Entropic Particle Filtering (ePF)

• Entropic Particle Filtering (ePF) is a particle-based Monte Carlo algorithm that combines Entropic Annealing (EA) and Look-ahead Modulation (LaM) to maintain diversity during inference-time scaling in language models.
• It dynamically mitigates premature exploitation by adjusting the resampling distribution through temperature control and predictive lookahead, balancing exploration with exploitation.
• ePF demonstrates significant improvements over standard Particle Filtering in complex math reasoning tasks, achieving up to a 50% relative increase in task rewards on challenging benchmarks.

Searching arXiv for the specified papers and closely related work to ground the article in current literature. arXiv search query: (Giannone et al., 7 Oct 2025) Fetching metadata from the arXiv API for the cited papers. $$H_n(t) = -\frac{\sum_{i=1}^N w_t^i \log w_t^i}{\log N},$$0

$$H_n(t) = -\frac{\sum_{i=1}^N w_t^i \log w_t^i}{\log N},$$1

Running a quick API lookup. $$H_n(t) = -\frac{\sum_{i=1}^N w_t^i \log w_t^i}{\log N},$$2

Running a quick API lookup. Entropic Particle Filtering (ePF) is a particle-based Monte Carlo algorithm introduced for inference-time scaling in LLMs, particularly on complex mathematical reasoning tasks. It is designed for the setting in which Particle Filtering (PF) is guided by process reward models (PRMs) and becomes vulnerable to premature exploitation: the filter can commit too early to locally promising trajectories, prune potentially correct hypotheses, and converge to suboptimal solutions under constrained computational budgets. ePF addresses this failure mode by integrating two mechanisms—Entropic Annealing (EA) and Look-ahead Modulation (LaM)—that preserve search diversity and incorporate a predictive signal about future trajectory quality. On several challenging math benchmarks, the method is reported to achieve up to a 50 % relative improvement in task reward over prior methods (Giannone et al., 7 Oct 2025).

1. Problem setting and definition

ePF is formulated in the context of inference-time scaling (ITS), where additional computation is allocated at generation time to improve language-model performance. Within this setting, PF is used as a sequential Monte Carlo procedure over partial reasoning trajectories. At each step $t$, PF maintains $N$ particles $\{z_t^i\}$, propagates them, scores them with a PRM, and resamples proportionally to the resulting weights. The paper identifies this pipeline as effective but fragile when PRMs are overconfident early in the reasoning process (Giannone et al., 7 Oct 2025).

The central failure mode is particle impoverishment. In the formulation given for ePF, overconfident resampling produces a sharply peaked distribution over particles, causing effective sample size and entropy to collapse. The resulting search becomes myopic: resampling depends only on current rewards and disregards future potential. The paper states that this is especially severe under constrained computational or memory budgets, and that it leads to convergence on locally optimal but globally suboptimal solutions (Giannone et al., 7 Oct 2025).

In this sense, ePF is not merely a replacement resampler; it is a PF variant intended to rebalance exploration and exploitation during multi-step reasoning. The method is specifically motivated by PRM-guided reasoning with LLMs, rather than by generic state estimation alone.

2. Diagnosed failure modes in standard PF

The ePF paper isolates two root causes for degraded PF performance in ITS. The first is a lack of diversity in the particle set due to overconfident resampling. The second is an inability to assess the potential of a reasoning path beyond its current reward. These are described respectively as particle impoverishment and premature exploitation or myopia (Giannone et al., 7 Oct 2025).

To quantify these effects, the method monitors several diagnostic quantities. The first is the normalized entropy of the weight distribution,

$$H_n(t) = -\frac{\sum_{i=1}^N w_t^i \log w_t^i}{\log N},$$

where $w_t^i$ are normalized particle weights at step $t$. The second is normalized Effective Sample Size (ESS). The paper also tracks variance of the resampling distribution, with high variance presented as an indicator of susceptibility to particle collapse (Giannone et al., 7 Oct 2025).

