Verbal Technical Analysis (VTA)

• VTA is a methodology that converts technical substrates into structured verbal representations, making complex signals actionable and modelable.
• It is applied across domains—from think-aloud usability testing and visualization to financial forecasting and retrieval-augmented reasoning—to enhance diagnostic and predictive accuracy.
• The approach bridges perception, reasoning, and action by structurally grounding verbal intermediates, though it faces challenges like calibration, error propagation, and modality gaps.

Verbal Technical Analysis (VTA) denotes a family of methods in which technically salient structure is externalized into natural language and then used as a primary analytic object. In the literature, the term has both narrow and broad uses. In usability research, it refers to the systematic analysis of think-aloud verbalizations for detecting user-experience problems; in financial forecasting, it denotes a hybrid framework that reasons verbally over price-derived annotations and conditions a time-series backbone on reasoning-derived attributes. Closely related work extends the same general pattern to chart-grounded explanation, retrieval-augmented reasoning, process supervision by verbal critique, and the diagnosis of reasoning failures in voice systems. Taken together, these works suggest that VTA is best understood as a design pattern in which verbal intermediates are treated not as ancillary explanations but as structured, modelable, and operationally consequential technical signals (Fan et al., 2022, Koa et al., 6 Nov 2025, Park et al., 2 May 2026, Chen, 23 Apr 2026, Wang et al., 20 Apr 2025, Lin et al., 30 Sep 2025).

1. Conceptual scope and recurrent structure

Across domains, VTA has three recurrent components. First, a technical substrate is converted into a verbal representation: user actions become utterance segments, chart elements become cited semantic references, retrieved documents become Verbal Annotations, reasoning traces become critique targets, and OHLCV sequences become textual technical-analysis attributes. Second, the verbal representation is structured rather than free-form: categories, inline citations, step labels, relevance scores, or tagged outputs constrain its semantics. Third, the verbal representation is used instrumentally for diagnosis, guidance, ranking, evaluation, or prediction rather than merely for post hoc explanation (Fan et al., 2022, Wang et al., 20 Apr 2025, Park et al., 2 May 2026, Chen, 23 Apr 2026, Koa et al., 6 Nov 2025).

Domain	Verbal object	Operational role
Usability testing	Think-aloud utterance segments	UX problem detection
Visualization	Chart-grounded explanations with [cite:id]	Comprehension and reasoning support
RAG	Verbal Annotation $(v,s)$	Retrieval–reasoning bridge
LLM reasoning	Step-indexed verbal critique	Iterative refinement
Voice systems	Voice-native episodes and error taxonomies	Modality-gap diagnosis
Finance	Reasoning traces over textual annotations	Interpretable forecasting

This diversity also implies that VTA is not a single canonical algorithm. A plausible implication is that the common denominator is the elevation of verbal structure to a first-class computational interface between perception, reasoning, and action.

2. Think-aloud usability analysis as an empirical foundation

In usability testing, VTA is grounded in the analysis of think-aloud utterances. The central empirical reference is a study comparing 9 native English speakers and 9 Chinese non-native English speakers, all students or recent graduates at US universities, across three concurrent think-aloud protocols: Classic Think-Aloud (CTA), Speech Communication (SC), and Interactive Think-Aloud (ITA). The study used three websites with known UX issues, task-based testing, Zoom recording, Otter.ai transcription with manual correction, and segment-level coding into Cooke’s five categories: Procedure, Reading, Observation, Explanation, and Others. A binary problem label was assigned per segment, and underlying issues were aggregated into actual problems with severity rated לפי Nielsen’s five-level scheme (Fan et al., 2022).

The principal result is that the verbalization patterns previously reported for native English speakers generalize to Chinese non-native English speakers with intermediate or higher proficiency. Language group had no significant effect on category distributions, with $F(1,16) = 0.062$, $p = 0.806$, partial $\eta^2 = 0.004$, while protocol had an overall effect, $F(2,32) = 4.621$, $p = 0.017$, partial $\eta^2 = 0.223$, concentrated mainly in Explanation utterances. Category prevalence differed strongly, with $F(4,64) = 44.440$, $p < 0.0001$, partial $\eta^2 = 0.71$; partial eta squared was defined as $F(1,16) = 0.062$0 (Fan et al., 2022).

