Chameleon: Adaptive Systems Across Disciplines

• Chameleon is a cross-disciplinary term defining adaptive systems that optimize performance under fluctuating conditions across various research fields.
• It spans applications from machine learning compilers and on-chip accelerators to multimodal models and privacy-preserving techniques, demonstrating measurable efficiency improvements.
• The term also underlies advancements in theoretical physics and robotics, illustrating controlled variability and context-aware behavior in dynamic experimental environments.

Chameleon is a recurrent research name used across machine learning systems, multimodal modeling, hardware acceleration, privacy engineering, wireless sensing, robotics, and theoretical physics. In the literature considered here, it denotes an optimizing compiler integrated into TVM, an early-fusion mixed-modal foundation model, a MatMul-free accelerator for on-chip few-shot and continual learning, privacy and adversarial systems for visual inputs, behavioral and forensic datasets, a 5G mmWave ISAC framework, a long-term LiDAR change-detection system spelled “Chamelion,” and a class of screened scalar-field theories in gravitation and dark-energy research (Ahn et al., 2020, Team, 2024, Blanken et al., 30 May 2025, Chow et al., 2024, Harry et al., 21 Jan 2026, Ratnakar et al., 19 Oct 2025, Gao et al., 18 Sep 2025, Jang et al., 9 Feb 2026, Khoury, 2013).

1. Adaptive computation and efficient machine-learning systems

One influential use of the name appears in “Chameleon: Adaptive Code Optimization for Expedited Deep Neural Network Compilation,” an optimizing compiler framework that accelerates deep neural network code generation by learning to adapt quickly to previously unseen schedule design spaces and by focusing costly hardware measurements on representative points (Ahn et al., 2020). Integrated into TVM, it formulates compilation as optimization over schedule knobs $\Theta$, applies Adaptive Exploration with a PPO actor–critic agent guided by an XGBoost cost model, and then applies Adaptive Sampling with K-means centroids, threshold-based Swift Meta-Search over $K\in[8,64]$ with $\text{Threshold}=2.5$, and sample synthesis based on the mode of non-invalid knob values. The detailed end-to-end results report AlexNet optimization time decreasing from 4.31 h to 1.20 h, VGG-16 from 11.18 h to 1.95 h, and ResNet-18 from 9.13 h to 2.13 h, corresponding to an average speedup of $\approx 4.54\times$, while average inference time improves by $\approx 4.5\%$; the abstract reports a 4.45x speed up in optimization time over AutoTVM and a 5.6% improvement in inference time (Ahn et al., 2020).

At the hardware level, “Chameleon: A MatMul-Free Temporal Convolutional Network Accelerator for End-to-End Few-Shot and Continual Learning from Sequential Data” names a unified learning and inference architecture for extreme-edge deployment (Blanken et al., 30 May 2025). The system reframes prototypical networks so that prototype-based distance classification becomes an equivalent fully connected layer, adds only $\sim 0.5\%$ area overhead to the inference logic, uses temporal convolutional networks with a greedy, dilation-aware execution and layer-wise FIFO activation storage, and implements a dual-mode 16×16 PE array with 4-bit signed log2 weights and shift-accumulate compute. Fabricated in TSMC 40-nm LP CMOS, it has a die area of 1.25 mm² and 71 kB of on-chip memory. The paper reports Omniglot few-shot learning accuracies of 96.8% ±1.6 for 5-way 1-shot and 98.8% ±0.5 for 5-way 5-shot, continual-learning final accuracy of 82.2% ±0.4 for learning 250 classes with 10 shots, 93.3% accuracy on 12-class Google Speech Commands at 3.1 $\mu$W in 4×4 mode, and 76.8 GOPS at approximately 6 TOPS/W in 16×16 mode (Blanken et al., 30 May 2025).

A third systems use appears in “Chameleon: Taming Dynamic Operator Sequences for Memory-Intensive LLM Training,” a swap-based memory optimization system designed for PyTorch-like Eager Mode, where operator sequences vary across iterations (Wang et al., 14 Sep 2025). Its pipeline combines a lightweight online profiler, a logical-layer policy generator that avoids costly per-operator timing, and an optimized executor with multi-feature fuzzy matching and a custom recordStream mechanism. The reported profiling overhead is 0.9% in lightweight mode and 34.6% in detailed mode, versus 219.7% for PyTorch’s built-in profiler, corresponding to an 84.25% reduction. The system enables training models up to 4× larger than device HBM and improves performance by up to 38.94% compared to recomputation or high-degree parallelism (Wang et al., 14 Sep 2025).

