FLEXI: Multifaceted Research Naming Convention

• FLEXI is a recurring research naming convention that signifies flexible resource allocation, adaptable model structures, and configurable systems across multiple domains.
• Its applications span diverse fields such as speech interaction, subgroup discovery, computational fluid dynamics, optical networking, and more, each with domain-specific metrics and evaluation protocols.
• The heterogeneous use of FLEXI highlights its role as an adaptable label that fosters innovations ranging from full-duplex LLM benchmarks to dynamic spectrum allocation and efficient CFD solvers.

FLEXI is a recurrent research name rather than a single standardized concept. In the arXiv literature, it denotes distinct methods, benchmarks, and software systems in spoken interaction, subgroup discovery, higher-education LLM infrastructure, computational fluid dynamics, and optical networking. A broader “Flexi-” naming pattern also appears in graph mining, neuromorphic acceleration, diffusion models, robust classification, spike-based vision, and survival analysis. The shared label therefore functions primarily as a naming convention for flexibility, configurability, or adaptive modeling, not as a unified technical framework.

1. Disambiguation and scope

The heterogeneity of the name is central to its interpretation. Representative usages include the following.

Usage	Domain	Core description
FLEXI	Speech interaction	Benchmark for full-duplex LLM-human spoken interaction (Ge et al., 26 Sep 2025)
FLEXI	Subgroup discovery	Optimal binning method for numeric and ordinal attributes (Nguyen et al., 2015)
FLEXI	Higher-education LLM systems	FernUni LLM Experimental Infrastructure (Zesch et al., 2024)
FLEXI	Computational fluid dynamics	High order discontinuous Galerkin framework (Krais et al., 2019)
Flexi-grid	Optical networking	Spectral-efficiency optimization in long-haul optical systems (Foggi et al., 2014)
Flexi-clique	Graph mining	Size-adaptive cohesive-subgraph model (Kim et al., 11 Feb 2026)
Flexi-Transducer	On-device ASR	Flexible latency/accuracy/compute trade-off for multi-domain scenarios (Mahadeokar et al., 2021)
FLEXI-Haz	Survival analysis	Flexible hazard model with a partially linear structure (Arie et al., 11 Dec 2025)

This distribution shows that “FLEXI” is best treated encyclopedically as a family of unrelated research artifacts. In most cases, the term signals one of three design intentions: flexible resource allocation, flexible model structure, or flexible operating points.

2. Spoken interaction, LLM infrastructure, and on-device ASR

In speech interaction research, FLEXI is the name of a benchmark for full-duplex Speech-to-Speech LLMs. It defines full-duplex S2S as interaction in which both user and model speak simultaneously, supporting barge-in and natural turn-taking rather than strict alternation, and it explicitly incorporates model interruption in emergency scenarios such as “My house is on fire!” (Ge et al., 26 Sep 2025). The benchmark organizes evaluation into six scenarios: Standard Turn-Taking, Pause Handling, User Interrupt, Model Interrupt, User Backchannel, and Model Backchannel. It formalizes turn-taking with speaker states $y_t^k$, Inter-Pausal Units, overlap, and interrupt events, and measures latency, quality, and conversational effectiveness through metrics including $$L=t_{\mathrm{start}}(\mathrm{response})-t_{\mathrm{end}}(\mathrm{user\_utterance})$$, Takeover Rate, Jump-In Rate, Turn Termination Rate, Topic Shift Score, Emergency Detection Score, Backchannel Rate, and coherence. Empirically, the benchmark reports substantial differences between Moshi, Freeze-Omni, Vita1.5, and Gemini. For example, in User Interrupt, Gemini reaches $\mathrm{TTR}=0.935$, latency $=1.766\,\mathrm{s}$, and $\mathrm{TSS}=4.51$, whereas the open-source models remain below $\mathrm{TTR}<0.6$; in Model Interrupt, open-source models interrupt more frequently but with lower $\mathrm{EDS}$, while Gemini records $\mathrm{TOR}=0$ and $\mathrm{EDS}=2.19$. The same paper proposes next token-pair prediction, in which a unified end-to-end transformer predicts a pair $(t^{\mathrm{ext}}_{n+1}, c_{n+1})$ with $$L=t_{\mathrm{start}}(\mathrm{response})-t_{\mathrm{end}}(\mathrm{user\_utterance})$$0, trained with a joint cross-entropy objective.