For VTA proper, the category–problem linkage is the key result. Observation utterances were the strongest indicator of UX problems, comprising 58.8% of problem-labeled segments for native speakers and 65.3% for non-native speakers. Procedure utterances were sometimes linked to problems, at 14.4% and 21.8% respectively, especially when phrased as questions or impasses. Explanation utterances were moderately linked to problems, at 14.4% for native speakers and 9.9% for non-native speakers, but became more informative under ITA, where they comprised 19.6% of problem segments versus 5.4% in CTA and 2.5% in SC. Reading had 0% association with problem encounters across all groups and protocols, making it an exclusion signal; Others were weak predictors unless they contained uncertainty or meta-comments about confusion or frustration (Fan et al., 2022).

The study’s problem counts reinforce the generalization claim. Native speakers produced 35 problem encounters and 15 actual problems; non-native speakers produced 41 encounters and 18 actual problems, with 14 common problems across groups and 5 unique problems that were all low severity. Protocol differences in actual problems were not significant. This supports three practical consequences stated in the study: recruit across proficiency levels of at least CEFR B1, favor CTA for scalable data collection because ITA’s extra finds were mostly low severity and require more moderator effort, and prioritize Observation while down-weighting Reading in AI-assisted analysis tools (Fan et al., 2022).

A methodological limitation is explicit. The coding relied on consensus but did not report inter-rater reliability statistics such as Cohen’s kappa or Krippendorff’s alpha. Any VTA implementation that aims at reproducible annotation or supervised learning inherits this issue unless reliability is measured directly.

3. Multimodal grounding: from AOIs to chart-centered dialogue

A second research line addresses a persistent weakness of purely verbal analysis: the difficulty of linking utterances to specific visual stimuli. In usability testing on scrollable website screenshots, synchronized gaze, mouse, and audio streams were used to connect spoken feedback to Areas of Interest (AOIs). Ten German-speaking participants viewed three city websites while answering prompts such as “What aspects do you find negative?”, “What stands out to you as positive?”, and “What elements do you find confusing?”. Eye tracking was recorded with a Tobii 4C Pro at 90 Hz, audio at 48 kHz, ASR was performed with Whisper (medium), and diarization with pyannote. AOIs were manually drawn, averaging 42 per page, and hit rates were defined as the percentage of verbally mentioned AOIs that were fixated by gaze or pointed by mouse during the utterance window (Murali et al., 2023).

The quantitative conclusion is unambiguous. Gaze achieved a mean hit rate of 66.2% (SE = 2.1), whereas mouse achieved 40.0% (SE = 5.6), with a paired-sample t-test of $F(1,16) = 0.062$1, $F(1,16) = 0.062$2. Positive and neutral statements had higher hit rates than negative statements for both modalities, with significant main effects of Sentiment, $F(1,16) = 0.062$3, $F(1,16) = 0.062$4, and Tracking method, $F(1,16) = 0.062$5, $F(1,16) = 0.062$6, but no interaction, $F(1,16) = 0.062$7, $F(1,16) = 0.062$8. Mouse behavior was dominated by scrolling, whereas gaze was more strongly associated with verbally mentioned content, supporting gaze-weighted multimodal VTA (Murali et al., 2023).

The same grounding problem appears in visualization research, but there it is framed as “mental bridging” between verbal explanation and chart elements. VizTA addresses this by representing each visual element as a 4-tuple $F(1,16) = 0.062$9 and by allowing query-side drag-and-drop references of the form [tag: [id: id_i, data: d_i]]. The agent responds with inline citations [cite: id_j], which are rendered as interactive labels that highlight the corresponding mark and show a tooltip on hover. Initialization includes chart specification, data description, chart knowledge $p = 0.806$0, chart data $p = 0.806$1, visual features, and an ID list $p = 0.806$2, enabling element-level and group-level grounding in box plots, density plots, violin plots, and quantile dot plots (Wang et al., 20 Apr 2025).

VizTA’s between-subject study with 24 participants showed higher SCQ correctness, with VizTA at $p = 0.806$3 and 95% CI $p = 0.806$4 versus Baseline at $p = 0.806$5 and 95% CI $p = 0.806$6, $p = 0.806$7, $p = 0.806$8. OEQ pass rate was $p = 0.806$9 for VizTA versus $\eta^2 = 0.004$0 for Baseline, $\eta^2 = 0.004$1, $\eta^2 = 0.004$2. VizTA users cited 150 meaningful chart-derived values versus 84 for Baseline, and inline citations were rated especially helpful for maintaining focus, with $\eta^2 = 0.004$3, $\eta^2 = 0.004$4 (Wang et al., 20 Apr 2025).

These works show two complementary forms of multimodal VTA. The usability line links verbalizations back to stimuli through synchronized behavioral traces; the visualization line makes the mapping explicit inside the dialogue interface itself. In both cases, the operational aim is the same: reduce ambiguity in reference and make verbal claims inspectable against concrete visual objects.

