---
title: 'CLARITI: Disambiguation in Research'
url: https://www.emergentmind.com/topics/clariti
type: topic
---

# CLARITI: Disambiguation in Research

Searching arXiv for the term and closely related variants to ground the article in the current literature.
CLARITI is not a single canonical research artifact in the current arXiv literature. Rather, the label appears as a spelling variant, acronymic shorthand, or stylistic near-homograph for several technically distinct systems: **CLARITY**, a clinical assistant for routing, inference, and triage in telehealth; **CLARITI**, an 8B clarification module for software engineering tasks; **CLARity**, a reinforcement-learning framework for improving reasoning consistency in expert LLMs trained on MCQs; and **CLARiTy**, a vision transformer for chest X-ray classification and weakly supervised localization [2510.02463] [2604.14624] [2510.09278] [2512.16700]. The term therefore requires contextual disambiguation: in medical dialogue systems it denotes a production clinical assistant, in software engineering it denotes a reward-trained clarification model, in LLM training it denotes a consistency-aware RL framework, and in medical imaging it denotes a class-localizing transformer architecture.

## 1. Terminological scope and disambiguation

Within the supplied literature, the string **“CLARITI”** is used in more than one way. In the clinical dialogue paper, the paper itself uses the name **CLARITY**, and the query form “CLARITI” is described as a spelling or acronym variant referring to the same system rather than to a separate module [2510.02463]. In the software-engineering paper, **CLARITI** is the canonical name and is expanded as **“CLARIficiation with TIered rewards”** [2604.14624]. In the chest X-ray paper, the query form “CLARITI” is explicitly tied to **CLARiTy**, expanded as **“Class Localizing and Attention Refining Image Transformer”** [2512.16700]. A further near-match, **CLARity**, names a reinforcement-learning framework for expert LLMs and expands to **“Consistency-aware Learning with Data-Augmented 2-Stage Reinforcement Strategy”** [2510.09278].

This naming convergence is substantive rather than merely orthographic. Each system operates in a different problem class, uses a different architecture, and is evaluated on different tasks. A plausible implication is that “CLARITI” functions in practice as an overloaded research label rather than a stable identifier for one model family.

A further source of confusion is **CLARO**, which is unrelated to these systems. CLARO is an 8-channel ASIC for the LHCb RICH upgrade and concerns high-rate photon readout rather than dialogue systems, clarification modules, or medical imaging [1511.09308].

## 2. CLARITY as a clinical assistant for routing, inference, and triage

**CLARITY** is a production-grade, AI-driven medical dialogue system deployed as a virtual clinical assistant on a large national inter-hospital telehealth platform [2510.02463]. Its stated goals are patient-to-specialist routing, structured clinical consultations, and severity assessment with emergency escalation. It is designed to replace part of the traditional “patient → general practitioner → specialist” pathway by collecting anamnesis, generating diagnostic hypotheses, returning a ranked specialist list with short explanations, and triggering emergency escalation when indicated.

Its architecture is explicitly hybrid. A **Finite State Machine (FSM)** serves as the central chat manager, while multiple LLM-based and ML-based microservices provide text generation, medical specialty selection, moderation, emergency detection, readiness estimation, and question detection [2510.02463]. The FSM is formally defined as a Mealy-type machine,
$$
M = (Q, \Sigma, \Omega, C, T, \delta, \lambda, q_0, q_d, q_{ca}, F),
$$
with bounded internal execution controlled by `N_ATTEMPTS`. High-level states include **Initialization**, **Information Collection**, **Diagnosis**, **Moderation**, **Emergency**, and **Free Dialogue**. This bounded state-machine design is described as central to controlling LLM behavior, because the LLM is always called within state-specific templates and bounded transitions rather than in an unconstrained conversational loop [2510.02463].

The information-collection subsystem combines a question generator, dialogue retrieval from a vector database, and a relevance classifier trained on 2,500 labeled examples. If cosine similarity to a historical dialogue exceeds 0.965, cached questions are reused; redundant questions are removed with a cosine cutoff of 0.86; and the relevance classifier attains precision 0.84 for the “relevant” label [2510.02463]. The routing core is the **Medical Specialty Selector**, which generates candidate diagnoses, maps each diagnosis to a specialty, and then produces brief explanations. Its evaluation uses pairwise precision and recall over multisets of specialist labels. Reported routing results are **Precision@1 = 77±0.3%**, **Recall@2 = 92±0.4%**, and **Recall@3 = 96±0.4%**, summarized in the paper as “77% precision for first recommendations” and “96% recall for top-3 recommendations” [2510.02463].

