BeLLA: BEV-Language Assistant for Autonomous Driving

• BeLLA is a BEV–language architecture that integrates a unified 360° spatial representation with an LLM to answer driving-scene questions.
• It employs a two-stage pipeline using a frozen BEV encoder, a deep convolutional projector, and LoRA-finetuned LLM to optimize spatial reasoning.
• Empirical evaluations show up to +9.3% improvement on spatial tasks, though limitations include handling appearance cues and temporal reasoning.

BeLLA, in the exact stylization used by recent autonomous-driving literature, denotes the “Bird’s Eye View Large Language Assistant,” an end-to-end architecture that connects unified $360^\circ$ bird’s-eye-view (BEV) representations with a LLM for question answering in autonomous driving (Mohan et al., 5 Dec 2025). In broader arXiv usage, however, closely related stylizations—BELLA, Bella, and BeLLA—name several unrelated systems in machine learning, heliophysics, high-intensity laser physics, and bioinformatics. The term therefore functions less as a single technical object than as a recurrent acronymic label whose meaning is domain-dependent.

1. BeLLA as a BEV–language architecture for autonomous driving

BeLLA was proposed to address a specific representational problem in driving-oriented vision-language systems. Existing work often relies on either single-view encoders or aggregated multi-view features, and the stated limitation is that such designs do not provide a unified spatial representation for ego-centric directions, object relations, and wider driving context (Mohan et al., 5 Dec 2025). BeLLA’s response is to place a BEV representation at the center of the language interface: multi-camera images are fused into a unified $360^\circ$ BEV feature map, that map is compressed into a token, and the token conditions an LLM for driving-scene question answering (Mohan et al., 5 Dec 2025).

This positioning matters because BEV is already the dominant geometric substrate for perception and planning in autonomous driving. In BeLLA, the claimed advantage is not merely better sensor fusion, but better language-grounded spatial reasoning: the LLM receives a representation already organized in ego-centric coordinates, rather than being forced to infer geometry from per-camera image features (Mohan et al., 5 Dec 2025). The paper explicitly evaluates this premise on NuScenes-QA and DriveLM and reports that BeLLA consistently outperforms existing approaches on question types requiring greater spatial reasoning, including relative object positioning and behavioral understanding, with up to $+9.3\%$ absolute improvement in certain tasks (Mohan et al., 5 Dec 2025).

A plausible implication is that BeLLA treats BEV not just as a perception encoding, but as a geometric intermediate language between multi-camera sensing and autoregressive reasoning. That role distinguishes it from image-centric driving VLMs that fuse appearance well but do not expose an explicitly ego-centric spatial topology to the LLM.

2. Architecture, alignment strategy, and training protocol

BeLLA is a two-stage pipeline. First, a frozen BEV encoder maps multi-camera inputs $\mathcal{I}=\{I_1,\dots,I_N\}$ to a BEV tensor

$$\mathbf{B} = \mathrm{BEVEnc}(\mathcal{I}) \in \mathbb{R}^{H\times W\times C}.$$

The paper uses BEVFormer as the BEV encoder and keeps it frozen during both pretraining and finetuning (Mohan et al., 5 Dec 2025). Second, a trainable projector $f_\theta$ compresses $\mathbf{B}$ into a single token in the LLM embedding space,

$$\mathbf{E}_{\mathrm{BEV}} = f_\theta(\mathbf{B}) \in \mathbb{R}^{1\times d},$$

using a convolutional stack, pooling, an MLP, normalization, and linear projection (Mohan et al., 5 Dec 2025). The projector depth is not incidental: the ablations show that a deeper convolutional projector substantially outperforms a linear projector or shallow convolutional variants (Mohan et al., 5 Dec 2025).

The first training stage is BEV–text alignment. The BEV encoder and LLM are frozen, and only the projector is optimized to generate automatically derived frame-level scene descriptions from NuScenes metadata and CAN-bus signals (Mohan et al., 5 Dec 2025). The pretraining objective is standard autoregressive language modeling,

$$\mathcal{L}_{\mathrm{pretrain}} = -\sum_{t=1}^{T}\log P(y_t \mid y_{<t}, \mathbf{E}_{\mathrm{BEV}}).$$

This stage is explicitly designed to teach the LLM to interpret the BEV token as a compact summary of spatial layout and motion status (Mohan et al., 5 Dec 2025).

