ReSum: Resummation & Compression Methods

• ReSum is a suite of multidisciplinary frameworks that address divergent perturbative expansions and information overload using techniques like multi-fidelity surrogates and contour-based resummation.
• The methodologies include RESuM for detector optimization, Meijer G–function resummation in quantum field theory, and unbiased cumulant expansions in lattice simulations, each targeting improved convergence and computational efficiency.
• ReSum strategies also extend to machine learning, where self-summarization and context compression methods enhance long-horizon reasoning and reduce memory constraints in large language models.

ReSum

ReSum refers to several algorithmic paradigms and frameworks, predominantly within high-energy physics, lattice field theory, and machine learning, that target the resummation of divergent or inefficiently convergent perturbative expansions, as well as the strategic compression or reflection of information for tractability in simulation- or reasoning-intensive domains. The term has been instantiated as both a named algorithm ("RESuM"—Rare Event Surrogate Model for detector optimization) and a generic label for advanced resummation methodologies and self-reflective summarization loops within LLM architectures.

1. The Rare Event Surrogate Model (RESuM) for Detector Design

The RESuM framework formalizes the Rare Event Design (RED) problem, encountered when optimizing design parameters $\theta\in\Theta$ for detectors where the figure-of-merit is the minimization of a rare-event background probability. In typical Monte Carlo simulation pipelines, each design $\theta$ is evaluated by generating $N$ events, each yielding a binary indicator $X_i\in\{0,1\}$ for background leakage, producing a metric $y(\theta)=m(\theta)/N$ with $m(\theta)=\sum_{i=1}^N X_i$. Given $m\ll N$, the variance $\mathrm{Var}[y]\approx \mathcal{O}(1/N)$ is substantial unless $N$ is made prohibitively large, making standard surrogate modeling difficult.

RESuM addresses this by constructing a multi-fidelity surrogate across both simulation cost and uncertainty, using a pretrained Conditional Neural Process (CNP) at the event level and a co-kriging Multi-Fidelity Gaussian Process (MF-GP) model at the metric level. The CNP injects nontrivial prior knowledge of the event-level leakage probability $t(\theta,\phi)$, outputting low-variance, continuous surrogate scores $\theta$0, replacing $\theta$1 with $\theta$2 to supply $\theta$3 as a smooth, low-noise estimator. The MF-GP fuses low-fidelity (LF) CNP surrogates, high-fidelity (HF) CNP surrogates, and high-fidelity raw simulation outputs in an autoregressive structure, yielding a holistic, uncertainty-aware surrogate for rare-event metrics across the full spectrum of computational fidelities (Schuetz et al., 2024).

RESuM employs an integrated variance reduction acquisition function, selecting the next design point $\theta$4 to minimize the expected posterior variance across the parameter space, further improving sample efficiency.

In the LEGEND NLDBD experiment, RESuM achieved a $\theta$5 reduction in neutron-induced background using only $\theta$6 of the computational resources required by a traditional grid search, demonstrating both the scalability of the approach in high-dimensional design and its adaptivity in handling rare-event-dominated metrics.

2. Resummation of Divergent Perturbative Series in Quantum Field Theory

Divergent perturbative expansions, which are ubiquitous in quantum field theory due to factorial growth of coefficients (Dyson's instability), are traditionally handled via Borel–Padé resummation. However, this approach is often limited by the requirement for many high-order coefficients, poor handling of branch cuts and renormalon singularities, and slow convergence.

The Meijer G–function resummation framework provides an alternative: given a truncated series $\theta$7, one constructs the Borel-transformed series, models the coefficient ratios with a rational ansatz, and fits a hypergeometric form whose Laplace inversion yields a Meijer G–function. This approach yields an analytic non-perturbative function $\theta$8 that smoothly extrapolates perturbative data into the strong-coupling regime and can even continue past renormalon singularities by extracting the physical real part from the resulting trans-series (Antipin et al., 2018).

Application to scalar $\theta$9 theory and gauge–fermion systems shows fast convergence to the correct asymptotics, accurate reproduction of instanton-dominated large-order behavior, clear diagnosis of UV/IR fixed points, and explicit mapping of phase diagrams (e.g., the conformal window in SU(3) as a function of fermion flavor number).

3. Algorithmic ReSum for Non-Borel Summable Series in 3D Supersymmetric Gauge Theories

In 3D $N$0 Chern–Simons–matter theories, perturbative expansions in $N$1 are generally not Borel-summable along the real axis because the Borel transform develops singularities (from, e.g., one-loop determinant poles) along the positive axis. The "ReSum" prescription replaces the ill-defined real-axis Laplace transform with a contour along the imaginary axis ($N$2, depending on sign of the Chern–Simons level). For example, for $N$3 SQCD on $N$4, the partition function after suitable variable transformation becomes explicitly a Laplace-type integral along $N$5, whose analytic continuation matches precisely the Borel sum of the perturbative expansion along this contour (Honda, 2016).

