SIMMER in Science: Multi-Domain Methods

• SIMMER is a multi-domain term referring to distinct research artifacts, from astrophysical convective burning and hydrogen simmering to AI planning benchmarks and simulation frameworks.
• In astrophysics, SIMMER characterizes pre-supernova convection and X-ray burst hydrogen burning with quantified turbulent spectra and ignition dynamics.
• In computational science, SIMMER encompasses an R discrete-event simulation package, a multimodal food image–recipe retrieval model, a safe-exploration method for RL, and an astronomical image-reduction pipeline.

Across the cited literature, SIMMER, simmer, Simmer, and SImMER do not denote a single concept. They refer instead to several unrelated research artifacts and physical regimes: stellar simmering before Type Ia supernova ignition, hydrogen simmer in thermonuclear X-ray bursts, a benchmark for latent failures in LLM executable planning, a discrete-event simulation framework for R, a multimodal embedding model for food image–recipe retrieval, a safe-exploration method for reinforcement learning, and a Python pipeline for reducing astronomical images of stars (Nonaka et al., 2011, Schwab et al., 2017, Zand et al., 2017, Lu et al., 12 Jun 2026, Ucar et al., 2017, Gomi et al., 17 Apr 2026, Sootla et al., 2022, Savel et al., 2022).

1. Scope of the term in research usage

The term appears in multiple capitalization schemes and disciplinary contexts. The following summary captures the principal usages documented in the cited papers.

Form	Domain	Definition
SIMMER	LLM planning	A benchmark for latent failures in executable kitchen planning with a symbolic world model comprising 77 actions, 262 unique objects, and approximately 46,800 possible interactions (Lu et al., 12 Jun 2026)
simmer	Discrete-event simulation	An R package for process-oriented DES with a C++ core and a trajectory-centered API (Ucar et al., 2017)
simmer	Network simulation	The same R package applied to three 5G scenarios, including crosshauling, mobile backhauling over FTTx, and massive IoT over LTE (Ucar et al., 2018)
SIMMER	Multimodal retrieval	“Single Integrated Multimodal Model for Embedding Recipes,” a single-encoder food image–recipe retrieval system based on VLM2Vec (Gomi et al., 17 Apr 2026)
Simmer	Safe RL	“Safe policy IMproveMEnt for RL,” based on safety-state augmentation and budget scheduling during training (Sootla et al., 2022)
SImMER	Astronomical software	A Python pipeline for reduction and analysis of point-source images from ShARCS and PHARO (Savel et al., 2022)
simmering / simmer	Astrophysics	Late-stage convective burning in Chandrasekhar-mass white dwarfs and residual rp-process burning in X-ray burst tails (Nonaka et al., 2011, Zand et al., 2017)

This distribution suggests that the label functions primarily as a local acronym or descriptive term within each subfield rather than as a transdisciplinary research program.

2. Astrophysical meanings: convective simmering and hydrogen simmer

In Type Ia supernova progenitors, simmering denotes the convective phase immediately preceding runaway ignition in a Chandrasekhar-mass white dwarf. High-resolution MAESTRO simulations extended earlier full-star calculations to effective resolutions of 4.34 km and 2.17 km and found that off-center ignition is likely, with radius 50 km most favored and a likely range of 40 to 75 km. The hottest-cell statistics stabilize over the final 100–200 s, nearly all candidate hot spots have outward radial velocity, and a multiple ignition scenario is not likely. At 2.17 km zoning, the convective core shows a narrow strong outward plume with solid angle approximately $\pi\,\mathrm{sr}$ rising at $(5$–$7)\times10^6\,\mathrm{cm\,s^{-1}}$, surrounded by lower-speed recirculation; outside the core, stably stratified layers support circumferential shear flows up to $\gtrsim2.5\times10^7\,\mathrm{cm\,s^{-1}}$ (Nonaka et al., 2011).