2. Multimodal modeling, personalization, and behavioral evaluation

In foundation modeling, “Chameleon: Mixed-Modal Early-Fusion Foundation Models” denotes a family of early-fusion, token-based, mixed-modal models that unify text, images, and code in a single autoregressive transformer (Team, 2024). The architecture largely follows Llama-2, uses RMSNorm, SwiGLU, RoPE, and, in the 34B configuration, grouped-query attention. Images are encoded into 1,024 tokens from a codebook of size 8,192, text uses a 65,536-token SentencePiece BPE vocabulary, and both the 7B and 34B variants use a 4,096-token context window. Mixed-modal stability is addressed with query–key normalization, Swin-block-style norm reordering, and a $z$-loss term $\lambda (\log Z)^2$ with $\lambda=10^{-5}$. The paper reports state-of-the-art image captioning performance, strong VQA, competitive text-only results against popular LLM baselines, and human-evaluation wins on long-form mixed-modal generation, including 55.2% “fully fulfills” task performance versus 37.6% for Gemini+ and 44.7% for GPT-4V+ (Team, 2024).

The name also designates a dataset for contextual psychological profiling in NLP. “Beyond Fixed Psychological Personas: State Beats Trait, but LLMs are State-Blind” introduces Chameleon as a dataset of 5,001 contextual psychological profiles from 1,667 Reddit users across 645 subreddits, with each profile spanning 26 dimensions drawn from Big Five personality, Schwartz Values, Self-Determination Theory, and DOSPERT risk attitudes (Harry et al., 21 Jan 2026). The dataset operationalizes Latent State–Trait theory through the decomposition $K\in[8,64]$0. Empirically, mean ICC is 0.27 for fused profiles, implying that approximately 72–74% of variance is within-person/contextual and only about 26–28% is between-person/stable trait. The study further reports that GPT-4o, Llama-3.1-8B, and Qwen2.5-14B are state-blind in generation, while reward models are state-aware but inconsistent: ArmoRM-8B rewards profiles such as Distressed (+0.76), whereas DeBERTa-RM and Skywork-RM-8B penalize them (−1.08 and −1.12, respectively) (Harry et al., 21 Jan 2026).

A related benchmark use appears in “The Chameleon Nature of LLMs: Quantifying Multi-Turn Stance Instability in Search-Enabled LLMs,” which studies “chameleon behavior” as stance shifts under contradictory multi-turn probes (Ratnakar et al., 19 Oct 2025). The Chameleon Benchmark Dataset contains 17,770 question–answer pairs across 1,180 conversations and 12 controversial domains. The paper introduces the Chameleon Score, an RMS aggregation of normalized stance-change frequency, confidence during stance changes, and Source Re-use Rate. Evaluations of Llama-4-Maverick, GPT-4o-mini, and Gemini-2.5-Flash report Chameleon Scores in the range 0.391–0.511, with GPT-4o-mini worst, and temperature variance below 0.004, indicating that the effect is not a sampling artifact (Ratnakar et al., 19 Oct 2025).

A fourth use, spelled in all caps in the paper title, is “CHAMELEON: On the Scene Diversity and Domain Variety of AI-Generated Videos Detection,” but a distinct personalization method also adopts the name. “Personalize Your LLM: Fake it then Align it” proposes CHAMELEON as a scalable personalization pipeline based on self-generated personal preference data and representation editing (Zhang et al., 2 Mar 2025). It selects representative history with PCA, generates personalized and neutral insights, synthesizes preference pairs, identifies personalized directions with SVD and non-personalized directions with CCS, and edits MLP-output representations at inference time. The paper reports that CHAMELEON improves instruction-tuned models and outperforms two personalization baselines by an average of 40% across two model architectures on LaMP tasks, while also generalizing to unseen users (Zhang et al., 2 Mar 2025).

3. Privacy, adversarial robustness, and media forensics

In privacy-preserving vision, “Chameleon” names a personalized facial obfuscation system built around the P3-Mask (Chow et al., 2024). The method learns a single user-specific mask from several images of one person via cross-image optimization rather than per-image perturbation, adds a perceptibility optimization based on an SSIM hinge penalty, and improves transferability through focal diversity-optimized ensemble learning over face-recognition models. Protection Success Rate is defined as $K\in[8,64]$1. Reported results show online protection time of 0.0076 seconds per face, versus 105.12 seconds for TIP-IM and Fawkes, and an SSIM example of 0.9493 for Chameleon compared with 0.8839 for OPOM, 0.8850 for TIP-IM, and 0.9612 for Fawkes. The paper also demonstrates authorized de-obfuscation: for M. Baccarin, FR accuracy averages 6.82% after protection, returns to 100% after correct unmasking, and drops to 1.14% with an incorrect mask (Chow et al., 2024).

