Sharpened Logarithmic Opinion Pool (SLOP)

• Sharpened Logarithmic Opinion Pool (SLOP) is a generalized logarithmic aggregation method that combines expert distributions using arbitrary real-valued weights, enabling both sharpening and inversion.
• It computes pooled distributions by summing weighted log-densities, which facilitates inference-time alignment and mitigates reward hacking through calibrated weight adjustments.
• Empirical evaluations in visual question answering and mathematical reasoning demonstrate that SLOP improves performance compared to standard sampling and baseline approaches.

Searching arXiv for the specified papers and closely related work on logarithmic opinion pools and SLOP. Sharpened Logarithmic Opinion Pool (SLOP) denotes a class of logarithmic pooling methods in which expert distributions are combined through weighted sums of log-densities, but with a notion of “sharpening” that relaxes the conventional constraints of standard logarithmic opinion pools. In the 2026 usage, SLOP is defined as an inference-time alignment distribution over generative and reward experts with arbitrary real-valued weights, including negative values, so that the pooled distribution can sharpen toward confident modes or invert anti-aligned experts (Wang et al., 13 May 2026). In earlier probability-aggregation and Gaussian-process literature, the same acronym is not standard: the 2015 Gaussian-process paper does not use the name SLOP, but its diversified log opinion pool mechanism can be read as a conceptual precursor that “sharpens” a transductive log opinion pool by reweighting informative, nonredundant experts (Cao et al., 2015). A separate 2012 Bayesian aggregation framework develops weighted logarithmic opinion pools with dependence-aware weights and motivates post-aggregation extremizing to correct underconfidence, again under a “sharpened” interpretation, though this terminology is not standard in that paper’s title (Kahn, 2012).

1. Definition and formal variants

A standard logarithmic opinion pool (LOP) combines expert distributions $p_i$ through

$\log p(x) \propto \sum_i w_i \log p_i(x),$

equivalently

$p(x) \propto \prod_i p_i(x)^{w_i}.$

This form appears explicitly in both the Gaussian-process expert literature and the 2026 alignment paper (Cao et al., 2015, Wang et al., 13 May 2026).

The 2026 SLOP formulation generalizes this by defining expert scores $s_i(x,y)$ as either generative log-likelihoods $\log p_i(y \mid x)$ or reward scores $r_i(x,y) \equiv \log q_i(y \mid x)$, with

$$\pi_\omega^*(y \mid x) = \mathrm{softmax}\left( \sum_{i=1}^m \omega_i s_i(x, y) \right) \propto \prod_{i=1}^m p_i(y \mid x)^{\omega_i},$$

where $\omega_i \in \mathbb{R}$ are arbitrary real weights (Wang et al., 13 May 2026). The key distinction from standard LOP is precisely this relaxation: standard LOP typically uses nonnegative weights summing to one, whereas SLOP allows negative weights and large magnitudes. In the 2026 paper, the term “sharpening” refers to this departure from conventional pooling constraints, enabling stronger concentration on selected modes and inversion of misaligned experts (Wang et al., 13 May 2026).

In the 2015 Gaussian-process setting, the LOP solution is obtained by minimizing a weighted sum of Kullback–Leibler divergences,

$$\min_p \sum_i w_i \, \mathrm{KL}(p \,\|\, p_i) \quad \text{subject to} \quad \int p(y \mid x)\,dy=1,\; w_i \ge 0,\; \sum_i w_i =1,$$

whose unique solution is again the geometric pool

$$p(y \mid x) \propto \prod_i p_i(y \mid x)^{w_i(x)}.$$

There, the weights are test-location dependent and remain nonnegative and normalized (Cao et al., 2015). This is not called SLOP in the paper; however, the later diversified reweighting mechanism is explicitly designed to increase diversity and reduce redundancy, which plausibly motivates describing it as a sharpened or diversified LOP in retrospect (Cao et al., 2015).

The 2012 Bayesian aggregation model yields a weighted LogOP on the log-odds scale for binary events, with an intercept determined by the prior and expert biases: $\log p(x) \propto \sum_i w_i \log p_i(x),$0 It also discusses a post-aggregation extremizing transformation,

$\log p(x) \propto \sum_i w_i \log p_i(x),$1

with center $\log p(x) \propto \sum_i w_i \log p_i(x),$2 equal to the prior log-odds and sharpening parameter $\log p(x) \propto \sum_i w_i \log p_i(x),$3 for underconfidence correction (Kahn, 2012). This is a different formalization from the 2026 one: rather than allowing arbitrary real-valued expert weights inside the pool, it sharpens the pooled output after aggregation.

