SOLE: Multifaceted Science & Technology

• SOLE is a multifaceted term defining exclusive reward models in RL, efficient transformer operations, 3D segmentation frameworks, and specialized instrumentation in solar physics and biomedical sensing.
• In reinforcement learning, SOLE-R1 uses a single pretrained video-language model to generate detailed progress estimates, achieving over 50% success on previously unseen tasks through hybrid SFT and RLVR training.
• Other SOLE applications include transformer hardware/software co-design yielding 30x to 90x acceleration, combinatory logic frameworks for termination proofs, and advanced soft X-ray and piezoresistive sensors in solar and gait analysis.

SOLE appears as a term or acronym in several distinct scientific and technical literatures, each with precise, domain-specific meaning. This entry surveys core usages in deep reinforcement learning, hardware/software co-design, 3D vision-LLMs, combinatory logic, solar physics instrumentation, and biomedical sensing, together with a brief note on historical–astronomical context and a fisheries case. Each section is organized by the most prominent SOLE conceptions in recent technical research.

1. SOLE as a Video-Language Reward Model for RL

SOLE, the Self-Observing Learner, refers to using a single pretrained video-language reasoning model as the sole reward signal in reinforcement learning (RL) for robotic control. The canonical reference is SOLE-R1, which outputs, at each timestep $t$, a pair $(m_t, p_t)$:

• $m_t$: a chain-of-thought (CoT) natural language explanation reasoning over observed video frames $o_t$ and a goal $g$, describing state changes and next subgoals.
• $p_t\in[-100,100]$: a scalar progress estimate, serving as the only reward signal $r_t$ for the RL agent.

Model Architecture

SOLE-R1 builds upon a vision-language transformer (e.g., Qwen3-VL-8B-Instruct) with three central components:

• Spatiotemporal encoder for video: framewise Visual Transformer (ViT-style) encodings, aggregated over a sliding window of $K$ frames, cross-attended with tokenized text.
• Goal and prior progress conditioning: input sequence is prepended with the natural-language goal, current window of video, and prior scalar $p_{t-1}$ (with stochastic masking to enforce non-myopic global reasoning).
• Autoregressive CoT decoding: > (free-form description of changes and intent) and <answer> (numeric progress $p_t$).

  Training and Data Synthesis

  The training pipeline consists of:

  • Large-scale trajectory and reasoning synthesis, generating both expert and non-expert (perturbed, regressed) video trajectories, with temporally aligned dense progress labels $(m_t, p_t)$0 constructed from geometric proxies or temporal normalization.

  • Synthesis of CoT traces $(m_t, p_t)$1 from progress signals, paraphrased to enforce linguistic diversity.
  • Hybrid supervised fine-tuning (SFT) on foundational, embodied, and synthesized traces, followed by RL from verifiable reward (RLVR), using a Generalized Ratio Policy Optimization (GRPO) loss with a reference model, parseability and numeric accuracy scores, and KL-regularization.

  Reward Signal Formalism

  At each $(m_t, p_t)$2: $(m_t, p_t)$3 with $(m_t, p_t)$4 a scaling factor and $(m_t, p_t)$5 a threshold. If policy acts at higher frequency than $(m_t, p_t)$6 labels, $(m_t, p_t)$7 is interpolated across frames.

  Experimental Results and Properties

  SOLE-R1 enables zero-shot RL on unseen tasks, using only observation and language goal, without demonstrations or explicit success signals. Across 24/50 previously unseen real or simulated manipulation tasks, SOLE-R1 achieves $(m_t, p_t)$8 success, outperforming GPT-5, Gemini-3-Pro, and specialist rewarders, while exhibiting robustness to reward hacking.

  Key capabilities are attributed to:

  • Learned generic "progress primitives" (approach, contact, place, articulate) from diverse training.
  • Per-step CoT anchoring, enforcing alignment between state changes and semantic progress.
  • Explicit negative examples from perturbed data, increasing distribution-shift robustness.
  • Hybrid SFT+RLVR for accurate and well-calibrated reward regression(Schroeder et al., 30 Mar 2026).

  2. SOLE in Transformer Hardware/Software Co-Design

  SOLE, or "SOftmax & LayerNorm E-fficient," is a hardware/software co-design scheme that replaces standard transformer Softmax and LayerNorm operations with two tightly coupled, highly quantized approximations:

  • E2Softmax: Implements Softmax by log$(m_t, p_t)$9-quantizing exponent output to 4 bits, computes $m_t$0 as $m_t$1 via shift-and-add logic, and uses an approximate, log-based divider for normalization, eliminating multipliers and LUTs.
  • AILayerNorm: Compresses 8-bit activations to 4 bits (with 1-bit tag), computes sums $m_t$2 and $m_t$3 in low precision (via a 16-entry LUT and shifting), uses power-of-two factors for channel alignment, and applies normalization and affine transformation in 8-bit arithmetic.

  SOLE demonstrates minimal accuracy loss without retraining:

  • Top-1 drops of only 0.2–0.9 pp for ImageNet and GLUE/SQuAD BERT workloads, for FP32 and INT8.
  • Stand-alone Softmax and LayerNorm acceleration of over $m_t$4x (up to $m_t$5x) vs NVIDIA 2080Ti GPU.
  • Energy- and area-efficiency improvements of 3-4x against state-of-the-art custom accelerators(Wang et al., 20 Oct 2025).