The second stage is BEV-conditioned driving QA. The BEV encoder remains frozen, the projector stays trainable, and the LLM is finetuned with LoRA adapters (Mohan et al., 5 Dec 2025). Given a question embedding sequence $\mathbf{Q}$, BeLLA replaces a reserved placeholder with $360^\circ$0 and trains the LLM to generate answer tokens with

$360^\circ$1

The paper reports AdamW optimization, learning rates of $360^\circ$2 for the projector and $360^\circ$3 for the LoRA-updated LLM, 10 epochs, batch size 2, and training on $360^\circ$4 NVIDIA H100 GPUs (Mohan et al., 5 Dec 2025).

The resulting system is described as end-to-end in the sense that the perception-to-language mapping from BEV features through the projector into the LLM is jointly optimized, although the BEV encoder itself is frozen (Mohan et al., 5 Dec 2025).

3. Benchmarks, empirical profile, and stated limitations

BeLLA’s principal empirical case is NuScenes-QA, a benchmark with over 460K QA pairs across 34K timesteps and discrete answer evaluation by Top-1 accuracy, and DriveLM–NuScenes, with 4,072 training frames and more than 377K QA pairs evaluated by BLEU-4, METEOR, ROUGE-L, and CIDEr (Mohan et al., 5 Dec 2025). The performance profile is asymmetric in a way consistent with the BEV interface: BeLLA is strongest on behavior, status, planning, and prediction, and weaker on appearance-centric perception questions that depend on color or texture (Mohan et al., 5 Dec 2025).

Benchmark	Key result	Paper
NuScenes-QA	Overall $360^\circ$5; Status $360^\circ$6	(Mohan et al., 5 Dec 2025)
DriveLM	Competitive overall with Qwen2 7B	(Mohan et al., 5 Dec 2025)
Ablation	No pretraining drops overall from $360^\circ$7 to $360^\circ$8	(Mohan et al., 5 Dec 2025)

On NuScenes-QA, BeLLA with LLaMA 3.2B achieves $360^\circ$9 overall accuracy, with $+9.3\%$0 on Status questions and $+9.3\%$1 on Object questions (Mohan et al., 5 Dec 2025). The paper emphasizes that Status is the clearest gain category: $+9.3\%$2 exceeds the best listed baseline by a wide margin and corresponds to the “up to $+9.3\%$3 absolute improvement” headline (Mohan et al., 5 Dec 2025). On DriveLM, BeLLA with Qwen2 7B is competitive overall and especially strong on Behavior questions, for which it reports BLEU-4 $+9.3\%$4, METEOR $+9.3\%$5, ROUGE-L $+9.3\%$6, and CIDEr $+9.3\%$7 (Mohan et al., 5 Dec 2025).

The ablation evidence is unusually direct. Removing BEV–text pretraining reduces NuScenes-QA overall accuracy from $+9.3\%$8 to $+9.3\%$9, and Status accuracy from $\mathcal{I}=\{I_1,\dots,I_N\}$0 to $\mathcal{I}=\{I_1,\dots,I_N\}$1 (Mohan et al., 5 Dec 2025). Likewise, replacing the deeper convolutional projector with a linear projector reduces overall accuracy from $\mathcal{I}=\{I_1,\dots,I_N\}$2 to $\mathcal{I}=\{I_1,\dots,I_N\}$3 (Mohan et al., 5 Dec 2025). These results suggest that the empirical contribution is concentrated less in the frozen BEV encoder than in the learned BEV-to-language interface.

The stated limitations are also tightly coupled to the design choice. Because BEV discards appearance cues, BeLLA struggles with color, texture, and fine-grained visual details such as traffic-light color (Mohan et al., 5 Dec 2025). It is single-frame rather than explicitly temporal, so temporal reasoning is indirect; it inherits failure modes from the frozen BEV encoder in night or adverse conditions; and it remains weak on precise coordinate-heavy questions (Mohan et al., 5 Dec 2025).

4. Reuse of the BELLA/Bella label in machine learning

Outside autonomous driving, the same label is reused for several unrelated ML systems.

