CampFIRE: Multidisciplinary Research Insights

• CampFIRE is a term applied across domains to denote distinct, context-specific constructs, from masked autoencoders in fluorescence microscopy to virtual reality thermal interfaces.
• In solar physics and cosmological hydrodynamics, CampFIRE describes phenomena like EUV brightenings and high-resolution galaxy formation models that bridge large-scale statistics with zoom-in simulations.
• CampFIRE underpins advances in machine learning and multi-agent reinforcement learning by enabling structured sparse CNN training and facilitating emergent exchange protocols among agents.

Searching arXiv for papers explicitly using the term “CampFIRE” across domains to ground the encyclopedia entry. CampFIRE is a term used in multiple research domains to denote distinct systems, models, and phenomena. In recent arXiv literature, it has referred to a channel-agnostic masked autoencoder for fluorescence microscopy, a virtual campfire interface for thermal feedback in virtual reality, a structured sparse training method for convolutional neural networks, a family of quiet-Sun extreme-ultraviolet brightenings observed by Solar Orbiter and their associated detection and magnetic-interpretation studies, a campfire-centered mechanism for emergent exchange in multi-agent reinforcement learning, and a high-resolution cosmological simulation suite within the FIRE project (Hurry et al., 24 Mar 2025, Ito et al., 2024, Gamboa et al., 2020, Chen et al., 2021, Alipour et al., 2022, Panesar et al., 2021, Kahil et al., 2022, Garbus et al., 2023, Samuel et al., 22 May 2026). The shared label does not denote a unified formalism; rather, it names domain-specific constructs whose commonality is primarily lexical. This suggests that any technical use of the term requires immediate contextual disambiguation.

1. CampFIRE in fluorescence microscopy

In fluorescence microscopy and high-content screening, CampFIRE denotes a channel-agnostic masked autoencoder introduced for out-of-distribution evaluation on JUMP-CP (Hurry et al., 24 Mar 2025). The model is explicitly designed to ingest arbitrary subsets and permutations of fluorescent channels and to learn reusable single-cell representations from cell-centred tiles of size $112 \times 112$ pixels. Training uses three Cell Painting channels—Nucleus (Nu), Actin + Golgi + membrane (Ac), and Mitochondria (M)—while Endoplasmic reticulum (ER) and nucleolus + cytoplasmic RNA (cyRNA) are held out for out-of-distribution channel evaluation (Hurry et al., 24 Mar 2025).

Architecturally, CampFIRE uses a ViT-style masked autoencoder with channel-agnostic patch embedding, a shared encoder, and a shared decoder with content-derived channel embeddings (Hurry et al., 24 Mar 2025). The patch embedding applies a 3D convolution with kernel and stride $(1,P,P)$, with $P=14$, and the same convolution is applied to each channel. A masking ratio $p_m = 0.8$ is used, and the decoder reconstructs the full multi-channel image using sinusoidal positional embeddings, Rotary Positional Embedding, and channel embeddings derived from average patch embeddings within each channel (Hurry et al., 24 Mar 2025). The training objective combines spatial mean-squared error with high-pass and low-pass Fourier-domain reconstruction losses: $$\mathcal{L} = - \frac{1}{|\mathcal{D}|}\sum_{i=1}^{|\mathcal{D}|} \left[ \lambda_{s} \mathrm{MSE}(\hat{y}_i, y_{i}) + \lambda_{\ell} \mathrm{MSE}(\mathcal{F}[\hat{y}_i| \ell], \mathcal{F}[y_{i}|\ell]) + \lambda_{h} \mathrm{MSE}(\mathcal{F}[\hat{y}_i| h], \mathcal{F}[y_{i}|h]) \right].$$ At inference time, the decoder is discarded and the single-cell embedding is the mean of the encoder’s final patch embeddings over all spatial locations and channels (Hurry et al., 24 Mar 2025).