Triage is operationalized through an emergency detector and a hard-coded emergency state. The emergency detector uses the full chat and a feature pipeline combining TF-IDF, one-hot critical words, and a binary LLM criticality feature, followed by PCA with \(n_c = 423\) and histogram-based gradient boosting [2510.02463]. In deployment, the threshold is tuned to keep **FPR = 0.02**; reported performance is **Precision = 0.72**, **Recall = 0.49**, and **\(F_1 = 0.59\)**. When an emergency is flagged, the FSM transitions to a scripted protocol such as “Call 103 immediately,” preventing the LLM from overriding escalation [2510.02463].

The deployment scale is notable. The second pilot processed **55,856 dialogues**, with **92.3% progression after dialogue initiation**, **54.7%** of dialogues yielding diagnosis plus referral, and appointment conversions of **26.8%** for online consults and **5.6%** for offline consults [2510.02463]. The first pilot reported **1,534 initiated consultations**, **62.13%** ending with specialist recommendation plus diagnostic hypothesis, a mean appointment duration of **2:13 minutes**, and referrals to **70 different specialties**. The paper further reports that CLARITY is **up to 3× shorter** than a human GP consultation and describes its first-attempt routing precision as surpassing typical human-level performance in initial routing [2510.02463].

## 3. CLARITI as a reward-driven clarification module for software engineering

In software engineering, **CLARITI** denotes a learned clarification module rather than a coding agent [2604.14624]. It sits upstream of a fixed downstream coding system, receives an underspecified issue, generates a small set of clarification questions, obtains answers from a simulated user, and passes the enriched context to **Seed OSS 36B in OpenHands**. The model itself is an **8B-parameter** LLM trained specifically for question generation in software engineering tasks [2604.14624].

The motivation is that real software issues are underspecified along several axes, but missing information is not uniformly valuable. The paper defines an **information need** as information absent from the issue description that is required for an agent to produce a correct patch without relying on unverifiable assumptions [2604.14624]. From expert annotations on SWE-Bench Verified, six categories are used: **Error Information**, **Expected Behavior**, **Implementation Details**, **External References**, **Reproduction Steps**, and **Version/Environment**. Clarification is formulated as a **single-turn, set-generation problem**: the model outputs a set of questions all at once rather than engaging in multi-turn ask–answer–ask interaction [2604.14624].

The paper’s empirical design is driven by two analyses. First, it uses **Shapley attribution** on 1,500 underspecified variants of 500 SWE-Bench Verified issues to quantify which information categories most affect downstream success. Mean absolute SHAP values establish an impact hierarchy led by **Error Information** at approximately **0.183**, followed by **Implementation Details** at approximately **0.0982**, with the remaining categories lower but non-zero [2604.14624]. Second, it studies question answerability by comparing generated questions against both underspecified and fully specified issue descriptions. Distributional analysis yields four strategic themes associated with answerable questions: **Ground in Evidence**, **Demand Specificity**, **Minimize Scope**, and **Ensure Actionability** [2604.14624].

These findings are operationalized as a **four-stage reward pipeline** trained with **Group Relative Policy Optimization (GRPO)** on top of **Qwen3-8B**, with a frozen SFT checkpoint as reference model and **Qwen3-32B** as judge [2604.14624]. The four reward components are:
- **Non-redundancy**
- **Diversity**
- **Answerability**
- **Task relevance**

The final reward is
$$
r_{\text{final}} = \frac{1}{4}\Big( r_{\text{non-redundancy}} + r_{\text{diversity}} + r_{\text{answerability}} + r_{\text{relevance}} \Big).
$$
Rejection filtering is applied stage-wise: if a generation fails Stage 1, 2, or 3, later stages are not evaluated and the total reward is set to zero [2604.14624]. Relevance is category-weighted according to the SHAP-derived impact hierarchy; answerability is estimated by asking whether the question can be answered from the fully specified original issue.

The main downstream results are reported on **250 underspecified issues**. **No clarification** yields **22.4%** success, **GPT-5 nano** reaches **29.6%**, **GPT-5** reaches **35.6%** with **5.1 questions per issue**, **CLARITI** reaches **36.8%** with **3.0 questions per issue**, and the **full issue** upper bound is **41.6%** [2604.14624]. The paper therefore reports that CLARITI matches or slightly exceeds the GPT-5 clarification baseline while using **41% fewer questions**, and that it recovers approximately **88%** of the fully specified performance ceiling. Its answerability score is **0.373** and its relevance score is **0.622**, both measured on clarification sets [2604.14624].

Category allocation explains much of this efficiency. CLARITI assigns **26.4%** of its questions to **Error Information**, compared with **10.2%** for GPT-5 and **6.0%** for GPT-5 nano [2604.14624]. The ablations are also strong: **SFT only** yields **24.2%** success, removing the answerability stage yields **24.4%**, removing the relevance stage yields **32.0%**, and the full four-stage system reaches **36.8%** [2604.14624]. The paper interprets this as evidence that both answerability and relevance are necessary; high-relevance but unanswerable questions degrade performance, and merely plausible question generation without RL is insufficient.