The same simulations quantified the convective core as turbulent with a Kolmogorov spectrum,

$E(k)=C_K\,\varepsilon^{2/3}\,k^{-5/3},$

with $C_K\approx1.5$, turbulent intensity $u' \simeq 16\,\mathrm{km\,s^{-1}}$, integral length scale $L \simeq 200\,\mathrm{km}$, and implied dissipation rate $$\varepsilon \simeq 2\times10^{11}\,\mathrm{cm^2\,s^{-3}}$$. The spectra collapse across resolutions to an inertial-range slope close to $k^{-5/3}$ for $(5$0, supporting a Kolmogorov cascade in the convective core (Nonaka et al., 2011).

A longer-timescale view of carbon simmering was developed in stellar-evolution calculations of single-degenerate Type Ia progenitors. In these models, the convective simmering phase lasts of order $(5$1–$(5$2 yr while the central temperature rises from roughly $(5$3 to $(5$4 and the central density increases from about $(5$5–$(5$6 to $(5$7–$(5$8. The paper emphasizes that weak-reaction rate choices shift the ignition density and final $(5$9 by $7)\times10^6\,\mathrm{cm\,s^{-1}}$0, whereas the mixing algorithm alters $7)\times10^6\,\mathrm{cm\,s^{-1}}$1 by $7)\times10^6\,\mathrm{cm\,s^{-1}}$2, and the treatment of the convective Urca process can change the total carbon burned by factors of several. In representative MESA models, diffusive mixing gives $7)\times10^6\,\mathrm{cm\,s^{-1}}$3 at $7)\times10^6\,\mathrm{cm\,s^{-1}}$4, advective mixing gives $7)\times10^6\,\mathrm{cm\,s^{-1}}$5, and the cumulative convective work can rise to $7)\times10^6\,\mathrm{cm\,s^{-1}}$6 by $7)\times10^6\,\mathrm{cm\,s^{-1}}$7 (Schwab et al., 2017).

A separate astrophysical usage appears in thermonuclear X-ray bursts, where hydrogen can continue to burn after the main flash through the rp process. The burst tail is modeled empirically as a power law plus a one-sided Gaussian,

$7)\times10^6\,\mathrm{cm\,s^{-1}}$8

In a sample of 1254 RXTE/PCA bursts, the power-law index lies between about 1.3 and 2.1 for 80% of bursts, the Gaussian is detected in half of all bursts, and the characteristic width is $7)\times10^6\,\mathrm{cm\,s^{-1}}$9. The Gaussian fluence fraction reaches up to about 60%, while 94% of bursts from ultracompact X-ray binaries lack the Gaussian component; the inferred rp-burning layer is underabundant in hydrogen by at least a factor of five relative to cosmic abundances (Zand et al., 2017).

3. SIMMER as a benchmark for latent failures in LLM planning

In machine learning, SIMMER is a benchmark for executable planning in a household kitchen domain that targets latent failures rather than only immediate execution errors. Its symbolic world model contains 77 canonical actions, 262 items across appliances, tools, cookware, ingredients, and fixtures, and approximately 46,800 semantically valid interactions derived from real-world cooking scripts. States are represented as

$\gtrsim2.5\times10^7\,\mathrm{cm\,s^{-1}}$0

with agent state $\gtrsim2.5\times10^7\,\mathrm{cm\,s^{-1}}$1, and each action is specified as

$\gtrsim2.5\times10^7\,\mathrm{cm\,s^{-1}}$2

Execution uses deterministic updates: if all predicates in $\gtrsim2.5\times10^7\,\mathrm{cm\,s^{-1}}$3 hold, effects map $\gtrsim2.5\times10^7\,\mathrm{cm\,s^{-1}}$4 to $\gtrsim2.5\times10^7\,\mathrm{cm\,s^{-1}}$5; otherwise the step is labeled an immediate failure (Lu et al., 12 Jun 2026).

The benchmark’s central contribution is its failure taxonomy. Immediate failures are explicit precondition violations at a step. Latent failures satisfy all explicit preconditions but leave the final state in violation of a higher-level hazard or goal constraint. SIMMER further distinguishes reversible from irreversible latent failures by checking whether a recovery sequence exists. The executor encodes latent failure conditions including contamination spill, uncooked served, and leaving appliances on. Operationally, evaluation proceeds through a two-phase state-machine executor: a step-by-step simulation phase that validates syntax and preconditions and applies effects, followed by a post-execution audit that scans the final state for hazards and tests reversibility through minimal “wash” or “cook” actions (Lu et al., 12 Jun 2026).