A central contribution of the paper is a controlled OOD evaluation scheme on JUMP-CP that isolates distribution shifts in plates, perturbagens, fluorescent channels, and cell types (Hurry et al., 24 Mar 2025). The training setup uses 20 TARGET2 plates and 237 COMPOUND plates, with 5 whole TARGET2 plates held out as OOD plates and 60 compounds held out entirely from training as OOD compounds (Hurry et al., 24 Mar 2025). Downstream evaluation uses linear probing on single-cell embeddings for 1-of-9 control classification and 1-of-60 held-out compound classification, with accuracy averaged over splits or folds (Hurry et al., 24 Mar 2025). Across all four compound prediction tasks, the authors report that CampFIRE outperformed ImageNet-pretrained DinoViT baselines (Hurry et al., 24 Mar 2025).

The held-out-channel results quantify the model’s channel transfer behavior. For the in-distribution channel set, CampFIRE achieves $0.460 \pm 0.016$ on controls, ID plates and $0.220 \pm 0.002$ on held-out compounds, OOD plates; with the OOD channel set $\{\mathrm{Nu}, \mathrm{ER}, \mathrm{cyRNA}\}$, the corresponding values are $0.44 \pm 0.01$ and $0.199 \pm 0.003$ (Hurry et al., 24 Mar 2025). In the hardest condition—held-out compounds on OOD plates with OOD channels—CampFIRE exceeds DinoViT-S8, which reaches $(1,P,P)$0 (Hurry et al., 24 Mar 2025). The paper also reports transfer to a macrophage CRISPR screen using a frozen CampFIRE encoder and a 2-layer MLP trained with triplet loss, with $(1,P,P)$1-score analysis indicating better separation of biologically meaningful controls than DinoViT-S8 (Hurry et al., 24 Mar 2025).

The microscopy CampFIRE is therefore situated at the intersection of masked autoencoders, microscopy foundation models, and controlled distribution-shift benchmarking. Its specific novelty lies in treating channels symmetrically in the encoder, using a content-derived channel embedding in a shared decoder, and combining this with random channel subsampling during training (Hurry et al., 24 Mar 2025).

2. CampFIRE in solar physics

In solar physics, “campfires” are small-scale extreme-ultraviolet brightenings in the quiet-Sun corona observed by the Extreme Ultraviolet Imager on Solar Orbiter (Chen et al., 2021, Alipour et al., 2022, Panesar et al., 2021, Kahil et al., 2022, Zhukov et al., 2021). They are observed in the 174 Å passband, dominated by Fe IX–Fe X emission near $(1,P,P)$2 MK, and were initially characterized as loop-like or flame-like structures with typical sizes of roughly $(1,P,P)$3 km to $(1,P,P)$4 km and lifetimes often below a few minutes (Alipour et al., 2022, Zhukov et al., 2021). Stereoscopic reconstruction places them at heights between $(1,P,P)$5 km and $(1,P,P)$6 km above the photosphere, suggesting that they are low-lying loops in the transition region and low corona rather than high coronal structures (Zhukov et al., 2021).

A numerical interpretation is given by the MURaM-based study of quiet-Sun coronal brightenings, which identifies seven synthetic events in a $(1,P,P)$7 domain evolved with 3D radiation MHD (Chen et al., 2021). Those simulated transients have lifetimes of about 2 minutes, lengths of $(1,P,P)$8–4 Mm, heights of $(1,P,P)$9–5 Mm, and temperatures above $P=14$0 MK, with one representative event reaching peak count rate $P=14$1 in synthetic EUI 174 Å and total detected photons $P=14$2 (Chen et al., 2021). Most are interpreted as component reconnection between interacting coronal loop bundles; one is associated with the untwisting of a highly twisted small flux rope (Chen et al., 2021). The same study estimates a dissipated energy of $P=14$3 per event and an average energy flux of $P=14$4, comparable to canonical quiet-Sun coronal heating requirements (Chen et al., 2021).

Observationally, the magnetic interpretation has focused on flux cancellation and small-scale loop systems. A study of 52 randomly selected campfires reports that most are rooted at the edges of photospheric magnetic network lanes, that 40 of 52 events show clear flux cancellation at rates of order $P=14$5, and that the average campfire length and width are $P=14$6 km and $P=14$7 km, respectively (Panesar et al., 2021). Their energies are estimated to lie in the range $P=14$8, and many are preceded by cool-plasma structures analogous to minifilaments, suggesting that magnetic flux cancellation is fundamental to the formation and triggering of most campfires (Panesar et al., 2021). A complementary study using Solar Orbiter/PHI line-of-sight magnetograms finds that 71% of 38 isolated events are confined between bipolar magnetic features exhibiting signatures of flux cancellation, while about a quarter are not clearly associated with such photospheric activity, implying that other heating mechanisms may also operate (Kahil et al., 2022).