A security-oriented use appears in “Chameleon: Adaptive Adversarial Agents for Scaling-Based Visual Prompt Injection in Multimodal AI Systems,” where Chameleon is an adaptive, feedback-driven attack framework against VLM preprocessing (Zeeshan et al., 4 Dec 2025). The attack exploits downscaling, modeled as convolution with a resampling kernel followed by subsampling, so that high-frequency perturbations alias into low-frequency, semantically meaningful patterns after resizing. The system queries Gemini-2.5-Flash iteratively and optimizes a reward $K\in[8,64]$2 over success, perceptual distance, and confidence. The abstract reports an Attack Success Rate of 84.5% across varying scaling factors, compared with 32.1% for static baselines, and a reduction of downstream decision-making accuracy by over 45% in multi-step tasks. Strategy-level results report 87.0% ASR for hill-climbing, 91.0% for the genetic algorithm, and mean normalized $K\in[8,64]$3 visual distances of 0.0847 and 0.0693, respectively (Zeeshan et al., 4 Dec 2025).

Media forensics supplies another dataset use. “Chameleon: On the Scene Diversity and Domain Variety of AI-Generated Videos Detection” constructs a benchmark of 1,200 five-second video segments: 600 real and 600 AI-generated, spanning News, Speech, and Recommendation domains and generated using Runway Gen 3, Kling, and Jimeng (Zeng et al., 9 Mar 2025). The dataset emphasizes scene switches, dynamic perspective changes, human actions, and environmental generation beyond face-centric manipulation. On this benchmark, deep detectors outperform large vision models: NPR attains AUC 0.8748, FreqNet 0.8605, and BNet 0.8364, whereas GPT-4V reaches 0.7332, GPT-4o 0.6488, Claude 3.5 Sonnet 0.5466, and Gemini-1.5-Flash 0.5067 (Zeng et al., 9 Mar 2025).

4. Integrated sensing, communication, and long-term mapping

In wireless systems, “Chameleon: Integrated Sensing and Communication with Sub-Symbol Beam Switching in mmWave Networks” proposes a 5G-NR-compliant mmWave ISAC framework that switches an additional sensing beam within each DMRS symbol while maintaining multi-user communication beams (Gao et al., 18 Sep 2025). Implemented on a 28 GHz software-defined radio testbed with IBM 8×8 phased array modules and USRP N310, the framework updates beamformers every 0.24 $K\in[8,64]$4s, yielding approximately 34 sub-symbol switches per DMRS symbol. It exploits the full 92.16 MHz DMRS bandwidth, supports sum downlink rates up to 0.799 Gbps across two users, produces 31×31-point 2D imaging in 0.875 ms, and, with machine learning on sensing CSI, reports median localization errors of 0.14 m in distance and 0.24° in angle, together with 99.0% material-classification accuracy (Gao et al., 18 Sep 2025).

In robotics and mapping, the related but intentionally misspelled “Chamelion: Reliable Change Detection for Long-Term LiDAR Mapping in Transient Environments” stands for “Change detection and long-term Map management in transient Environments, using LiDAR and designed for Online operation” (Jang et al., 9 Feb 2026). The system addresses online detection of positive changes, negative changes, and static structure by concatenating prior-map and current-scan points as 4D sparse tensors $K\in[8,64]$5, processing them with a MinkowskiEngine backbone, and separating outputs into a class head and a cross-visibility confidence head. Training data are produced with a composition-based augmentation strategy that inserts object snapshots from a single session into scans and maps to synthesize pseudo-labeled multi-session changes. On a custom real-world dataset, reported scan-wise IoU values are 0.728, 0.484, and 0.767 across three sequences, while map-wise F1 reaches 0.747, 0.893, and 0.816. Runtime is 14.83 Hz on an RTX 3060 GPU and 2.74 Hz on a Jetson Orin NX with 8 GB (Jang et al., 9 Feb 2026).

5. Chameleon scalar fields in gravitation and dark-energy physics

In theoretical physics, a chameleon is a scalar field whose self-interactions and coupling to matter make its effective mass depend on ambient density. In the Einstein frame, the action takes the form

$K\in[8,64]$6

with matter coupling typically written as $K\in[8,64]$7, and an effective potential

$K\in[8,64]$8

The density-dependent minimum $K\in[8,64]$9 and mass $\text{Threshold}=2.5$0 underpin screening, while the thin-shell parameter

$\text{Threshold}=2.5$1

controls suppression of the fifth force outside compact bodies (Khoury, 2013). The review “Chameleon Field Theories” establishes two cosmological no-go theorems: the chameleon force range at cosmological density today is at most of order Mpc, and the conformal factor $\text{Threshold}=2.5$2 is essentially constant over the last Hubble time. It also derives a quantum-stability bound

$\text{Threshold}=2.5$3

for gravitational-strength coupling, while fifth-force experiments imply $\text{Threshold}=2.5$4, leaving a narrow window $\text{Threshold}=2.5$5 at laboratory densities (Khoury, 2013).