The benchmark comprises 100 cooking scripts spanning 12 techniques, with 2–18 natural-language steps and averages of 26.6 actions and 31.5 objects per task. Six LLMs were evaluated with deterministic decoding at $\gtrsim2.5\times10^7\,\mathrm{cm\,s^{-1}}$6: GPT-5.4, Gemini 3 Flash, Claude Opus 4.6, Llama 3.3 70B, DeepSeek V3.2, and Qwen 3.5 27B. The principal metrics are

$\gtrsim2.5\times10^7\,\mathrm{cm\,s^{-1}}$7

together with LatentRate, IrreversibleRate, and AvgFailures. Across models, error-free plans remain under 17%, immediate failures affect 66–100% of plans, latent failures affect 29–56%, and irreversible failures appear in 20–45% of plans. Averaged over models, about 50% of latent failures arise from state propagation, about 30% from implicit precondition ignorance, and about 20% from post-goal neglect. A prompting strategy termed counterfactual foresight simulation, which requires explicit checking of hold state, location, container state, and appliance state before each action, reduces immediate failures by up to 80%, latent failures by up to 72%, and irreversible failures by up to 75% on the top two models (Lu et al., 12 Jun 2026).

4. simmer as a discrete-event simulation framework in R

In statistics and operations research, simmer is an R package for discrete-event simulation built around a process-oriented modeling paradigm. Its central abstraction is the trajectory, a list of Activities chained with the pipe operator %>% and executed by arrivals. The architecture combines an R API with a C++ simulation core implemented via Rcpp. Internally, a Simulator object maintains an event list as a C++ multiset ordered first by timestamp and then by fixed event-type priority, thereby resolving simultaneous events such as release-before-seize conflicts. Process classes include Generator, Arrival, Manager, and Task, and the system automatically monitors arrival, resource, and attribute changes for post-hoc analysis (Ucar et al., 2017).

At the R level, a typical workflow instantiates an environment with simmer("MySim"), defines one or more trajectories, attaches them to generators, adds resources, runs the model, and retrieves monitored data with get_mon_arrivals(env), get_mon_resources(env), or get_mon_attributes(env). The package supports reuse and composition of trajectories, along with dynamic parameters specified as R functions. In an M/M/1 setting the paper relates the monitored outputs to classical queueing formulas such as

$\gtrsim2.5\times10^7\,\mathrm{cm\,s^{-1}}$8

and reports that simmer without monitoring is roughly on par with SimPy and SimJulia, while monitoring adds only modest overhead. A batched generator improves throughput: as batch size increases from 1 to about 40, performance improves, and at $\gtrsim2.5\times10^7\,\mathrm{cm\,s^{-1}}$9 simmer becomes 1.6×–1.9× faster than SimPy on the same workload. The main performance limitation is repeated C++-to-R callbacks; replacing a fixed timeout(delay=1) by timeout(function() 1) slows a toy test from about 0.5 s to about 3.2 s for $E(k)=C_K\,\varepsilon^{2/3}\,k^{-5/3},$0 arrivals (Ucar et al., 2017).

A subsequent paper demonstrates simmer as a prototyping tool for communication-network research through three 5G-inspired use cases. The first models crosshauling of fronthaul and backhaul traffic through tandem XPFEs under no differentiation, strict priority without preemption, and strict priority with preemption. The second models mobile backhauling over FTTx, including a TDM-PON upstream channel of $E(k)=C_K\,\varepsilon^{2/3}\,k^{-5/3},$1, dynamic bandwidth allocation, and periodic RRH reservations. The third studies energy efficiency in massive IoT over LTE with $E(k)=C_K\,\varepsilon^{2/3}\,k^{-5/3},$2 NB-IoT devices, random access collisions, retransmissions, and synchronization windows. Reported scenario statistics include approximately 10 s runtime and $E(k)=C_K\,\varepsilon^{2/3}\,k^{-5/3},$3 maximum events for Scenario 1, approximately 150 s and $E(k)=C_K\,\varepsilon^{2/3}\,k^{-5/3},$4 events for Scenario 2, and approximately 150 s and $E(k)=C_K\,\varepsilon^{2/3}\,k^{-5/3},$5 events for Scenario 3, with total simulation code below 100 lines in all cases (Ucar et al., 2018).