A separate use of the name denotes the FernUni LLM Experimental Infrastructure, an institutional platform for hosting open LLMs on university-owned servers (Zesch et al., 2024). The stated motivation is that cloud-hosted APIs can expose sensitive user data to third parties, incur unpredictable per-token fees, create inequities, and tie workflows to external providers. The pilot deployment uses a bare-metal Kubernetes cluster on Server A and Docker-Compose on Server B, with Ollama as the model-serving component, Open WebUI as the chat front end, Traefik as reverse proxy and load balancer, Checkmk plus a custom nvidia_smi plugin for monitoring, and CAIPI middleware for Moodle integration. Both servers reside in the university data center and run Ubuntu 22.04 LTS. Reported throughput values on Server B are 127 tokens/s for Phi3 (4 B), 77 for Gemma (7 B), 94 for Mistral (7 B), and 82 for Llama3 (8 B). The infrastructure is evaluated not only as systems software but also as an institutional governance model emphasizing on-premise deployment, rate limiting, SSO integration, logging, and GDPR-oriented prompt sanitization.

Flexi-Transducer addresses a related but distinct problem: a single compact on-device ASR model serving multiple use-cases or domains with different latency and accuracy requirements (Mahadeokar et al., 2021). It uses a streaming Emformer encoder with domain-specific segment sizes, Alignment Restricted RNN-T loss, and a domain indicator vector. The paper sets $$L=t_{\mathrm{start}}(\mathrm{response})-t_{\mathrm{end}}(\mathrm{user\_utterance})$$1 for VoiceCommand and $$L=t_{\mathrm{start}}(\mathrm{response})-t_{\mathrm{end}}(\mathrm{user\_utterance})$$2 for Dictation, fixes the left context to $$L=t_{\mathrm{start}}(\mathrm{response})-t_{\mathrm{end}}(\mathrm{user\_utterance})$$3 chunks, and restricts alignments through a right-buffer parameter $$L=t_{\mathrm{start}}(\mathrm{response})-t_{\mathrm{end}}(\mathrm{user\_utterance})$$4. The domain indicator $$L=t_{\mathrm{start}}(\mathrm{response})-t_{\mathrm{end}}(\mathrm{user\_utterance})$$5 is injected into both encoder and predictor. The reported effect is a single model that can provide very low latency for voice commands and stronger accuracy for dictation, with results such as Dict WER $$L=t_{\mathrm{start}}(\mathrm{response})-t_{\mathrm{end}}(\mathrm{user\_utterance})$$6 and V_Cmd WER $$L=t_{\mathrm{start}}(\mathrm{response})-t_{\mathrm{end}}(\mathrm{user\_utterance})$$7 under domain-specific Emformer context, and further latency improvements when the domain vector is added.

3. High-order computational fluid dynamics and related exascale workflows

In computational fluid dynamics, FLEXI is an open-source, high-order consistent simulation tool chain for solving the compressible Navier-Stokes equations in a high performance computing setting (Krais et al., 2019). The framework solves general hyperbolic-parabolic systems of conservation laws in strong form,

$$L=t_{\mathrm{start}}(\mathrm{response})-t_{\mathrm{end}}(\mathrm{user\_utterance})$$8

using a discontinuous Galerkin spectral-element method on curvilinear hexahedral elements. The approximate solution is expanded in a nodal basis of degree $$L=t_{\mathrm{start}}(\mathrm{response})-t_{\mathrm{end}}(\mathrm{user\_utterance})$$9, numerical interface fluxes are supplied by Riemann solvers such as Roe, Lax-Friedrichs, and HLL, viscous terms use BR1 or BR2 liftings, and explicit low-storage Runge-Kutta methods provide time integration. The framework also emphasizes HOPR for mesh handling, POSTI for post-processing, HDF5-based parallel I/O, shock capturing through DG-to-FV subcell switching, split-form DGSEM on LGL nodes, and unusually extensive support for reproducibility through solution-file metadata that records the complete input deck, Git commit, local patches, and CMake flags.