4. Retrieval, reasoning, and verbal supervision

In retrieval-augmented reasoning, VTA appears as an explicit bridge between retrieved evidence and downstream reasoning. Verbal-R3 replaces raw passage injection with Verbal Annotations: analytic narratives stating how a document relates to a query and whether it supports the answer. The Verbal Reranker takes $\eta^2 = 0.004$5 and returns $\eta^2 = 0.004$6, where $\eta^2 = 0.004$7 is the annotation and $\eta^2 = 0.004$8 is a pointwise relevance score; ties are broken by the score-token logit $\eta^2 = 0.004$9. Selected items are injected in a regularized format, "[Doc i] v_i (Relevance score: s_i)", within an <information> block. The Generator then interleaves reasoning, retrieval, and answering through <search> and <answer> tags (Park et al., 2 May 2026).

The reranker is trained by supervised fine-tuning on synthetic query–document–annotation triplets distilled from GPT-OSS-120B, with loss

$F(2,32) = 4.621$0

The Generator is optimized with GRPO under a hierarchical reward with $F(2,32) = 4.621$1, $F(2,32) = 4.621$2, $F(2,32) = 4.621$3, and hyperparameters $F(2,32) = 4.621$4, $F(2,32) = 4.621$5, $F(2,32) = 4.621$6. By default, retrieval uses $F(2,32) = 4.621$7 documents and keeps $F(2,32) = 4.621$8 annotated contexts (Park et al., 2 May 2026).

Empirically, Verbal Annotations improve context use rather than mere surface compatibility. With Search-R1 3B, average EM rises from 38.75 with naive injection to 41.92 with Verbal Annotations, whereas paraphrasing retrieved contexts yields 37.78. For Qwen2.5-3B, Verbal-R3 increases average EM/F1 from 38.75/46.31 to 45.36/54.66; for Qwen2.5-7B, from 41.90/50.10 to 48.30/57.28. Gains are larger in multi-hop than single-hop settings: for the 3B model, average F1 improves by 26.91% in multi-hop versus 9.67% in single-hop; for the 7B model, 20.26% versus 8.20%. Test-time scaling with $F(2,32) = 4.621$9 and $p = 0.017$0 reduces reranker calls by 45.2% for 3B and 53.8% for 7B compared with naive majority voting without unique-query extraction (Park et al., 2 May 2026).

A parallel development applies VTA to the internal steps of reasoning itself. Verbal Process Supervision (VPS) defines a generate–critique–refine loop in which a stronger supervisor produces step-indexed verbal critiques over a weaker actor’s chain-of-thought. Formally, the actor initializes with $p = 0.017$1, receives critique $p = 0.017$2, and refines via $p = 0.017$3, up to round budget $p = 0.017$4. The supervisor labels each step as Correct, Partially correct, or Incorrect, emits targeted minimal fixes, preserves correct steps, and returns CONVERGED when all steps and the final answer are correct (Chen, 23 Apr 2026).

VPS establishes critique granularity as a distinct axis of inference-time scaling. On GPQA Diamond, GPT-5.4 (High) | GPT-5.4 (Low) reaches 94.9% at $p = 0.017$5, surpassing the reported 94.1% state of the art without parameter updates. On AIME 2025, VPS lifts weak actors from 11.7–26.7% to 63.3–90.0%, with the GPT-5.4 Nano pair rising from 26.7% to 90.0% by $p = 0.017$6. At matched compute, VPS exceeds Reflexion by +8.5 points on GPQA, +10.0 on AIME, and +12.1 on LiveCodeBench V6; it exceeds Self-Consistency@5 by +5.0 points on GPQA and +8.3 points on LiveCodeBench. The gains correlate with supervisor headroom, with Pearson $p = 0.017$7 between $p = 0.017$8 and VPS improvement (Chen, 23 Apr 2026).

The contrast between Verbal-R3 and VPS is instructive. In Verbal-R3, verbal analysis is applied to external evidence; in VPS, it is applied to internal reasoning trajectories. Both systems nonetheless rely on the same principle: a structured verbal intermediate can localize technical relevance more effectively than unstructured context or scalar feedback.