  3. SOLE: Segment Any 3D Object with Language

  SOLE is a multimodal, open-vocabulary 3D segmentation framework integrating geometric point clouds with natural language prompts. The architecture features:

  • A backbone combining sparse 3D-UNet (MinkowskiNet) with projected 2D CLIP features.
  • A cross-modality decoder performing self- and cross-attention across point, visual, and text features, yielding mask logits and semantic embeddings per predicted entity.
  • Hierarchical cross-modal loss, including mask–visual, mask–caption, and mask–entity associations via contrastive supervision.

  Trained without class-level annotation, SOLE achieves high AP on ScanNetv2/ScanNet200/Replica and approaches fully supervised performance. It supports rich, free-form instructions and provides robust segmentation for novel and cross-domain categories. Main limitations involve ambiguity in co-occurring objects and degraded feature quality for highly occluded targets(Lee et al., 2024).

  4. SOLE in Combinatory Logic: Sole Combinatory Calculus

  In rewriting theory, "sole combinatory calculus" refers to systems with one combinator and application, formalized as a singleton TRS $m_t$6:

  • Signature: $m_t$7; terms are either the constant $m_t$8 or applications thereof.
  • Rewrite rule: $m_t$9 with $o_t$0 and $o_t$1 a non-erasing expression in $o_t$2 and $o_t$3.

  Termination analysis is achieved by synthesizing tree automata with a final sink state and encoding closure/reachability/rewriting constraints as SAT problems. This method has fully resolved eight previously open non-erasing combinator cases—identifying explicit automata and infinite self-embedding counterexamples. This demonstrates that termination is generally unprovable for broad families of non-erasing singleton combinators and establishes a reusable SAT-driven framework for TRS non-termination proofs(Nakano et al., 2024).

  5. SOLE in Solar Physics: Solar Low-Energy X-ray Spectrometer

  SOLE (often SoLEXS) is a soft X-ray spectrometer on India's Aditya-L1 mission positioned at the Sun–Earth L1 point. It features:

  • Two Silicon Drift Detectors (SDDs), aperture areas 7.1 mm² and 0.1 mm², for dynamic range from A-class to X-class flares.
  • 2–22 keV coverage, 170 eV FWHM at 5.9 keV, 1 s cadence, 100% observational duty cycle (no Earth's shadow effects).
  • On-board and ground calibration (energy-channel relations, FWHM vs $o_t$4, spectral redistribution, ARF, deadtime).
  • Cross-calibration with GOES-XRS and Chandrayaan-2/XSM confirms $o_t$5 radiometric accuracy.

  SOLE provides continuous spectroscopic data instrumental for studies into coronal heating (e.g., nanoflares), flare energetics, and CME-flare coupling, enabled by its real-time 1 s full-Sun coverage and robust instrumental calibration(Sarwade et al., 30 Sep 2025).

  SOLE-Related Solar Symbolism in Historical Astronomy

  The Sun ("sole") also appears in Roman urban foundation practices—as in Augusta Taurinorum (Turin)—where the sunrise azimuth aligns with city axes, and in Augustan propaganda via solar symbolism and festival calendrics, reconstructed with mathematical corrections for atmospheric refraction, horizon altitude, and astronomical event timing(Caranzano et al., 2019).

  6. SOLE in Biomedical Sensing: Pressure-Sensing Shoe Sole

  SOLE refers to a soft, flexible, piezoresistive sensor array integrated into a shoe sole for gait analysis. Features include:

  • Ecoflex/Graphene composite as the sensor medium, sandwiched between copper electrodes, yielding negative piezoresistivity.
  • 15×15 mm², ~1.25 mm thick sensors, arrayed (5 per insole) under anatomical load zones.
  • ESP32 microcontroller ADC readout, Bluetooth LE wireless transmission, and a MATLAB GUI for real-time heatmap, COP trajectory, stride/symmetry analysis.
  • Sensitivity $o_t$6, ~120 ms response time, $o_t$76% hysteresis, $o_t$81,000 cycle durability.

  This platform is intended for gait-quality monitoring and rehabilitation in both healthy and clinical populations (foot disorder, neuromotor pathology)(Adeel et al., 24 Jan 2025).

  7. Additional Notation: "Sole" in Learning Theory

  In learning theory, "sole supervision" denotes purely supervised learning objectives in the absence of adversarial or GAN-style discriminator feedback. Sole supervision may suffer from vanishing gradients near minima, slowing convergence. Adversarial augmentation maintains a persistent gradient norm, strictly lowering expected empirical risk and accelerating convergence (rigorously established under Jacobian and loss smoothness conditions)(Rout, 2019).

  ---

  In all scientific domains above, SOLE denotes an object, framework, or instrumentation with either the central role of exclusive reward (as in RL), resource-efficient architecture (as in transformer hardware), language-driven segmentation, reductionist calculi (as in combinators), instrumentation (as in solar physics), biomechanical sensing, or as technical jargon in fisheries or learning theory. Each advances domain-specific state-of-the-art methods or theory, frequently delivering substantive empirical or analytical advantages over prior baselines.