Subsequent work repositions FLEXI explicitly for exascale CFD (Blind et al., 2023). The framework is reported to scale excellently to over 500 000 CPU cores on HAWK at HLRS. To support a three-dimensional wing-segment shock-buffet case, the paper adds a compressibility-aware wall model, a finite-volume-based shock capturing method, and the TensorFlow-Fortran-Binding (TFFB) for deploying trained machine learning models inside Fortran solvers. The wall-model study on a NACA 64A-110 airfoil compares WRLES, WMLES-Spalding, and WMLES-ODE; the shock-capturing study validates a hybrid DGSEM/FV method on a linear compressor cascade at $\mathrm{TTR}=0.935$0 and $\mathrm{TTR}=0.935$1; and the TFFB layer provides an ISO_C_BINDING-mediated interface from Fortran to the TensorFlow C-API.

The same solver has been extended to particle-laden flows (Kopper et al., 2022). Here FLEXI remains the Eulerian-phase solver while a Lagrangian particle module adds point-particle dynamics, one- and two-way coupling, ray-tracing-based boundary intersection handling on curved elements, and large-scale parallel particle communication. The paper reports weak scaling up to 256 nodes and 32 768 cores with approximately $\mathrm{TTR}=0.935$2 parallel efficiency, about $\mathrm{TTR}=0.935$3 elements, and about $\mathrm{TTR}=0.935$4 particles. The implementation relies on Hilbert space-filling-curve ordering, MPI-3 shared memory for mesh data, a Cartesian background mesh for host-cell search, and warm-restart load balancing.

FLEXI also appears as the flow simulator inside reinforcement-learning-based active flow control (Kurz et al., 28 Mar 2025). In that setting, flow simulations are conducted in parallel using the high-order discontinuous Galerkin framework FLEXI, while Relexi manages PPO training and graph-neural-network policies. For the two-dimensional cylinder benchmark, the setup uses 150 curved quadrilaterals and polynomial degree $\mathrm{TTR}=0.935$5, yielding 3 750 DOFs per unknown. The reported training regime uses $\mathrm{TTR}=0.935$6 environments on 512 cores, and a full training of 2 000 PPO iterations completes in approximately 48 hours of wall-clock time. Both MLP- and GNN-controlled policies achieve $\mathrm{TTR}=0.935$7 from an uncontrolled baseline $\mathrm{TTR}=0.935$8, while the GNN policy yields a smaller drag-oscillation amplitude.

4. Flexi-grid optical networking and spectrum allocation

In optical communications, FLEXI usually refers not to a software framework but to flexi-grid optical systems, in which fiber spectrum is divided into finer slots and channels may occupy variable-width contiguous groups. The spectral-efficiency optimization problem in such systems is formulated explicitly in terms of achievable information rate and spectral efficiency (Foggi et al., 2014). For one polarization and carrier, the sampled received signal is modeled as

$\mathrm{TTR}=0.935$9

and spectral efficiency is defined as

$=1.766\,\mathrm{s}$0

with carrier spacing $=1.766\,\mathrm{s}$1 and symbol time $=1.766\,\mathrm{s}$2. The paper studies flexi-grid WDM on a 12.5 GHz grid with $=1.766\,\mathrm{s}$3, WSS filters modeled as 3rd-order super-Gaussian filters of 35.75 GHz 3 dB bandwidth, and limited-complexity MAP detection with channel shortening and trellis memory $=1.766\,\mathrm{s}$4 or $=1.766\,\mathrm{s}$5. Its central result is that a limited-complexity trellis receiver can either extend the reach of Nyquist WDM flexi-grid networks or enable denser-than-Nyquist packing. At $=1.766\,\mathrm{s}$6 and 2400 km, the paper reports $=1.766\,\mathrm{s}$7 for symbol-by-symbol detection and $=1.766\,\mathrm{s}$8 for MAP with $=1.766\,\mathrm{s}$9, together with a reach improvement of more than $\mathrm{TSS}=4.51$0 longer for 25% OH FEC.