The statistical characterization of campfires has been advanced by automatic detection methods. A Zernike-moment and support-vector-machine pipeline applied to 50 HRIEUV 174 Å images detects 8678 campfires with length scales between 400 km and 4000 km, yielding a birthrate of $P=14$9 (Alipour et al., 2022). Of these, 3300 are seen in at least two frames and 23 last longer than the 245 s observing window (Alipour et al., 2022). About 80% are formed within 5 Mm of supergranular boundaries, 27% are found in coronal bright points, and the probability distribution functions for total intensity, peak intensity, projected area, and duration follow power laws with indices between 2 and 3 (Alipour et al., 2022). This has been interpreted as a possible signature of self-organization or self-organized criticality in the campfire formation process (Alipour et al., 2022).

The solar-physics usage of CampFIRE is therefore not a single instrument or algorithm but a research program centered on quiet-Sun EUV brightenings, their 3D geometry, their magnetic drivers, and their possible role in coronal heating (Chen et al., 2021, Alipour et al., 2022, Panesar et al., 2021, Kahil et al., 2022, Zhukov et al., 2021).

3. CampFIRE in virtual reality thermal interfaces

In haptics and virtual reality, CampFIRE appears as a broader concept associated with a virtual campfire experience, most explicitly in the context of “HeatFlicker,” a compact virtual campfire device that recreates flickering heat using non-contact thermal radiation and asymmetric vibration (Ito et al., 2024). Users insert both forearms into a covered box while an LCD displays virtual hands near a virtual fire, visible-light LEDs radiate heat from slanted side walls, a camera tracks the hands, and a handheld vibrator provides asymmetric acceleration patterns that induce a directional pulling illusion (Ito et al., 2024).

The central phenomenon is a thermal illusion of moving heat: users feel as if heat is shifting position or direction even though the physical heat source is stationary (Ito et al., 2024). Preliminary observations with three participants, including the first author, indicate that when the direction of asymmetric vibration changes, the perceived location of the hottest region or the perceived direction of heat flow changes in synchrony with the pulling sensation (Ito et al., 2024). The work is grounded in thermal referral, apparent thermal motion, and asymmetric-vibration pulling illusions, but it does not introduce a formal psychophysical model or explicit signal equations (Ito et al., 2024).

The physical system includes multiple visible-light LEDs, a fan for thermal management, a speaker for crackling fire sounds, a black curtain to occlude the hardware, an LCD panel for the visual scene, and a camera behind the display for hand tracking (Ito et al., 2024). Control is described conceptually rather than parametrically: LEDs are driven via pulse-width modulation for time-varying heat, the vibrator uses asymmetric vibration stimuli from earlier work, and the direction of the pulling illusion is periodically reversed to create a flickering effect (Ito et al., 2024). Exact values for vibration frequency, amplitude, duty cycle, and the mapping from visual fire intensity to actuator signals are not specified (Ito et al., 2024).

The paper explicitly frames the system as a demonstration rather than a full user study. No quantitative evaluations, detection rates, or psychophysical protocols are reported, and future work is identified as measuring the relationship between pulling direction and perceived direction of thermal movement and quantifying the intensity of the flickering thermal illusion (Ito et al., 2024). In this usage, CampFIRE denotes a multisensory virtual campfire concept in which apparent thermal movement is generated through cross-modal interaction rather than physical motion of the heat source (Ito et al., 2024).

4. CampFIRE in machine learning for sparse CNN training

In deep learning systems research, Campfire denotes a compressible, regularization-free, structured sparse training method for convolutional neural networks (Gamboa et al., 2020). The method aims to produce sparse weight matrices during training using gradual magnitude-based pruning, while avoiding explicit sparsity regularization and preserving accelerator-friendly structure (Gamboa et al., 2020). It is motivated by the hardware inefficiency of unstructured sparsity and by the overhead introduced by regularization-based or dynamically rewired sparse training methods (Gamboa et al., 2020).