Subsequent work studies how screening depends on source geometry, astrophysical environment, and experimental configuration. “The shape dependence of chameleon screening” numerically solves for axisymmetric sources in a spherical vacuum chamber and finds that deviations from spherical symmetry can increase the chameleon acceleration experienced by a test particle by up to a factor of $\text{Threshold}=2.5$6, with optimized Legendre-shaped sources reaching an enhancement of approximately 2.8 relative to a sphere for $\text{Threshold}=2.5$7 and $\text{Threshold}=2.5$8 (1711.02065). “Chameleon Halo Modeling in f(R) Gravity” incorporates screening into a halo model by introducing a chameleon mass threshold $\text{Threshold}=2.5$9 and a rapidity parameter $\approx 4.54\times$0, with

$\approx 4.54\times$1

to capture the enhanced abundance of halos around the screening threshold and the corresponding nonlinear power-spectrum transition (Li et al., 2011). “A Compendium of Chameleon Constraints” then unifies astrophysical and laboratory exclusions, concluding that, with $\approx 4.54\times$2 fixed near the dark-energy scale, only extremely weakly coupled chameleons survive, and that for $\approx 4.54\times$3 and $\approx 4.54\times$4 most parameter space is excluded (Burrage et al., 2016).

The same mechanism has motivated a diverse experimental literature. “A chameleon helioscope” argues that solar chameleons, emitted mainly below a few keV, can undergo total internal reflection from dense mirror coatings at grazing incidence, permitting X-ray telescope optics to focus chameleons themselves before conversion in a magnetic field (Baker et al., 2011). “Chameleon Induced Atomic Afterglow” studies chameleons trapped in an optical cavity after photon–chameleon production; the trapped standing wave can excite atomic transitions and generate photon afterglow even after the laser and magnetic field are turned off (Brax et al., 2010). “Chameleon effect and the Pioneer anomaly” observes that spacecraft may be unscreened while planets are thin-shelled, but concludes quantitatively that the chameleon contribution to the Pioneer 10/11 anomalous acceleration is negligible, with $\approx 4.54\times$5 under the Khoury–Weltman estimate and at most a few percent using a Cassini-based bound (Anderson et al., 2012). “Towards a UV Completion for Chameleon Scalar Theories” proposes that the volume modulus in string compactifications, including a KKLT-type potential, can realize the chameleon mechanism and translates experimental screening constraints into bounds on KKLT parameters (Hinterbichler et al., 2010).

6. Cross-disciplinary semantics and disambiguation

Despite the shared title, these works are not variants of a single method or theory. They span compiler optimization, early-fusion multimodal generation, on-chip learning hardware, personalization, search-enabled LLM evaluation, privacy masks, adversarial prompt injection, AI-generated video detection, mmWave ISAC, long-term LiDAR mapping, and scalar-field gravitation (Ahn et al., 2020, Team, 2024, Blanken et al., 30 May 2025, Zhang et al., 2 Mar 2025, Ratnakar et al., 19 Oct 2025, Chow et al., 2024, Zeeshan et al., 4 Dec 2025, Zeng et al., 9 Mar 2025, Gao et al., 18 Sep 2025, Jang et al., 9 Feb 2026, Khoury, 2013). One paper explicitly notes that the spelling “Chamelion” is intentional and mnemonic rather than a typographical variation (Jang et al., 9 Feb 2026).

A plausible implication is that the name is consistently chosen to emphasize environment sensitivity, adaptive behavior, or rapid context switching. In the compiler work, adaptation occurs over previously unseen schedule design spaces; in the edge accelerator, one architecture switches between inference and learning modes; in the psychological and stance benchmarks, the central issue is state dependence across contexts or conversational turns; in the privacy and adversarial systems, the object is to alter how models perceive visual inputs under transformations; in mmWave networking, beamformers switch at sub-symbol granularity; and in the scalar-field literature, the field’s mass and force range change with ambient density (Ahn et al., 2020, Blanken et al., 30 May 2025, Harry et al., 21 Jan 2026, Ratnakar et al., 19 Oct 2025, Chow et al., 2024, Zeeshan et al., 4 Dec 2025, Gao et al., 18 Sep 2025, Khoury, 2013). On that reading, “Chameleon” functions less as a disciplinary keyword than as a recurring label for systems and theories whose defining property is controlled variability under changing conditions.