A complementary line of work studies routing and spectrum assignment under dynamic traffic (Lohani et al., 2021). The network is modeled as an undirected graph $\mathrm{TSS}=4.51$1, each link spectrum is sliced into 12.5 GHz slots, and a bitmap $\mathrm{TSS}=4.51$2 records slot availability. The defining constraints are spectrum contiguity, spectrum continuity, and non-overlap. The proposed adaptive algorithms prioritize the maximum run-length of consecutive free slots rather than only static distance or hop count. Type II, the joint RSA variant, integrates routing and feasibility checking by propagating path bitmaps and pruning any expansion for which $\mathrm{TSS}=4.51$3. In simulations on NSFNET and USNET, Type II and Type III outperform shortest-path and $\mathrm{TSS}=4.51$4-shortest-path baselines; for NSFNET at roughly 15 Erlangs, shortest-path blocks about $\mathrm{TSS}=4.51$5 of requests while Type II blocks about $\mathrm{TSS}=4.51$6, and at roughly 30 Erlangs, $\mathrm{TSS}=4.51$7-SP blocks about $\mathrm{TSS}=4.51$8 while Type II blocks about $\mathrm{TSS}=4.51$9.

A third strand replaces direct simulation with Markov modeling (Erreygers et al., 2018). The paper constructs exact CTMCs for low- and high-bit-rate traffic classes, reduced-state approximate CTMCs, and reduced-state imprecise Markov chain models that give guaranteed lower and upper bounds on blocking probabilities. In the detailed model, the state records how many superchannels carry exactly $\mathrm{TTR}<0.6$0 class-1 flows and no class-2 flow; blocking probabilities are then defined as sums of steady-state probabilities over blocking events $\mathrm{TTR}<0.6$1 and $\mathrm{TTR}<0.6$2. The reduced model aggregates states to triples $\mathrm{TTR}<0.6$3, and the imprecise model replaces uncertain microscopic transition rates with interval-valued rate operators $\mathrm{TTR}<0.6$4 and $\mathrm{TTR}<0.6$5. The result is a tractable performance-evaluation framework that can provide policy-specific or policy-independent bounds on blocking.

5. Pattern mining, dense subgraphs, and size-adaptive cohesion

In subgroup discovery, FLEXI is a binning method designed to construct one-dimensional binary features for numeric and ordinal attributes by directly optimizing a subgroup-quality measure $\mathrm{TTR}<0.6$6 (Nguyen et al., 2015). For a family $\mathrm{TTR}<0.6$7 of all possible adjacent-bin partitions of an attribute $\mathrm{TTR}<0.6$8, the primary objective is

$\mathrm{TTR}<0.6$9

To make this tractable, the method first pre-partitions sorted values into $\mathrm{EDS}$0 equal-frequency initial bins $\mathrm{EDS}$1, then applies dynamic programming with

$\mathrm{EDS}$2

The paper instantiates the framework with WRAcc, z-score, KL-divergence, Hellinger, and QR-divergence, with precomputation ranging from $\mathrm{EDS}$3 for WRAcc and z-score to $\mathrm{EDS}$4 for QR. Reported experiments on synthetic and real-world data sets show that FLEXI outperforms state of the art with up to 25 times improvement in subgroup quality, improves average WRAcc by up to 300% over several baselines, and remains stable for $\mathrm{EDS}$5.