5. SIMMER as a multimodal model for food image–recipe retrieval

In multimodal retrieval, SIMMER denotes Single Integrated Multimodal Model for Embedding Recipes. The model replaces the conventional dual-encoder architecture with a single unified encoder $E(k)=C_K\,\varepsilon^{2/3}\,k^{-5/3},$6 instantiated by VLM2Vec, so that both food images and structured recipe text are embedded by the same MLLM backbone. The recipe modality is explicitly structured around title, ingredients, and instructions, and the method uses separate prompt templates for image-as-query, image-as-candidate, recipe-as-query, and recipe-as-candidate. Training is contrastive in both directions with a unidirectional InfoNCE loss,

$E(k)=C_K\,\varepsilon^{2/3}\,k^{-5/3},$7

where $E(k)=C_K\,\varepsilon^{2/3}\,k^{-5/3},$8, cosine similarity is used, and the total loss is the sum of the two directional terms. A component-aware augmentation scheme further adds title-only, ingredients-only, and instructions-only recipe variants to the full recipe during training (Gomi et al., 17 Apr 2026).

The experiments use Recipe1M with approximately 238K paired image–recipe samples for training and 51K each for validation and test. Evaluation follows both 1k and 10k candidate settings, using medR and $E(k)=C_K\,\varepsilon^{2/3}\,k^{-5/3},$9. Fine-tuning uses Adam with learning rate $C_K\approx1.5$0, global batch size 128, LoRA rank 16, and 2,000 steps, reported as about 20 hr on 8×A6000 GPUs. The best model, VLM2Vec-V1-7B, improves 1k image-to-recipe $C_K\approx1.5$1 from 81.8% to 87.5% and 10k image-to-recipe $C_K\approx1.5$2 from 56.5% to 65.5% relative to the previous best method; for recipe-to-image, it reaches 85.1% at 1k and 61.5% at 10k. The paper also reports that V1-7B zero-shot performance at 1k image-to-recipe rises from 40.8% to 87.5% after fine-tuning, while component-aware augmentation leaves full-recipe performance at about 87.5% but improves partial-input robustness by up to about 3–4% $C_K\approx1.5$3 (Gomi et al., 17 Apr 2026).

The ablation analysis identifies recipe instructions as especially informative: instructions-only gives $C_K\approx1.5$4, ingredients plus instructions reaches 85.9%, title plus instructions 80.6%, and title plus ingredients 46.9% with augmentation. The qualitative examples emphasize that the model can distinguish subtle toppings, faint fillings, and procedural cues more effectively than dual-encoder baselines. A plausible implication is that SIMMER’s gains arise not only from contrastive fine-tuning but from exploiting a backbone whose hidden space is already jointly pretrained over vision and language (Gomi et al., 17 Apr 2026).

6. Simmer as a safe-exploration method in reinforcement learning

In reinforcement learning, Simmer stands for Safe policy IMproveMEnt for RL. The method starts from a constrained MDP with reward discount $C_K\approx1.5$5, safety discount $C_K\approx1.5$6, nonnegative safety cost $C_K\approx1.5$7, and safety budget $C_K\approx1.5$8. Its key device is a scalar safety state $C_K\approx1.5$9 that tracks remaining budget:

$u' \simeq 16\,\mathrm{km\,s^{-1}}$0

By induction,

$u' \simeq 16\,\mathrm{km\,s^{-1}}$1

so $u' \simeq 16\,\mathrm{km\,s^{-1}}$2 if and only if no violation has yet occurred. This lifts the original MDP to an augmented state space $u' \simeq 16\,\mathrm{km\,s^{-1}}$3 and converts sparse, unknown safety signals into an explicit budget-tracking variable (Sootla et al., 2022).