Campfire organizes pruning in a pruning era, during which sparsity is gradually increased, after which the mask is fixed and zeroed weights remain zero for the remainder of training (Gamboa et al., 2020). Pruning is structured at several granularities: window pruning inside each $p_m = 0.8$0 kernel, CK pruning at the whole-kernel level across channel–filter coordinates, a combined scheme applying window and CK pruning together, and block plus fine-grained pruning in fully connected layers (Gamboa et al., 2020). The sparsity schedule is: $p_m = 0.8$1 with target sparsity $p_m = 0.8$2, initial sparsity $p_m = 0.8$3, pruning start epoch $p_m = 0.8$4, pruning-era length $p_m = 0.8$5, and exponent $p_m = 0.8$6 in the reported experiments (Gamboa et al., 2020).

On ImageNet with ResNet-50 v1.5, the dense baseline reaches 76.29% top-1 accuracy (Gamboa et al., 2020). With CK pruning starting at epoch 40, the model achieves 75.33% at 60% sparsity, 74.92% at 70% sparsity, and 74.16% at 80% sparsity (Gamboa et al., 2020). Combined pruning yields 75.38% at 60% sparsity and 75.07% at 70% sparsity (Gamboa et al., 2020). With a longer pruning era from epoch 40 to 70, CK and combined pruning at 60% sparsity reach 75.52% and 75.56%, respectively, while remaining within about one percentage point of the dense baseline (Gamboa et al., 2020). The paper therefore states that with 70% target sparsity, over 75% top-1 accuracy is achievable (Gamboa et al., 2020).

The method is also evaluated under FGSM adversarial attacks. For ResNet-50 v1.5 at $p_m = 0.8$7, the dense model achieves 30.47% top-1 accuracy, while sparse Campfire variants at 60–70% sparsity remain close, typically around 28.3–29.2%, indicating that structured sparse training does not introduce severe robustness degradation beyond the clean-accuracy loss (Gamboa et al., 2020). The main systems implication is that Campfire yields fixed, structured sparsity patterns that are directly amenable to compression and zero-skipping on hardware accelerators (Gamboa et al., 2020).

5. CampFIRE in multi-agent reinforcement learning

In multi-agent reinforcement learning, CampFIRE refers to the campfire mechanism in The Trading Game, a foraging environment in which an always-safe central region and a harsh day/night cycle induce repeated congregation and downtime (Garbus et al., 2023). The environment contains four agents, two resource types—fruits and greens—and a central campfire consisting of a $p_m = 0.8$8 always-safe core surrounded by a $p_m = 0.8$9 halo of dim light (Garbus et al., 2023). Outside the campfire region, a nighttime light penalty applies when $$\mathcal{L} = - \frac{1}{|\mathcal{D}|}\sum_{i=1}^{|\mathcal{D}|} \left[ \lambda_{s} \mathrm{MSE}(\hat{y}_i, y_{i}) + \lambda_{\ell} \mathrm{MSE}(\mathcal{F}[\hat{y}_i| \ell], \mathcal{F}[y_{i}|\ell]) + \lambda_{h} \mathrm{MSE}(\mathcal{F}[\hat{y}_i| h], \mathcal{F}[y_{i}|h]) \right].$$0, with default penalty parameter $$\mathcal{L} = - \frac{1}{|\mathcal{D}|}\sum_{i=1}^{|\mathcal{D}|} \left[ \lambda_{s} \mathrm{MSE}(\hat{y}_i, y_{i}) + \lambda_{\ell} \mathrm{MSE}(\mathcal{F}[\hat{y}_i| \ell], \mathcal{F}[y_{i}|\ell]) + \lambda_{h} \mathrm{MSE}(\mathcal{F}[\hat{y}_i| h], \mathcal{F}[y_{i}|h]) \right].$$1 and per-step light reward $$\mathcal{L} = - \frac{1}{|\mathcal{D}|}\sum_{i=1}^{|\mathcal{D}|} \left[ \lambda_{s} \mathrm{MSE}(\hat{y}_i, y_{i}) + \lambda_{\ell} \mathrm{MSE}(\mathcal{F}[\hat{y}_i| \ell], \mathcal{F}[y_{i}|\ell]) + \lambda_{h} \mathrm{MSE}(\mathcal{F}[\hat{y}_i| h], \mathcal{F}[y_{i}|h]) \right].$$2 (Garbus et al., 2023).