Flexi-clique generalizes the idea of flexible structure from one-dimensional binning to graph cohesion (Kim et al., 11 Feb 2026). For a simple undirected graph $\mathrm{EDS}$6 and $\mathrm{EDS}$7, a vertex subset $\mathrm{EDS}$8 is a Flexi-clique if and only if $\mathrm{EDS}$9 is connected and

$\mathrm{TOR}=0$0

The model is explicitly non-hereditary and non-quasi-hereditary, the decision problem is NP-hard, and no constant-factor approximation exists unless $\mathrm{TOR}=0$1. The paper introduces a fast heuristic, the Flexi-Prune Algorithm (FPA), based on core decomposition and connectivity-aware peeling, and an exact Efficient Branch-and-Bound Algorithm (EBA) with six pruning rules. Experiments on 9 real-world graphs and LFR benchmarks up to $\mathrm{TOR}=0$2M and $\mathrm{TOR}=0$3M show that FPA achieves at least $\mathrm{TOR}=0$4 of OPT on most real graphs, $\mathrm{TOR}=0$5 on Erdős, and $\mathrm{TOR}=0$6 on Florida, while EBA solves exact OPT in less than 1 s on $\mathrm{TOR}=0$7k and less than 20 s on $\mathrm{TOR}=0$8k.

The common theme across these two uses is size-adaptive optimization. In subgroup discovery, the adaptive object is the discretization of a scalar attribute under an arbitrary quality measure. In graph mining, it is the minimum-degree threshold $\mathrm{TOR}=0$9, which grows sub-linearly with subgraph size and thereby departs from fixed-density notions such as clique, $\mathrm{EDS}=2.19$0-plex, and $\mathrm{EDS}=2.19$1-quasi-clique.

6. Extended “Flexi-” family: representation learning, hardware, vision, and statistics

The broader “Flexi-” family includes earlier methodological uses of the prefix and several later architectures. An early example is flexi-Weighted Least Squares, which unifies phylogenetic trees, multidimensional scaling, and Neighbor Nets through

$\mathrm{EDS}=2.19$2

with polynomial weights $\mathrm{EDS}=2.19$3 and exponential weights $\mathrm{EDS}=2.19$4 (Waddell et al., 2010). In that formulation, Sammon stress is recovered as the special case $\mathrm{EDS}=2.19$5, ordinary least-squares MDS as $\mathrm{EDS}=2.19$6, and model selection proceeds through likelihood, AIC, and BIC. The paper’s empirical example on population-genetic $\mathrm{EDS}=2.19$7 distances reports that fWLS trees are favored by BIC, whereas Neighbor Net fits best numerically but lacks robust reticulate structure under residual resampling.

In surrogate modeling for parametric PDEs, Flexi-VAE is a variational autoencoder with a neural propagator for single-shot prediction (Rafiq et al., 14 May 2025). It trains encoder, decoder, and propagator jointly, compares a Direct Concatenation Propagator and a Positional Encoding Propagator, and argues that DCP yields a more disentangled and physically meaningful latent space. On Burgers’ equation and advection-diffusion, the reported runtime advantage over an AE-LSTM baseline reaches over $\mathrm{EDS}=2.19$8 on CPU and $\mathrm{EDS}=2.19$9 on GPU for large temporal shifts, while preserving accurate long-horizon forecasts.

In edge neuromorphic computing, Flexi-NeurA is a configurable neuromorphic accelerator paired with the Flex-plorer design-space exploration tool (Farahani et al., 20 Feb 2026). The core architecture consists of a Configurable Neuron Unit, an AER Management Unit, a lightweight Controller, and an SPI Slave interface. It supports event-driven operation, time-multiplexed neuron updates, AER packet streaming, and design-time choices of neuron model, topology, bit-widths, layer sizes, and reset policy. Flex-plorer uses simulated annealing over candidate precision settings with

$(t^{\mathrm{ext}}_{n+1}, c_{n+1})$0

For a three-layer 256–128–10 fully connected network with LIF neurons mapped onto two processing cores, the paper reports $(t^{\mathrm{ext}}_{n+1}, c_{n+1})$1 accuracy on MNIST, $(t^{\mathrm{ext}}_{n+1}, c_{n+1})$2 inference latency, 1 623 logic cells, 7 BRAMs, and $(t^{\mathrm{ext}}_{n+1}, c_{n+1})$3 of total power.