5. Voice-interactive reasoning and modality-induced constraints

VERA extends VTA into the evaluation of real-time voice assistants. It is a benchmark of 2,931 voice-native episodes organized into five tracks—Math, Web, Science, Long-Context, and Factual—designed to preserve reasoning difficulty while adapting prompts for speech interaction. Voice systems receive spoken prompts at 24 kHz TTS; text baselines receive the same content in text form. Outputs are judged with an LLM-as-a-judge protocol using GPT-4o, each evaluated three times with majority vote over {Correct, Incorrect, Not Attempted}. Macro-averaged accuracy is defined as

$p = 0.017$9

and the Voice Reasoning Gap is formalized as

$\eta^2 = 0.223$0

Judge reliability was validated by 1,000 human spot-checks, with GPT-4o reaching 97.8% agreement with humans overall (Lin et al., 30 Sep 2025).

The central result is a large and consistent modality gap. The best text models reach 54.0% macro-averaged accuracy, whereas voice systems reach 11.3%; on competition mathematics, a leading text model achieves 74.8% while its voice counterpart reaches 6.1%. The paired track-level comparison between GPT-5 text and GPT-realtime voice yields gaps of 68.7 points on Math, 11.5 on Web, 29.2 on Science, 71.5 on Context, and 20.9 on Factual, with McNemar’s test confirming $\eta^2 = 0.223$1 and a macro-averaged difference of approximately 40.4 percentage points with 95% CI $\eta^2 = 0.223$2 (Lin et al., 30 Sep 2025).

Latency analysis shows why VERA matters for VTA. Time-To-First-Response (TTFR) reveals a low-latency plateau: voice systems at $\eta^2 = 0.223$3 s TTFR cluster near roughly 10% accuracy, and approaching text-level performance requires sacrificing real-time interaction. The decoupled cascade LiveAnswer, combining a GPT-5 reasoner with a fast narrator, improves macro accuracy to 27.0% but still remains far from text and incurs user-perceived TTFR of 10.50 s, with TSTT at 9.68 s, TTFRpartial at 0.83 s, and TLLM+TTS at 63.40 s. Increasing “thinking time” within a native voice model did not help: Audio Flamingo 3’s thinking mode increased latency from 2.40 s to 15.14 s while reducing macro accuracy from 1.7% to 1.5% (Lin et al., 30 Sep 2025).

VERA’s failure analysis is also explicitly verbal. It distinguishes knowledge errors such as unsupported_fact and entity_confusion, reasoning errors such as computation_error and logical_contradiction, and understanding/output errors such as off_target, no_final_answer, and fabricated_context. Native streaming systems tend toward fluent but incorrect answers; end-to-end speech models toward disengagement; cascades toward mis-grounding between reasoning and narration. This suggests that in voice systems, VTA is not only a mechanism for producing explanations but a framework for diagnosing the architectural tension between speaking in real time and reasoning reliably.

6. Financial VTA: verbal reasoning over time series

The explicit naming of VTA appears in financial forecasting, where it is defined as a framework that combines verbal reasoning with latent reasoning to produce accurate and interpretable stock time-series forecasts. The input is a historical window

$\eta^2 = 0.223$4

and the output is

$\eta^2 = 0.223$5

where $\eta^2 = 0.223$6 is the verbal reasoning trace and $\eta^2 = 0.223$7 are future adjusted closing prices. A textual annotator maps $\eta^2 = 0.223$8 to $\eta^2 = 0.223$9 using statistical descriptors and technical attributes such as SMA, EMA, Momentum, RSI, MACD, Williams %R, CCI, ADX, Bollinger Bands, and the Stochastic Oscillator (Koa et al., 6 Nov 2025).

The architecture has three parts. A reasoning LLM, based on Qwen2.5-7B-Instruct, receives prompts constructed from $F(4,64) = 44.440$0 and $F(4,64) = 44.440$1 and generates a reasoning trace and a predicted sequence. A forecasting backbone, based on GPT-2 repurposed for time-series forecasting, aligns time tokens with principal word embeddings via cross-attention and dual-branch transformers. Finally, joint conditional training derives descriptive attribute classes $F(4,64) = 44.440$2 from the reasoning output and conditions the backbone forecast with classifier-free guidance: $F(4,64) = 44.440$3 The classifier-free unconditioning probability is $F(4,64) = 44.440$4 and the guidance scale is $F(4,64) = 44.440$5 (Koa et al., 6 Nov 2025).

Optimization is driven by Time-GRPO, a GRPO-style RL objective using inverse-MSE reward. For a sampled forecast, the reward is given as either

$F(4,64) = 44.440$6

or

$F(4,64) = 44.440$7

combined with a format reward enforcing > ... tags and KL regularization. Training proceeds in three stages: cold-start RL, supervised fine-tuning on filtered low-MSE traces, and final RL refinement. The forecasting branch additionally uses feature-alignment and output-consistency losses, with total objective

$F(4,64) = 44.440$8

LoRA is used for parameter-efficient fine-tuning (Koa et al., 6 Nov 2025).