This yields exact, non-perturbative results for partition functions and BPS observables from perturbative data alone, without requiring explicit non-perturbative saddle-point contributions. This formalism applies to a broad class of observables and manifolds, and inherently incorporates the resolution of semi-classical saddles (e.g., Higgs branch, vortices) via contour choice.

4. Resummation Techniques in Lattice Field Theory and Statistical Mechanics

In finite-density lattice field theory, Taylor expansions in chemical potential $N$6 are limited by the difficulty of calculating higher-order derivatives and the poor convergence at large $N$7. Standard exponential resummation of low-order charge density correlators introduces a bias due to stochastic estimation of correlators, especially as $N$8 and $N$9 grow. A cumulant expansion, where only finitely many unbiased products of correlators are retained (truncated at order $X_i\in\{0,1\}$0), provides superior convergence properties—the cumulant truncation character recovers the Taylor expansion up to $X_i\in\{0,1\}$1 and non-perturbatively incorporates resummed contributions from lower-order terms without bias (Mitra et al., 2022).

Numerically, unbiased cumulant resummation tracks higher-order Taylor results much more closely than either a low-order Taylor expansion or the biased exponential resummation, particularly at large $X_i\in\{0,1\}$2.

5. ReSum in Machine Learning: Self-Summarization and Context Compression

A distinct family of "ReSum" methods has emerged in contemporary LLM research, focusing on the strategic compression of reasoning trajectories via self-summarization actions, both to improve long-horizon reasoning and to manage memory constraints.

In RLVR (Reinforcement Learning with Verifiable Rewards), ReSum augments the token-level action space with "summarization-trigger" tokens, allowing LLMs to periodically condense their ongoing reasoning into short summary tokens. A contrastive rollout paradigm generates paired continuations with and without summarization, and a summarization-aware advantage function incentivizes precise and beneficial summarization events. Empirically, this approach yields a consistent $X_i\in\{0,1\}$3 absolute increase in exact-match accuracy on difficult mathematical benchmarks, with an $X_i\in\{0,1\}$4 reduction in average output length, demonstrating both improved reasoning efficiency and reduced overthinking effects (Wang et al., 11 Jun 2026).

Other ReSum paradigms within the LLM web agent literature condition agent states on periodic summaries of extended search trajectories, circumventing the fixed context-length bottleneck and enabling web agents to reason and explore for indefinite horizons. Reinforcement learning with grouped policy optimization (ReSum-GRPO) further adapts the agent's policy to operate efficiently in a summary-conditioned state, with average absolute performance gains of 4.5–8.2\% compared to untuned context-chaining or naïve ReAct baselines (Wu et al., 16 Sep 2025).

6. Broader Applicability and Impact

The underlying principles of ReSum—whether in rare event surrogate modeling, resummation of divergent series, or adaptive summarization of reasoning—are unified by goals of (i) extracting robust, uncertainty-aware predictions or policies from limited or intractable data, and (ii) achieving significant computational savings compared to naïve, brute-force or fixed-order algorithms.

The RED and RESuM formalisms are directly transferable to complex simulation-based scientific settings beyond neutrino detection, including but not limited to gravitational-wave event rate modeling, materials discovery (defect or failure rate prediction), and any design optimization over rare-event-dominated observables where exact simulation is costly or the underlying variance is high (Schuetz et al., 2024).

Advanced resummation and context compression strategies are likewise central to high-precision theoretical predictions in quantum field theory, non-equilibrium statistical mechanics, and the design of scalable, tractable large-scale machine learning systems.

7. Summary Table: Major ReSum Instantiations

Domain	Formalism/Model	Key Technical Approach
Detector Optimization	RESuM (RED + MF-GP + CNP)	Co-kriging, neural process prior, active learning, variance reduction (Schuetz et al., 2024)
QFT	Meijer G–function Resummation	Hypergeometric ansatz for Borel sum, Laplace inversion, non-perturbative analytic continuation (Antipin et al., 2018)
QFT (3D SUSY)	ReSum contour for Borel–Laplace integrals	Imaginary-axis Borel summation, uniform convergence (Honda, 2016)
Lattice QCD	Unbiased Cumulant Expansion	Truncated cumulant sum, finite unbiased correlator products (Mitra et al., 2022)
LLM/RL	Self-Summarization (ReSum, ReSum-GRPO)	Summarization-aware RL, context compression, grouped policy objectives (Wang et al., 11 Jun 2026, Wu et al., 16 Sep 2025)

The evolution and convergence of ReSum paradigms across disciplines demonstrate the centrality of resummation, compression, and surrogate modeling to both theoretical and computational advances in contemporary science.