2. The 2026 inference-time alignment formulation

The 2026 formulation situates SLOP within inference-time alignment. It starts from KL-regularized reward maximization,

$\log p(x) \propto \sum_i w_i \log p_i(x),$4

whose optimizer has the tilted form

$\log p(x) \propto \sum_i w_i \log p_i(x),$5

This connects reward tilting to product-of-experts modeling: the reference model $\log p(x) \propto \sum_i w_i \log p_i(x),$6 and a reward-induced alignment distribution $\log p(x) \propto \sum_i w_i \log p_i(x),$7 are pooled logarithmically (Wang et al., 13 May 2026).

The paper then introduces temperature adjustment of the reference model. The generalized optimum is

$\log p(x) \propto \sum_i w_i \log p_i(x),$8

or equivalently

$\log p(x) \propto \sum_i w_i \log p_i(x),$9

Here $p(x) \propto \prod_i p_i(x)^{w_i}.$0 acts as an inverse temperature on the reference model, while $p(x) \propto \prod_i p_i(x)^{w_i}.$1 controls reward tilt (Wang et al., 13 May 2026). This two-expert temper-and-tilt form is a special case of SLOP with $p(x) \propto \prod_i p_i(x)^{w_i}.$2.

The resulting perspective unifies several families of methods. Best-of-$p(x) \propto \prod_i p_i(x)^{w_i}.$3, Soft Best-of-$p(x) \propto \prod_i p_i(x)^{w_i}.$4, and related heuristics become approximations to sampling from tilted or pooled target distributions. Product-of-experts guidance and direct logit steering become special cases of summing weighted log-densities. Reward-model ensembling becomes principled under a single probabilistic pooling rule (Wang et al., 13 May 2026).

A central feature is the admissibility of negative weights. The paper states that when a proxy reward is anti-aligned, allowing $p(x) \propto \prod_i p_i(x)^{w_i}.$5 treats that proxy as an “anti-expert,” so its high-scoring outputs are actively suppressed rather than reinforced (Wang et al., 13 May 2026). This is one of the most substantive differences from classical LogOP, where nonnegativity constraints are standard.

3. Sampling, decoding, and approximation theory

Direct sampling from the SLOP distribution is generally intractable for large output spaces. The 2026 paper therefore extends Soft Best-of-$p(x) \propto \prod_i p_i(x)^{w_i}.$6 (SBoN) to the multi-expert SLOP setting (Wang et al., 13 May 2026). Let $p(x) \propto \prod_i p_i(x)^{w_i}.$7 be the reference generative expert. Candidates $p(x) \propto \prod_i p_i(x)^{w_i}.$8 are drawn i.i.d. from $p(x) \propto \prod_i p_i(x)^{w_i}.$9, and each candidate receives pseudo-reward

$s_i(x,y)$0

Selection then uses

$s_i(x,y)$1

The subtraction of the reference score accounts for the fact that proposals are sampled from $s_i(x,y)$2 rather than from the target pooled distribution (Wang et al., 13 May 2026).

The approximation guarantee stated in the paper is

$s_i(x,y)$3

which provides asymptotic consistency of the SBoN extension as the number of candidates grows (Wang et al., 13 May 2026). This guarantee gives SLOP a theoretical footing absent from many ad hoc inference-time reranking procedures.

The same work also gives an autoregressive token-level view. If all experts provide token conditionals $s_i(x,y)$4, then per-step SLOP decoding takes the form

$s_i(x,y)$5

with logits

$s_i(x,y)$6

This makes SLOP compatible with classifier-free-guidance-style logit adjustments and direct token-level steering (Wang et al., 13 May 2026).

A practical distinction follows. Candidate-level SLOP is appropriate when some experts can only score completed sequences. Token-level SLOP is a direct-decoding alternative when all experts expose token probabilities (Wang et al., 13 May 2026). This suggests that SLOP is less a single decoding algorithm than a distributional principle that can be instantiated by reranking or by guided generation.

4. Weight calibration and reward hacking mitigation

The 2026 paper places calibration at the center of SLOP’s practical use. Its motivating concern is reward hacking: optimizing strongly toward an imperfect proxy reward may move outputs into modes that score well under the proxy while failing on the gold objective (Wang et al., 13 May 2026). In an ensemble setting, correlated experts can amplify the same failure modes.

To mitigate this, the paper proposes calibrating SLOP weights on a small set of prompts with verifiable gold rewards. For calibration prompts $s_i(x,y)$7, candidate responses $s_i(x,y)$8 are scored by gold rewards $s_i(x,y)$9 and expert scores $\log p_i(y \mid x)$0. Candidate logits are

$\log p_i(y \mid x)$1

with selection probabilities

$\log p_i(y \mid x)$2

The empirical objective is

$\log p_i(y \mid x)$3

optimized by gradient ascent, with initialization $\log p_i(y \mid x)$4 (Wang et al., 13 May 2026).