The reported empirical results are state of the art across U.S., Chinese, and European equities. Compared with CALF, VTA improves StockNet from MSE/MAE 0.0674/0.1738 to 0.0659/0.1701; China A50 from 0.2412/0.2843 to 0.2265/0.2737; EURO STOXX 50 from 0.0762/0.1957 to 0.0748/0.1929; and Dow Jones from 0.1092/0.2181 to 0.1040/0.2120. Aggregated across all data, MSE improves from 0.1235 to 0.1178 and MAE from 0.2180 to 0.2122. In ablations, cold-start RL yields a small average MSE improvement of about 1.6%, SFT markedly improves MSE, second-stage RL yields about 20.3% average MSE improvement over the base model, and conditioning delivers the best final performance (Koa et al., 6 Nov 2025).

Portfolio evaluation further ties verbal reasoning to economic utility. Using 10-day predicted returns with daily rebalancing, VTA attains returns of 0.2409, volatility of 0.1185, drawdown of −0.0883, and the top Sharpe ratio of 1.7190. Expert evaluation by 25 industry experts gives VTA the highest average scores on clarity (3.74), depth (4.04), accuracy (3.37), coherence (3.60), and relevance (3.71), above GPT-4.1 mini and DeepSeek-R1 on all five dimensions (Koa et al., 6 Nov 2025).

This financial formulation is narrower than the umbrella usage found elsewhere, but it crystallizes the strongest possible version of the VTA claim: verbal technical analysis can be optimized not only for interpretability but also for downstream predictive accuracy.

7. Limitations, ambiguities, and future directions

The literature assigns VTA a heterogeneous scope. In some work it is a direct term of art; in others it is a unifying interpretation imposed on adjacent methods. This suggests that VTA currently functions more as a research program than as a universally stabilized formalism. The benefit of this breadth is cross-domain transfer of design patterns; the cost is terminological ambiguity.

Several empirical limitations are explicit. The think-aloud usability study used a small sample of 18, a student population, a single moderator, three websites, and consensus coding without reported reliability statistics (Fan et al., 2022). The gaze–mouse linking study used only 10 participants, static scrollable screenshots rather than interactive websites, and did not report calibration details (Murali et al., 2023). VizTA currently covers four families of 1D distributional visualizations and notes that complex multi-view dashboards or non-standard encodings may require stronger grounding and knowledge integration (Wang et al., 20 Apr 2025). Verbal-R3 depends on two LLMs, can suffer error propagation from imperfect annotations, and was evaluated on offline corpora rather than live web retrieval (Park et al., 2 May 2026). VPS reports mostly single-run point estimates, depends on non-trivial supervisor headroom, and degrades when decisive errors are not linguistically expressible, as in parts of code synthesis (Chen, 23 Apr 2026). VERA shows that speech fidelity is not the main bottleneck in voice reasoning and that decoupled cascades introduce their own grounding and consistency errors under latency constraints (Lin et al., 30 Sep 2025). Financial VTA depends on heuristic indicator schemas, needs multi-stage training for strong gains, and remains vulnerable to regime shifts and news-driven market behavior (Koa et al., 6 Nov 2025).

The forward directions proposed by these works are notably convergent. Usability VTA calls for expansion to other non-native groups, retrospective think-aloud, moderator-effect analysis, and reliability measurement (Fan et al., 2022). Multimodal linkage work points toward DOM-driven AOIs, mobile settings, and remote webcam tracking (Murali et al., 2023). VizTA suggests richer visual feedback, multi-agent planning, and broader literacy studies (Wang et al., 20 Apr 2025). Verbal-R3 motivates stronger retrieval-quality evaluation and safeguards such as confidence scoring and human oversight (Park et al., 2 May 2026). VPS motivates hybrid verbal–executable supervision and adaptive stopping (Chen, 23 Apr 2026). VERA highlights asynchronous “thinker–talker” architectures with structured intermediate state and confidence-gated narration (Lin et al., 30 Sep 2025). Financial VTA points to multi-modal fusion with news, options, fundamentals, portfolio-level reasoning, and uncertainty-aware conditioning (Koa et al., 6 Nov 2025).

A plausible synthesis is that future VTA systems will be increasingly hybrid: verbally explicit, structurally constrained, multimodally grounded, and coupled to external checks or specialized backbones. In that form, VTA is less a single task than a general methodology for making technical intermediates both legible and operational.