Variant	Expansion or use	Paper
BELLA	Budget-Efficient LLM Selection via Automated skill-profiling	(Okamoto et al., 2 Feb 2026)
BELLA	Black-box model Explanations by Local Linear Approximations	(Radulovic et al., 2023)
Bella	Bayesian Low-Rank LeArning	(Doan et al., 2024)

In LLM systems research, BELLA denotes “Budget-Efficient LLM Selection via Automated skill-profiling,” a framework for choosing which LLM to use under explicit cost or latency constraints (Okamoto et al., 2 Feb 2026). Its architecture has four stages in the main text—benchmarking, critic-based skill profiling, skill clustering into capability matrices, and skill-aware cost-aware selection—and formalizes model choice through a capability matrix $\mathcal{I}=\{I_1,\dots,I_N\}$4, a task requirement matrix $\mathcal{I}=\{I_1,\dots,I_N\}$5, and a cost vector $\mathcal{I}=\{I_1,\dots,I_N\}$6 (Okamoto et al., 2 Feb 2026). The selection rule filters models by capability threshold

$\mathcal{I}=\{I_1,\dots,I_N\}$7

and then solves

$\mathcal{I}=\{I_1,\dots,I_N\}$8

The paper presents BELLA as skill-based, cost-aware, interpretable, and multi-skill, and situates it in financial reasoning with benchmarks such as FLaME (Okamoto et al., 2 Feb 2026).

In explainability research, BELLA instead stands for “Black box model Explanations by Local Linear Approximations” and denotes a deterministic, model-agnostic, post-hoc method for explaining individual predictions of regression black-box models (Radulovic et al., 2023). It avoids synthetic perturbations, constructs neighborhoods from real data, and optimizes the lower bound of the confidence interval of the Berry–Mielke universal $\mathcal{I}=\{I_1,\dots,I_N\}$9 score,

$$\mathbf{B} = \mathrm{BEVEnc}(\mathcal{I}) \in \mathbb{R}^{H\times W\times C}.$$0

to balance fidelity and generality (Radulovic et al., 2023). The method yields factual linear explanations and counterfactual explanations, and the paper reports higher generality and robustness than LIME and SHAP in its experimental setup (Radulovic et al., 2023).

In Bayesian deep learning, “Bella” denotes “Bayesian Low-Rank LeArning,” a framework that makes Bayesian neural networks practical by restricting Bayesian inference to low-rank perturbations around a pretrained model (Doan et al., 2024). For each dense layer,

$$\mathbf{B} = \mathrm{BEVEnc}(\mathcal{I}) \in \mathbb{R}^{H\times W\times C}.$$1

with rank $$\mathbf{B} = \mathrm{BEVEnc}(\mathcal{I}) \in \mathbb{R}^{H\times W\times C}.$$2, so the number of trainable parameters per particle drops from $$\mathbf{B} = \mathrm{BEVEnc}(\mathcal{I}) \in \mathbb{R}^{H\times W\times C}.$$3 to $$\mathbf{B} = \mathrm{BEVEnc}(\mathcal{I}) \in \mathbb{R}^{H\times W\times C}.$$4 (Doan et al., 2024). The paper implements both low-rank ensembles and low-rank SVGD, reports applicability to ImageNet, CAMELYON17, DomainNet, CLIP, and LLaVA, and states that the method can achieve performance close to full ensembles or SVGD while using about $$\mathbf{B} = \mathrm{BEVEnc}(\mathcal{I}) \in \mathbb{R}^{H\times W\times C}.$$5 of the trainable parameters per particle in many experiments (Doan et al., 2024).

These ML usages are technically unrelated. Their only commonality is the recycled acronymic surface form.

5. Other domain-specific BELLA systems in heliophysics, laser physics, and bioinformatics

In heliophysics, BELLA denotes the “BayEsian LocaLisation Algorithm,” a Bayesian multilateration method for locating solar radio burst sources from multi-spacecraft time-of-arrival information (Cañizares et al., 2024). Its posterior is written as

$$\mathbf{B} = \mathrm{BEVEnc}(\mathcal{I}) \in \mathbb{R}^{H\times W\times C}.$$6