The paper notes an important theoretical subtlety: the calibration objective may fail to admit a finite maximizer, because the supremum can occur as weights diverge. Weight decay regularization is introduced to discourage such divergence in practice (Wang et al., 13 May 2026). This is not merely a numerical detail; it implies that “sharpening” can become arbitrarily extreme unless constrained by regularization or by the geometry of the data.

The same work analyzes optimal weights under a Gaussian score model. If the log-posterior ratios on the correct answer are jointly Gaussian with mean $\log p_i(y \mid x)$5 and covariance $\log p_i(y \mid x)$6, then the weight vector

$\log p_i(y \mid x)$7

maximizes the detection margin $\log p_i(y \mid x)$8 (Wang et al., 13 May 2026). This yields a principled covariance-corrected weighting rule in which expert quality and inter-expert correlation jointly determine the optimum. A plausible implication is that SLOP’s calibrated weights play a role analogous to Fisher-style linear discrimination in log-score space.

5. Earlier antecedents in logarithmic pooling

Although the 2026 paper introduces the explicit acronym “SLOP,” two earlier lines of work provide important antecedents.

The 2015 paper on transductive Gaussian-process experts develops a log opinion pool framework for combining independently trained GP experts at test time (Cao et al., 2015). It proves that the generalized product of experts (gPoE-GP) is exactly the LOP minimizer under weighted KL aggregation, without requiring conditional independence assumptions or a shared prior. For Gaussian expert posteriors $\log p_i(y \mid x)$9, the pooled posterior remains Gaussian, with precision-weighted formulas

$r_i(x,y) \equiv \log q_i(y \mid x)$0

The paper then introduces diversified log opinion pool of GP experts (dLOP-GP), which modifies the entropy-based gPoE weights by ascending a diversity term based on pairwise symmetric KL divergences between expert posteriors (Cao et al., 2015).

That update is constructed from the approximation

$r_i(x,y) \equiv \log q_i(y \mid x)$1

where $r_i(x,y) \equiv \log q_i(y \mid x)$2 rewards closeness to the unknown target predictive distribution and

$r_i(x,y) \equiv \log q_i(y \mid x)$3

rewards disagreement among experts via the symmetric-KL matrix $r_i(x,y) \equiv \log q_i(y \mid x)$4 (Cao et al., 2015). The resulting one-step ascent on $r_i(x,y) \equiv \log q_i(y \mid x)$5 sharpens the pool toward informative, nonredundant experts. The paper itself calls this diversification rather than sharpening, but conceptually it is a clear precursor to later SLOP language.

The 2012 Bayesian aggregation paper provides a different antecedent. There, a Gaussian generative model over experts’ debiased log-odds yields dependence-aware weights for a weighted LogOP, with properties such as the external Bayesian property when weights sum to one (Kahn, 2012). The same account explicitly motivates extremizing the pooled posterior to counter dependence-induced underconfidence. In that formulation, SLOP is not an ensemble with arbitrary real-valued expert weights but a post-aggregation transformation that scales pooled log-odds away from the prior center (Kahn, 2012).

These antecedents illustrate that “sharpened logarithmic opinion pool” has at least two technical meanings in the literature-adjacent discourse: internal sharpening through unconstrained real-valued weights inside the pool, and external sharpening through an extremizing map applied after a standard pool.

6. Empirical behavior and task domains

The 2026 paper evaluates SLOP in two settings: reward-guided visual question answering and mathematical reasoning (Wang et al., 13 May 2026).

For ScienceQA multiple-choice VQA, the reference VLM is LLaVA-1.5-7B, with exact temper-and-tilt sampling over the answer set $r_i(x,y) \equiv \log q_i(y \mid x)$6. Calibration uses 200 samples from the test split, with evaluation on the remaining 4041 samples. The reported baselines are 63.4% for reference sampling $r_i(x,y) \equiv \log q_i(y \mid x)$7 and 66.1% for greedy decoding $r_i(x,y) \equiv \log q_i(y \mid x)$8. Proxy-only optimization $r_i(x,y) \equiv \log q_i(y \mid x)$9 performs well only when the proxy is accurate, and degrades rapidly otherwise. Calibrated SLOP “generally dominates individual models,” degrades gracefully as proxy accuracy falls, and uses $$\pi_\omega^*(y \mid x) = \mathrm{softmax}\left( \sum_{i=1}^m \omega_i s_i(x, y) \right) \propto \prod_{i=1}^m p_i(y \mid x)^{\omega_i},$$0 when proxy quality drops below random, thereby converting the proxy into an anti-expert (Wang et al., 13 May 2026).

