staying as close as possible to the original probabilities. The approach projects onto the full correlation polytope (i.e. the convex hull of all binary outer products aaTaa^T with a∈{0,1}Na \in \lbrace 0,1\rbrace^N) up to a global rescaling. This projection is formulated as a nonnegative least squares problem over 2N2^N atoms and solved by column generation. The output is coherent for any stopping tolerance, which only bounds the suboptimality of the solution (i.e., its distance to the LLM-generated probabilities). On coin and dice tasks with four LLMs, the method eliminates incoherence and lowers RMSE in 35 of 40 settings.

Critical strengths
  • S1: Well-motivated problem that closes a real gap. The paper tackles a relevant problem: LLM probabilities are increasingly used jointly rather than in isolation. This was well-motivated by the forecasting, decision-support system and analyst examples in the introduction. In these settings, incoherence creates meaningless conditional probabilities. This justifies the need for exact coherence with high fidelity to the original probabilities, which the paper provides.
  • S2: Strong guarantees and high empirical performance. Coherence holds for every input and every tolerance, not only approximately and the method provides a certificate for the coherence of the output. The suboptimality of the solution is bound by a user-defined tolerance τ\tau that allows to get higher fidelity by investing more compute. Empirically, the projection also lowers RMSE to the ground-truth probabilities in most settings (35 of 40), including all 30 settings for Gemma 2, GLM-4.5 and Qwen 2.5, and it achieves the lowest mean RMSE for these three models.
Critical weaknesses
  • W1: Limited methodological contribution and novelty. The paper’s main contribution is formulating coherence enforcement as a projection, (i.e. an NNLS problem over the cone of the correlation polytope), which is arguably straight-forward. All other components exist already (the characterization of coherence, column generation, and Lawson–Hanson active-set NNLS). The guarantees in Propositions 1 and 2 follow fairly directly from these standard tools. In addition, the running time is exponential in the number of events in the worst case, which limits the method’s applicability as a general solution.
Questions for authors
  • Q1: Effect on stronger models. The projection increases RMSE for GPT-OSS, the most accurate model, on all five coin premises. Could you evaluate additional strong models, such as larger or frontier models, to show whether the accuracy gains persist as raw estimates improve?
  • Q2: Comparison to Zhu et al. Your event families follow Zhu et al. (2025), and all four evaluated models are open-weight, so their embedding-based approach with an axiomatically constrained autoencoder seems directly applicable. How does this method perform in accuracy and remaining incoherence? Does applying your projection on top of their recovered probabilities give further gains?
  • Q3: Comparison to better elicitation. How does the projection compare to stronger elicitation strategies, such as chain-of-thought prompting or asking the model for the joint probabilities of several events in one query? It would also help to see whether projection still improves accuracy when applied on top of these strategies. The paper argues they are complementary, but without results it is unclear how much incoherence they already remove and how much the projection adds.
Reviewer_qyfhOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a post-processing method that makes the pairwise event probabilities produced by an LLM exactly coherent. The authors first show that LLM outputs (marginals and pairwise intersections) are not coherent, i.e., no joint distribution realizes them, and therefore need to be converted into coherent matrices. Deciding coherence is NP-complete, so computing the exact projection is NP-hard. They solve the projection with column generation, which gradually builds a coherent matrix by adding “possible worlds” one at a time. Empirically, the algorithm runs much faster on LLM-produced matrices than on random matrices of the same size.

They provide bounds on how close the algorithm’s output is to the optimal projection.

Furthermore, they show that their method attains the lowest RMSE with respect to the true probabilities (J*) in most settings (29 of 40), and that, unlike the non-trivial baselines, it always returns a coherent output.

Critical strengths

The math, problem statement, and approach are clear.

Critical weaknesses

The choice of LLMs is limited. I would like the authors to consider a wider set of LLMs, including frontier models with reasoning enabled, and models that can use tools (e.g., Python). The paper also does not state whether reasoning was enabled for GPT-OSS and GLM-4.5, which both support it.

How does the method compare to the baselines with stronger models? Table 3 already shows that for GPT-OSS, the raw output is better than the projection on all five coin premises. Oracle scaling also does well there, but its RMSE is identical to the raw output on 9 of 10 premises. This suggests the fit is degenerate for this model and should be explained.

If larger LLMs, or those capable of reasoning, already produce nearly coherent and accurate probabilities, the necessity of this method is unclear. The authors should quantify how much incoherence remains for strong models.

The reason LLM output matrices are easier to optimize is not clear to me, and an analysis would be valuable: what properties do LLM outputs have that random matrices lack? Note also that random U[0,1] matrices are likely much farther from the coherent set than LLM outputs. A fairer comparison would use coherent matrices perturbed by noise, matched to the LLM outputs’ distance to the coherent set.

A more extensive study, covering more LLMs and tasks beyond coin flips and dice, would be very useful.

Questions for authors

How would a random matrix (e.g., the (A + Aᵀ)/2 with Aᵢⱼ ~ U[0,1] matrices used in Fig. 3) score in Table 2? If its distance to the coherent set is comparable to that of the evaluated LLMs, this is a significant concern. It would suggest that the chosen models, which are mostly small (e.g., Gemma-2-2B, Qwen2.5-7B), produce outputs that are barely more coherent than noise. The reported improvements would then mainly reflect correcting near-random predictions, rather than refining informative but slightly incoherent judgments. This concern is reinforced by GPT-OSS, the strongest model in the study, where the projection increases error on all five coin premises. I would ask the authors to (i) add a random-matrix reference row to Table 2 (and the corresponding projected RMSE to Table 3), and (ii) evaluate stronger models to show that the method remains beneficial when the raw outputs are far from random.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.