with the physical model $$\mathbf{B} = \mathrm{BEVEnc}(\mathcal{I}) \in \mathbb{R}^{H\times W\times C}.$$7 and PyMC sampling to infer source position and effective propagation speed (Cañizares et al., 2024). In a Type III burst case using STEREO A/B and Wind, BELLA tracked the source from $$\mathbf{B} = \mathrm{BEVEnc}(\mathcal{I}) \in \mathbb{R}^{H\times W\times C}.$$8 to $$\mathbf{B} = \mathrm{BEVEnc}(\mathcal{I}) \in \mathbb{R}^{H\times W\times C}.$$9 over $f_\theta$0–$f_\theta$1 MHz, inferred an apparent solar-wind speed of $f_\theta$2, and found agreement with goniopolarimetry, analytical TDOA, SEMP, and HUXt (Cañizares et al., 2024). A subsequent five-spacecraft application using Parker Solar Probe, STEREO A, Wind, Solar Orbiter, and Mars Express tracked a 4 December 2021 event over $f_\theta$3–$f_\theta$4 MHz, inferred a Parker-spiral path with $f_\theta$5 and a $f_\theta$6 East footpoint, and used the inferred apparent path to quantify radio-wave scattering (Cañizares et al., 20 Feb 2025).

In high-intensity laser physics, BELLA denotes the Berkeley Lab Laser Accelerator petawatt facility in its dual-beamline configuration for strong-field QED (Turner et al., 2022). The upgraded BELLA PW system is described as a 1 Hz, petawatt-class Ti:sapphire laser centered near $f_\theta$7 nm, with up to $f_\theta$8 J available at the target chamber, pulse durations of $f_\theta$9–$\mathbf{B}$0 fs, and peak power up to $\mathbf{B}$1 PW at the chamber (Turner et al., 2022). The dual beamlines enable collisions between GeV-class laser-wakefield-accelerated electron beams and intense laser pulses, with simulated access to a nonlinear quantum parameter up to $\mathbf{B}$2, and the production of GeV-class, mrad-divergence positron beams via the Breit–Wheeler process (Turner et al., 2022).

In genomics, BELLA is not expanded in the LOGAN paper, but is identified as “a long-read many-to-many overlapping and alignment software” and “long-read many-to-many overlapper and aligner” (Zeni et al., 2020). Its pipeline consists of k-mer pruning, overlap detection via sparse matrix–matrix multiplication, seed-and-extend pairwise alignment, and adaptive-threshold filtering (Zeni et al., 2020). LOGAN replaces BELLA’s SeqAn X-drop alignment kernel with a GPU implementation and reports that this integration improves overall BELLA runtime by up to $\mathbf{B}$3 while preserving equivalent results (Zeni et al., 2020).

Taken together, these examples show that BELLA/BeLLA functions as a highly overloaded acronym across scientific computing, experimental physics, and space science. The shared label does not imply methodological continuity.

6. Orthographic collisions and bibliographic disambiguation

Not every occurrence of the string “Bella” or “La Bella” in the literature refers to an acronymic system. In climatology, “La Bella” is the name of a pluviometric station in the Colombian Eje Cafetero, located at $\mathbf{B}$4 and $\mathbf{B}$5 m above sea level, and used in a study of sunspot cycles and regional precipitation (González-Lozano, 2015). In arithmetic geometry, several papers discuss “Bellaïche–Stevens” $\mathbf{B}$6-adic $\mathbf{B}$7-functions and Bellaïche’s results on the eigencurve; these uses involve the surname Bellaïche rather than a BELLA acronym (Benois et al., 2020, Benois et al., 2024, Charlton et al., 2024). In topology, one paper gives a partial answer to a long-standing question of Angelo Bella, again referencing a surname rather than a named system (Carlson, 2024).

This bibliographic overlap is not merely cosmetic. In arXiv indexing and keyword search, exact-stylization differences—BeLLA, BELLA, Bella, Bellaïche, and La Bella—can collapse into a single retrieval set. For technical reading, the relevant disambiguation is therefore by expansion and field: “Bird’s Eye View Large Language Assistant” in autonomous driving (Mohan et al., 5 Dec 2025), “Budget-Efficient LLM Selection via Automated skill-profiling” in LLM routing (Okamoto et al., 2 Feb 2026), “Bayesian Low-Rank LeArning” in Bayesian deep learning (Doan et al., 2024), “BayEsian LocaLisation Algorithm” in heliophysics (Cañizares et al., 2024), the Berkeley laser facility in strong-field QED (Turner et al., 2022), and the long-read overlapper/alignment tool in bioinformatics (Zeni et al., 2020).

A common misconception is to treat BeLLA as a single cross-domain framework. The literature does not support that reading. What it supports is an acronym family reused independently in several research programs, with the exact stylization “BeLLA” currently anchored most specifically to the BEV–LLM autonomous-driving architecture introduced in late 2025 (Mohan et al., 5 Dec 2025).