For GSM8K, the paper pools four LLM experts—Gemma-3-1B, Qwen2-1.5B, Qwen3-0.6B, and Phi-3.5-mini—using token-averaged log-likelihoods to avoid length bias. Calibration uses 200 problems and evaluation uses 1119 held-out test problems. SLOP and a hard-selection variant consistently improve over reference sampling and strong Best-of-$$\pi_\omega^*(y \mid x) = \mathrm{softmax}\left( \sum_{i=1}^m \omega_i s_i(x, y) \right) \propto \prod_{i=1}^m p_i(y \mid x)^{\omega_i},$$1 baselines, with gains depending on which model serves as the reference (Wang et al., 13 May 2026).

Reference	Baseline	SLOP
Gemma-3-1B	43.70±0.10	60.13±1.27
Qwen2-1.5B	44.08±0.10	64.23±0.95
Qwen3-0.6B	63.58±0.10	76.05±0.31
Phi-3.5-mini	81.51±0.08	83.91±0.42

The hard-selection variant is slightly stronger still in most of these experiments, while gains are modest when the reference is already the strongest model (Wang et al., 13 May 2026). The paper also reports that covariance-based weights of the form $$\pi_\omega^*(y \mid x) = \mathrm{softmax}\left( \sum_{i=1}^m \omega_i s_i(x, y) \right) \propto \prod_{i=1}^m p_i(y \mid x)^{\omega_i},$$2 can be competitive with calibrated weights when the covariance is well-conditioned (Wang et al., 13 May 2026).

In the Gaussian-process expert literature, empirical results are reported on KIN40K, SARCOS, and UK-APT using SMSE and SNLP metrics. There, dLOP-GP generally improves over gPoE-GP, particularly under subset schemes that induce expert diversity such as independent hyperparameters or varied kernels (Cao et al., 2015). This suggests a recurring pattern across domains: sharpening or diversification is most beneficial when experts are heterogeneous rather than redundant.

7. Conceptual distinctions, limitations, and interpretation

Several distinctions are necessary to avoid conflating different uses of the same phrase.

First, SLOP in the 2026 sense is not simply a synonym for logarithmic opinion pooling. Standard LOP usually constrains weights to be nonnegative and normalized. The 2026 SLOP explicitly abandons those constraints, allowing arbitrary real weights and thereby permitting both sharpening and inversion (Wang et al., 13 May 2026). By contrast, the 2015 GP framework relies crucially on nonnegative weights summing to one, because those constraints underwrite the weighted-KL variational characterization and ensure safe fallback to the prior far from data (Cao et al., 2015).

Second, SLOP differs from mixture-of-experts. A mixture forms convex averages of probabilities; a product-of-experts or logarithmic pool adds log-densities and therefore sharpens on regions of consensus. The 2026 paper stresses that this can improve alignment, but also reduce diversity if left unregularized (Wang et al., 13 May 2026).

Third, negative weights are powerful but hazardous. In the alignment setting, a negative weight can suppress an anti-aligned proxy; however, miscalibration of the gold signal or instability in weight estimation can redirect the model toward undesirable behavior (Wang et al., 13 May 2026). In older LogOP treatments, unconstrained weights are mathematically admissible for some properties such as external Bayesianity, but their probabilistic interpretation is less conservative (Kahn, 2012).

Fourth, expert correlation is a recurring limitation. The Gaussian score analysis in the 2026 paper shows that ensemble gains saturate with correlation, while the 2015 dLOP-GP method is designed explicitly to counter redundancy among correlated GP experts (Wang et al., 13 May 2026, Cao et al., 2015). The 2012 generative aggregation model likewise treats dependence as central, since positive dependence damps effective reinforcement and can motivate post-hoc extremizing (Kahn, 2012).

Finally, calibration dependence remains a central caveat. The 2026 formulation requires a calibration set with verifiable gold rewards; if the calibration target is itself misaligned, SLOP may be tuned toward the wrong objective (Wang et al., 13 May 2026). The 2012 sharpening approach similarly requires validation-based selection of the sharpening parameter $$\pi_\omega^*(y \mid x) = \mathrm{softmax}\left( \sum_{i=1}^m \omega_i s_i(x, y) \right) \propto \prod_{i=1}^m p_i(y \mid x)^{\omega_i},$$3, and over-extremizing can harm proper scoring rules (Kahn, 2012). This suggests that sharpening is not intrinsically beneficial; its value depends on whether the unsharpened pool is underconfident, overconservative, or contaminated by poorly weighted experts.

In contemporary usage, therefore, Sharpened Logarithmic Opinion Pool most precisely refers to the 2026 inference-time alignment framework in which generative and reward experts are combined through a logarithmic pool with arbitrary real-valued weights, calibrated to preserve alignment while mitigating reward hacking (Wang et al., 13 May 2026). Earlier work on GP expert pooling and Bayesian expert aggregation provides closely related mechanisms—diversification, dependence correction, and extremizing—that illuminate the broader technical lineage of the idea (Cao et al., 2015, Kahn, 2012).