Training is organized epoch-wise around a scheduled initial budget $u' \simeq 16\,\mathrm{km\,s^{-1}}$4. At each epoch, rollouts are collected in the augmented MDP, an empirical safety statistic $u' \simeq 16\,\mathrm{km\,s^{-1}}$5 is computed, the base RL solver updates the policy under the empirical constraint, and a controller selects the next budget. Two scheduling mechanisms are defined. PI-Simmer uses a filtered tracking error with proportional, integral, and anti-windup terms, followed by clipping. Q-Simmer formulates budget selection as a small tabular MDP over discrete budget levels and actions $u' \simeq 16\,\mathrm{km\,s^{-1}}$6, with shaped rewards and $u' \simeq 16\,\mathrm{km\,s^{-1}}$7-greedy Q-learning. The framework is instantiated for both average-cost constraints, where one requires $u' \simeq 16\,\mathrm{km\,s^{-1}}$8, and probability-one constraints, where reward shaping applies a penalty $u' \simeq 16\,\mathrm{km\,s^{-1}}$9 whenever $L \simeq 200\,\mathrm{km}$0 and $L \simeq 200\,\mathrm{km}$1 (Sootla et al., 2022).

Empirical evaluation covers a safe pendulum swing-up task, a static point-goal environment, and multiple Safety-Gym tasks including PointPush1, PointGoal1, PointButton1, CarPush1, CarGoal1, and CarButton1, with baselines such as CPO, L-PPO, L-TRPO, PID-Lagrangian, LAMBDA, and PO-PPO. In the almost-sure setting, PI-Simmer and Q-Simmer reduce the number of violating rollouts by an order of magnitude over PO-PPO alone with only minor return loss. In the average-cost setting, simply adding the safety state stabilizes L-PPO, eliminates large cost spikes, and yields near-zero average violations; Simmer-L-PPO further reduces empirical cost-rate without degrading return. On Safety-Gym, Simmer combined with PID-Lagrangian outperforms baselines on cost-rate and matches or exceeds return, with especially pronounced gains in PointPush1 and CarPush1 (Sootla et al., 2022).

7. SImMER as an astronomical image-reduction pipeline

In observational astronomy, SImMER is an open-source Python pipeline for reducing and analyzing images of stellar point sources. The first public version supports the ShARCS camera on the Shane 3-m telescope and the PHARO camera on the Hale 5.1-m telescope, and it provides dark subtraction, flat-fielding, sky subtraction, image registration, FWHM measurement, contrast-curve calculation, and generation of tables and plots. The codebase is modular, pip- and conda-installable, and organized into modules such as darks.py, flats.py, sky.py, registration.py, contrast.py, plotting.py, insts/, and run_night.py (Savel et al., 2022).

A core component is image registration, for which the pipeline exposes several modes. Quick-look registration uses peak_local_max with a binary search over detection threshold and is reported at about 12 ms per frame. Empirical-PSF registration models the on-axis PSF as a rotated elliptical 2D Gaussian and uses emcee together with DAOFind for precise centroiding. Saturated-star registration follows a rotational-symmetry criterion,

$L \simeq 200\,\mathrm{km}$2

searching for the center that minimizes the residual. The pipeline also includes a multi-source helper for comparable-brightness binaries, FWHM estimation from either Gaussian parameters or radial profiles, and a contrast-curve routine based on annuli, angular wedges, injected 2D Gaussians, and the magnitude-contrast conversion

$L \simeq 200\,\mathrm{km}$3

Reported validation includes contrast curves agreeing with the original IDL-based code within $L \simeq 200\,\mathrm{km}$4 over $L \simeq 200\,\mathrm{km}$5–$L \simeq 200\,\mathrm{km}$6 and median FWHM of about 3 px ($L \simeq 200\,\mathrm{km}$7) for ShARCS AO-on K-band data (Savel et al., 2022).

The mixed-case styling is therefore not incidental: SImMER is a specific astronomical software package, distinct from the lower-case R package simmer and from the machine-learning systems named SIMMER or Simmer. Across the cited literature, the shared label marks a convergence of naming rather than a convergence of method.