The reported empirical evidence is that PRMs tend to be overconfident, and that high variance or low entropy in the particle weights is highly correlated with poor final task success. This diagnostic framing is important because ePF does not treat poor search outcomes as an isolated resampling artifact; instead, it ties them to calibration properties of the reward model and to the sequential nature of long-horizon reasoning.

A common misconception is to equate PF failure in this setting with insufficient particle count alone. The ePF formulation argues more specifically that even with particle-based search, overconfident local scoring can destroy diversity before the task provides enough information to discriminate promising trajectories. This makes the issue structural rather than purely a matter of scaling $N$.

3. Entropic Annealing

Entropic Annealing (EA) is the first of the two mechanisms that define ePF. Its stated purpose is to dynamically mitigate premature particle collapse by flattening the resampling distribution when diversity is low. Rather than always resampling from the raw PRM-induced softmax, EA introduces a temperature parameter $\beta_t$ and adjusts it according to diversity (Giannone et al., 7 Oct 2025).

The adaptive temperature schedule is given as

$\beta_t^{-1} = \frac{N}{ESS(t)} (1 - t/T),$

where $ESS(t) = \sum_{i=1}^N (w_t^i)^{-2}$, and $N$0 is the reasoning trajectory length. The annealed resampling distribution is

$N$1

If $N$2, entropic annealing is invoked; the details note $N$3 with the example $N$4 and refer to Algorithm 2 in the appendix (Giannone et al., 7 Oct 2025).

The behavior described for EA is explicitly stage-dependent. Early in a trajectory, when ESS is usually low, temperature is high and $N$5 is low, so particles are resampled nearly uniformly. Later in the trajectory, temperature anneals back to $N$6, which concentrates computation on high-scoring hypotheses. The paper presents this as a way to prevent the filter from locking into a narrow set of reasoning paths before sufficient evidence has accumulated (Giannone et al., 7 Oct 2025).

The implementation also uses systematic rather than multinomial resampling. In the reported account, this reduces random fluctuations and preserves diversity further. The method therefore changes both the shape of the resampling distribution and the resampling scheme itself.

4. Look-ahead Modulation

Look-ahead Modulation (LaM) is the second mechanism in ePF and is intended to address PF’s inherent myopia. Whereas EA acts on the entropy of the current resampling distribution, LaM injects a one-step predictive signal about a trajectory’s immediate successors (Giannone et al., 7 Oct 2025).

For each particle $N$7 at step $N$8, the method samples a one-step lookahead $N$9, scores that successor with the PRM to obtain $\{z_t^i\}$0, and then forms modulated weights

$\{z_t^i\}$1

followed by

$\{z_t^i\}$2

The resampling distribution is thus influenced jointly by the current PRM score and the predicted quality of the next-step continuation (Giannone et al., 7 Oct 2025).

The lookahead states are discarded after modulation. The paper emphasizes this point to indicate that the main filter remains consistent with the underlying model dynamics. LaM is characterized as a light-weight extension compared to full auxiliary particle filters (APF), with small computational overhead relative to the gains, especially because it is invoked during diversity collapse rather than at every step (Giannone et al., 7 Oct 2025).

Taken together, EA and LaM operate on different failure modes. EA preserves diversity when the current particle population starts to collapse; LaM reduces overcommitment to trajectories that look good only under short-horizon reward evaluation. This suggests a division of labor within ePF: one mechanism stabilizes the particle population, and the other improves the quality of the ranking signal used for resampling.

5. Algorithmic workflow and empirical behavior

The paper summarizes ePF as an iterative procedure over $\{z_t^i\}$3: propagate particles $\{z_t^i\}$4; score each particle with the PRM using $\{z_t^i\}$5; compute normalized softmax weights; invoke entropic annealing if diversity is low; optionally apply LaM; and resample using systematic resampling (Giannone et al., 7 Oct 2025).