In diffusion modeling, Flexi-UNet is the resolution-adaptive UNet backbone inside Scale Space Diffusion (Mukhopadhyay et al., 9 Mar 2026). The key idea is that noisy states whose information content is equivalent to a low-resolution image should not necessarily traverse the full encoder-decoder. Flexi-UNet therefore activates only those encoder and decoder stages compatible with the timestep’s mapped resolution $(t^{\mathrm{ext}}_{n+1}, c_{n+1})$4, using zero-filled skip-connections and lightweight $(t^{\mathrm{ext}}_{n+1}, c_{n+1})$5 projections for bypassed stages. On CelebA-64, the 2-level configuration reports FID $(t^{\mathrm{ext}}_{n+1}, c_{n+1})$6, train time about 62 h, and 50 GFLOPs per iteration, compared with FID $(t^{\mathrm{ext}}_{n+1}, c_{n+1})$7, about 70 h, and about 60 GFLOPs for DDPM-$(t^{\mathrm{ext}}_{n+1}, c_{n+1})$8. The summary conclusion is a reduction of 20–50% in FLOPs and 15–50% in wall-clock time at a tolerable FID cost.

In high-speed vision, SpikeMM describes “Flexi-Magnification” of micro-motions using spike cameras (Zhang et al., 2024). The pipeline replaces frame-based motion magnification with asynchronous spike-stream processing, multi-level temporal extraction, self-supervised blind-spot networks, implicit neural representation upsampling, and downstream motion-magnification back-ends. The paper evaluates flow consistency and motion smoothness across scenes such as TuningFork, ShortRuler, LongRuler, and Balloon, with average values improving from 0.3779 and 0.3186 for TFI baselines to 0.0115 and 0.0152 for SpikeMM in the reported tables.

In robust classification and survival analysis, the same naming convention marks flexible weighting or partially linear structure. Flexi-Fuzz-LSSVM defines a piecewise “flat-then-decay” membership weight

$(t^{\mathrm{ext}}_{n+1}, c_{n+1})$9

combines it with local class probability and imbalance ratio, and plugs the result into LSSVM (Akhtar et al., 2024). On UCI and KEEL data, the paper reports average accuracies of $$L=t_{\mathrm{start}}(\mathrm{response})-t_{\mathrm{end}}(\mathrm{user\_utterance})$$00 for Flexi-Fuzz-LSSVM-I and $$L=t_{\mathrm{start}}(\mathrm{response})-t_{\mathrm{end}}(\mathrm{user\_utterance})$$01 for Flexi-Fuzz-LSSVM-II, and on ADNI it reports $$L=t_{\mathrm{start}}(\mathrm{response})-t_{\mathrm{end}}(\mathrm{user\_utterance})$$02 and $$L=t_{\mathrm{start}}(\mathrm{response})-t_{\mathrm{end}}(\mathrm{user\_utterance})$$03, with $$L=t_{\mathrm{start}}(\mathrm{response})-t_{\mathrm{end}}(\mathrm{user\_utterance})$$04 accuracy for CN vs AD using Flexi-Fuzz-II. FLEXI-Haz, by contrast, is a baseline-free survival model with hazard

$$L=t_{\mathrm{start}}(\mathrm{response})-t_{\mathrm{end}}(\mathrm{user\_utterance})$$05

where a linear component preserves interpretability and a ReLU network captures time-covariate interactions without the proportional hazards assumption (Arie et al., 11 Dec 2025). Theoretical results include minimax-optimal convergence of the nuisance component up to log factors, root-$$L=t_{\mathrm{start}}(\mathrm{response})-t_{\mathrm{end}}(\mathrm{user\_utterance})$$06 consistency and asymptotic normality for $$L=t_{\mathrm{start}}(\mathrm{response})-t_{\mathrm{end}}(\mathrm{user\_utterance})$$07, and semiparametric efficiency.

Taken together, these derived usages show that “Flexi” has become a productive research prefix for methods that expose design-time configurability, adaptive weighting, resolution-aware computation, or flexible structural assumptions. The unifying idea is nominal rather than formal: each method operationalizes flexibility in a domain-specific way.