## 4. CLARity as a consistency-aware RL framework for expert LLMs

A different system, spelled **CLARity**, is a reinforcement-learning framework for training expert LLMs in scarce-data domains such as law and medicine [2510.09278]. Its starting point is the observation that standard outcome-based RL on multiple-choice questions can improve answer accuracy while degrading reasoning quality, especially the consistency between chain-of-thought and final answer. On the legal MCQ dataset **JEC-QA**, the paper reports that after standard GRPO-based RL, the proportion of responses with logical fallacies or inconsistencies rises from **7%** to **31%** [2510.09278].

The central construct is **reasoning consistency**: whether the options endorsed in the model’s reasoning match the options chosen in its final answer [2510.09278]. To exploit this, CLARity uses a small general-purpose LLM as a parser rather than as a domain expert. The expert model outputs `<think> ... </think>` and `<answer> ... </answer>`, and a separate **Qwen2.5-7B-Instruct** instance parses the reasoning into a set such as `{A}`, `{ABD}`, or `{}`. The core consistency reward is
$$
R_{\text{consistency}} = \mathbb{I}_{\text{Ans}_{\text{parsed}} = \text{Ans}_{\text{gen}}},
$$
while answer correctness is
$$
R_{\text{ans}} = \mathbb{I}_{\text{Ans}_{\text{gen}} = A_i}.
$$
In Stage 2, the implementation uses **0** for consistent outputs and **\(-1\)** for inconsistent ones, plus a similar small penalty for malformed format [2510.09278].

The training pipeline is explicitly **two-stage**. **Stage 1 (Refine)** shapes outputs into explicit option-wise reasoning using only a format reward and a structure reward; it does not optimize correctness [2510.09278]. **Stage 2 (Monitor)** drops the rigid per-option structural constraint, retains the `<think>/<answer>` format, and adds consistency and answer-correctness rewards:
$$
R = 0.1 R_{\text{format}} + 1.0 R_{\text{consistency}} + 1.0 R_{\text{ans}}.
$$
This is combined with **dynamic data reformulation**, which deconstructs easy MCQs into propositions, rewrites them with **DeepSeek-V3**, and regroups them into harder composite MCQs according to a curriculum based on estimated pass rate [2510.09278].

The framework is evaluated with **Qwen2.5-3B-Instruct** and **Qwen2.5-7B-Instruct** as expert models on **JEC-QA** and **MedQA-USMLE** [2510.09278]. The primary metrics are **Accuracy (Acc)**, **Consistency ratio (Cons%)**, and **Acc\(^+\)**, the fraction of responses that are both correct and consistent. Averaged over law and medical MCQs, the reported values for the 7B model are:

| Method | Acc\(^+\) | Cons% | Acc |
|---|---:|---:|---:|
| Vanilla Qwen2.5-7B | 35.5 | 90.5 | 39.4 |
| Standard MCQ RL | 51.4 | 75.3 | 61.1 |
| PRM RL | 51.1 | 82.3 | 57.8 |
| CLARity\(_{7B}\) | 58.9 | 91.8 | 61.3 |

These results support the paper’s headline claim that CLARity improves **consistency by 16.5 points** and **Acc\(^+\)** by **7.5 points** over standard MCQ RL while maintaining essentially the same high accuracy [2510.09278]. The paper further reports that, on validation variants involving iterative judgment and shuffled MCQs, **CLARity\(_{7B}\)** achieves average **Acc\(^+\) = 52.0** and **Cons% = 94.6**, close to **GPT-4o** at **52.2** and **95.2**, respectively [2510.09278]. Human evaluations also show gains in correctness, professionalism, and readability, despite supervision being based only on consistency [2510.09278].

A notable practical claim is cost-effectiveness. The reward model is a general-purpose **7B** model with no domain-specific fine-tuning. Very small reward models perform poorly: reported inconsistency detection rates against a DeepSeek-V3 ground truth are **21%** for **Qwen2.5-0.5B**, **40%** for **Qwen2.5-1.5B**, and **94%** for **Qwen2.5-7B** [2510.09278]. This suggests a minimum viable reward-model capacity, but still at substantially lower cost than large PRM-style supervision.

## 5. CLARiTy as a vision transformer for chest X-ray pathology classification and localization

In medical imaging, **CLARiTy** is a **pure Vision Transformer–based architecture** for **multi-label classification of 14 thoracic pathologies** and **weakly supervised localization of 8 pathologies** on chest radiographs [2512.16700]. It is designed for the common setting in which image-level labels are abundant but region-level annotations are scarce. The model’s principal idea is to use ViT self-attention as a class-specific localization signal and then refine it via a dedicated module, **SegmentCAM**, using explicit anatomical priors [2512.16700].