This workflow preserves the canonical PF structure—propagation, weighting, resampling—while altering the logic that determines when and how resampling should exploit current scores. The diagnostics are intrinsic to the procedure rather than post hoc evaluation metrics: entropy, ESS, and variance directly govern intervention by EA and, in the reported implementation, the additional lookahead mechanism.

Empirically, the method is evaluated on GSM8K, MATH500, DEEPMATH, OMNIMATH, AIME-2024, and AIME-2025, using Qwen2.5/3 and Llama models, both generalist and specialist. The paper reports consistent gains over classic PF and over strong baselines including Self-Consistency, Best-of-N, and Beam Search, with larger gains as task complexity increases or compute budgets tighten (Giannone et al., 7 Oct 2025).

One highlighted result is on AIME-2025 with Qwen3-1.7B and 12k sequence length: PF reaches 26.6% pass@1, Best-of-N 28.8%, and ePF 38.9% pass@1. The summary table also reports GSM8K pass@1 of 94.3 for PF, Best-of-N, and ePF, and MATH500 pass@1 of 65.1 for the same three methods. For ePF + LaM, the sample table gives $\{z_t^i\}$6 on AIME-2025, with other entries omitted there (Giannone et al., 7 Oct 2025).

Algorithm	AIME-2025 pass@1 (%)	Notes
PF	26.6	Qwen3-1.7B, 12k sequence length
Best-of-N	28.8	Same setting
ePF	38.9	Same setting

The reported ablations state that EA is most effective when annealing is dynamically scheduled using ESS, and that LaM provides robust additive improvement on top of EA. The paper further states that ePF with LaM achieves the same or better performance as PF with up to 4x fewer particles, which it interprets as improved sample efficiency and scalability (Giannone et al., 7 Oct 2025).

6. Relation to other “entropic” PF methods and broader significance

The term “entropic” appears in more than one PF research line, but the mechanisms differ. In "Differentiable Particle Filtering via Entropy-Regularized Optimal Transport" (Corenflos et al., 2021), entropy regularization is used in an optimal transport resampling scheme to make PF differentiable with respect to model and proposal parameters. There, the transport plan is computed via entropy-regularized optimal transport and Sinkhorn iterations, yielding a differentiable ensemble transform: $\{z_t^i\}$7 That work addresses non-differentiability of resampling in state-space models and variational inference (Corenflos et al., 2021).

By contrast, ePF uses entropy as a diagnostic and control variable for search diversity in PRM-guided reasoning. Its Entropic Annealing adjusts the entropy of the resampling distribution when ESS or entropy indicates collapse; its Look-ahead Modulation supplies a predictive signal about successor quality (Giannone et al., 7 Oct 2025). The shared terminology therefore does not imply a shared mechanism. One line uses entropy-regularized optimal transport to enable end-to-end differentiation; the other uses entropy-aware annealing to prevent premature exploitation during inference-time search.

This distinction matters conceptually. The differentiable PF literature focuses on gradient flow through the resampling step, low-variance gradient estimates, and consistency as $\{z_t^i\}$8 and $\{z_t^i\}$9 (Corenflos et al., 2021). The ePF literature focuses on exploration–exploitation balance, particle diversity, and robustness under constrained inference budgets in mathematical reasoning (Giannone et al., 7 Oct 2025). A plausible implication is that “entropic” PF methods now span at least two separate methodological agendas: differentiability in sequential inference, and diversity preservation in search over reasoning trajectories.

Within ITS for LLMs, the significance attributed to ePF is practical as well as methodological. The paper states that even non-thinking, generalist LLMs become competitive with specialist “reasoner” models when paired with ePF, and that the majority of the additional compute comes from occasional lookahead passes and the use of systematic resampling, with modest actual overhead because the interventions are only engaged when diversity drops (Giannone et al., 7 Oct 2025). In that formulation, ePF serves as a robust PF variant for long-horizon reasoning in which local reward estimates are informative but not reliably calibrated.