Each day lasts 24 steps and each night also lasts 24 steps, with episodes lasting 180 steps (Garbus et al., 2023). Agents can move, pick up all fruits or all greens on the current cell, or drop 0.5 units of fruit or greens (Garbus et al., 2023). At each step, they automatically consume 0.1 units of the resources they carry; consuming both resource types yields reward 1.0, whereas consuming only one resource type yields 0.1 (Garbus et al., 2023). This reward structure makes exchange valuable, but no explicit trade or communication action exists (Garbus et al., 2023).

Under PPO with separate policies for each agent, the campfire’s enforced congregation enables the emergence of a reciprocal exchange protocol called DROP–SWAP (Garbus et al., 2023). Agents form pairs, stand on adjacent cells, both drop 0.5 units of their respective resources, then move to pick up the partner’s drop, repeating this pattern over multiple night steps (Garbus et al., 2023). In the base setting with five resources per patch, each pair exchanges about 1.5 units per agent per night, or about 9 fruits and 9 greens across three nights and two pairs per episode (Garbus et al., 2023). Ablation results show that with a weaker night penalty $$\mathcal{L} = - \frac{1}{|\mathcal{D}|}\sum_{i=1}^{|\mathcal{D}|} \left[ \lambda_{s} \mathrm{MSE}(\hat{y}_i, y_{i}) + \lambda_{\ell} \mathrm{MSE}(\mathcal{F}[\hat{y}_i| \ell], \mathcal{F}[y_{i}|\ell]) + \lambda_{h} \mathrm{MSE}(\mathcal{F}[\hat{y}_i| h], \mathcal{F}[y_{i}|h]) \right].$$3, exchange is less reciprocal, and with no night penalty $$\mathcal{L} = - \frac{1}{|\mathcal{D}|}\sum_{i=1}^{|\mathcal{D}|} \left[ \lambda_{s} \mathrm{MSE}(\hat{y}_i, y_{i}) + \lambda_{\ell} \mathrm{MSE}(\mathcal{F}[\hat{y}_i| \ell], \mathcal{F}[y_{i}|\ell]) + \lambda_{h} \mathrm{MSE}(\mathcal{F}[\hat{y}_i| h], \mathcal{F}[y_{i}|h]) \right].$$4, stable reciprocal exchange does not emerge even though one-way giving may occur (Garbus et al., 2023).

The same framework also yields emergent anti-cheating behavior and a tolerated-theft-like pattern. Agents successfully rescind exchange offers when a direct partner defects, but they are more variably robust to third-party interference (Garbus et al., 2023). In one reported trial, one agent repeatedly drops extra greens to distract an interfering third agent so that a main exchange with another partner can proceed; when this offering is prevented, the interfering agent positionally blocks the exchange (Garbus et al., 2023). The paper interprets this as analogous to tolerated theft, despite the absence of punishment, combat, or larceny mechanics (Garbus et al., 2023). In this usage, CampFIRE is therefore an environmental design that creates the social conditions under which embodied exchange protocols can emerge.

6. CampFIRE in cosmological hydrodynamics

In computational astrophysics, CampFIRE designates the high-resolution zoom-in component of the BonFIRE + CampFIRE simulation strategy for galaxy formation at cosmic dawn within the FIRE project (Samuel et al., 22 May 2026). Two CampFIRE runs re-simulate the same $$\mathcal{L} = - \frac{1}{|\mathcal{D}|}\sum_{i=1}^{|\mathcal{D}|} \left[ \lambda_{s} \mathrm{MSE}(\hat{y}_i, y_{i}) + \lambda_{\ell} \mathrm{MSE}(\mathcal{F}[\hat{y}_i| \ell], \mathcal{F}[y_{i}|\ell]) + \lambda_{h} \mathrm{MSE}(\mathcal{F}[\hat{y}_i| h], \mathcal{F}[y_{i}|h]) \right].$$5 overdense subvolume of a larger BonFIRE box: CampFIRE-6k with $$\mathcal{L} = - \frac{1}{|\mathcal{D}|}\sum_{i=1}^{|\mathcal{D}|} \left[ \lambda_{s} \mathrm{MSE}(\hat{y}_i, y_{i}) + \lambda_{\ell} \mathrm{MSE}(\mathcal{F}[\hat{y}_i| \ell], \mathcal{F}[y_{i}|\ell]) + \lambda_{h} \mathrm{MSE}(\mathcal{F}[\hat{y}_i| h], \mathcal{F}[y_{i}|h]) \right].$$6 and CampFIRE-800 with $$\mathcal{L} = - \frac{1}{|\mathcal{D}|}\sum_{i=1}^{|\mathcal{D}|} \left[ \lambda_{s} \mathrm{MSE}(\hat{y}_i, y_{i}) + \lambda_{\ell} \mathrm{MSE}(\mathcal{F}[\hat{y}_i| \ell], \mathcal{F}[y_{i}|\ell]) + \lambda_{h} \mathrm{MSE}(\mathcal{F}[\hat{y}_i| h], \mathcal{F}[y_{i}|h]) \right].$$7 (Samuel et al., 22 May 2026). Both use FIRE-3 physics in GIZMO with meshless finite mass hydrodynamics and Planck-2018 cosmology (Samuel et al., 22 May 2026).

The scientific purpose is to combine the large statistics of BonFIRE with the higher resolution of CampFIRE, thereby predicting galaxy properties over $$\mathcal{L} = - \frac{1}{|\mathcal{D}|}\sum_{i=1}^{|\mathcal{D}|} \left[ \lambda_{s} \mathrm{MSE}(\hat{y}_i, y_{i}) + \lambda_{\ell} \mathrm{MSE}(\mathcal{F}[\hat{y}_i| \ell], \mathcal{F}[y_{i}|\ell]) + \lambda_{h} \mathrm{MSE}(\mathcal{F}[\hat{y}_i| h], \mathcal{F}[y_{i}|h]) \right].$$8 at $$\mathcal{L} = - \frac{1}{|\mathcal{D}|}\sum_{i=1}^{|\mathcal{D}|} \left[ \lambda_{s} \mathrm{MSE}(\hat{y}_i, y_{i}) + \lambda_{\ell} \mathrm{MSE}(\mathcal{F}[\hat{y}_i| \ell], \mathcal{F}[y_{i}|\ell]) + \lambda_{h} \mathrm{MSE}(\mathcal{F}[\hat{y}_i| h], \mathcal{F}[y_{i}|h]) \right].$$9 (Samuel et al., 22 May 2026). The CampFIRE region is overdense, with $0.460 \pm 0.016$0, and is centered on a halo reaching $0.460 \pm 0.016$1 at $0.460 \pm 0.016$2 (Samuel et al., 22 May 2026). Snapshots are output every 15 Myr, and the simulations explicitly resolve clustered star formation and feedback on sub-10-pc scales in the dense interstellar medium (Samuel et al., 22 May 2026).

A notable result is that galaxy formation emerges through clustered, bursty star formation, with halo-scale star formation efficiencies

$0.460 \pm 0.016$3

reaching 10–30% in high-mass halos and exceeding 1% in a subset of low-mass halos (Samuel et al., 22 May 2026). The latter systems are identified as ultra-compact galaxies with typical stellar half-mass radii $0.460 \pm 0.016$4, stellar masses $0.460 \pm 0.016$5, and age dispersions $0.460 \pm 0.016$6 (Samuel et al., 22 May 2026). The combined BonFIRE+CampFIRE UV luminosity function at $0.460 \pm 0.016$7 is reported to be in broad agreement with observations for $0.460 \pm 0.016$8, with a faint-end turnover near $0.460 \pm 0.016$9, although the simulations slightly overpredict the abundance of brighter galaxies (Samuel et al., 22 May 2026).

The paper also introduces a simple Pop III model with critical metallicity $0.220 \pm 0.002$0, implemented by boosting massive-star feedback below that metallicity to mimic a top-heavy IMF (Samuel et al., 22 May 2026). The resulting Pop III star-formation-rate density is in broad agreement with other simulation and semi-analytic estimates, while Pop III galaxies remain faint and rare relative to the observed bright JWST population (Samuel et al., 22 May 2026). In this context, CampFIRE is not a campfire metaphor but the high-resolution half of a cosmological simulation suite designed to bridge volume and resolution at early times (Samuel et al., 22 May 2026).