The backbone is **ViT-S/16** with **12 transformer blocks**, **6 attention heads per layer**, **embedding dimension \(D = 384\)**, and **patch size \(16 \times 16\)** [2512.16700]. Two configurations are reported:
- **CLARiTy-S-16-224**: input \(224 \times 224\), patch grid \(14 \times 14\)
- **CLARiTy-S-16-512**: input \(512 \times 512\), patch grid \(32 \times 32\)

Instead of a single CLS token, CLARiTy uses **14 class-specific tokens**, one per pathology [2512.16700]. The self-attention matrix is partitioned so that the **class→patch** block yields class-specific attention maps, which are averaged over the last **\(p = 8\)** layers:
$$
\boldsymbol{A}^{(i,j,c)} = \frac{1}{p}\sum_{l}^{p}\boldsymbol{A}_l^{(i,j,c)}.
$$
At the output, class-token logits \(\boldsymbol{\ell}_C\) are combined with SegmentCAM foreground logits \(\boldsymbol{\ell}_f\) by averaging:
$$
\hat{\boldsymbol{\ell}} = \tfrac{1}{2}(\boldsymbol{\ell}_C + \boldsymbol{\ell}_f).
$$

**SegmentCAM** operates on the final patch tokens and has two roles: foreground segmentation and background activation suppression [2512.16700]. It produces heatmaps \(\boldsymbol{H}\), feature maps \(\boldsymbol{F}\), and classwise foreground masks \(\boldsymbol{S}_f\). Three losses constrain the heatmaps using automatically generated anatomical masks: **mask proximity** \(\mathcal{L}_{\text{PRX}}\), **mask confinement** \(\mathcal{L}_{\text{CNF}}\), and **foreground area** \(\mathcal{L}_{\text{AC}}\). A background activation suppression term,
$$
\mathcal{L}_{\mathrm{BAS}} = \frac{1}{\max\left(\sum_c y^{(c)}, 1\right)} \sum_c y^{(c)} \frac{ \operatorname{softplus}(\ell_b^{(c)}) }{ \operatorname{softplus}(\ell_P^{(c)}) + \varepsilon },
$$
penalizes signal retained in background-only features for positive classes [2512.16700]. The model also uses an **orthogonal class token (OCT) loss** over the last **\(q = 8\)** transformer layers and a learned **attention pooling** head instead of global average pooling. The full objective combines class-token classification loss, CAM losses, and OCT regularization [2512.16700].

Training uses **NIH ChestX-ray14** with **112,120 frontal CXRs**, **30,805 patients**, and 14 pathology labels [2512.16700]. Bounding boxes exist for **880 images** and **984 boxes** across 8 pathologies, used only for localization evaluation [2512.16700]. **CLARiTy-S-16-224** is initialized from **DINO** self-supervised pretraining, while **CLARiTy-S-16-512** is initialized from the 224 model with interpolated positional embeddings. The system is further distilled from a **ConvNeXtV2-B** teacher, whose reported NIH validation Macro AUC is **0.833** at 224 and **0.850** at 512 [2512.16700].

The main test results separate classification from localization. For classification, **Macro AUC** is **0.799** for **CLARiTy-S-16-224** and **0.818** for **CLARiTy-S-16-512** [2512.16700]. For localization on the NIH localization test subset, **Macro IoU Accuracy@0.5** is **0.303** for the 224 model and **0.318** for the 512 model [2512.16700]. The paper highlights that the **50.7%** gain in the abstract refers to the relative improvement of **CLARiTy-S-16-224** over **ThoraX-PriorNet-224** at IoU 0.5, and that **CLARiTy-S-16-512** improves over **ThoraX-PriorNet-512** by **46.5%** relative at the same threshold [2512.16700]. Particularly large gains are reported for small pathologies such as nodules and masses: at IoU 0.5, the 512 model reaches **0.393** for nodules and **0.419** for masses [2512.16700].

The computational footprint remains moderate for the lower-resolution model. **CLARiTy-S-16-224** has **27.03 M parameters** and **5.99 GFLOPs**, whereas **CLARiTy-S-16-512** has **27.35 M parameters** and **37.84 GFLOPs** [2512.16700]. The paper explicitly positions the 224 variant as suitable for lower-resource settings while still surpassing baselines decisively in localization [2512.16700].

## 6. Cross-cutting patterns, distinctions, and common misconceptions

Although these systems share a near-identical name, their methodological cores are different. **CLARITY** uses a **hybrid FSM + LLM + classical ML microservices** design for safe medical dialogue orchestration [2510.02463]. **CLARITI** in software engineering uses a **single-turn clarification policy** trained with **tiered intrinsic rewards** derived from empirical analyses of information impact and user answerability [2604.14624]. **CLARity** for expert LLM training uses **consistency-aware RL** plus **dynamic MCQ reformulation** to prevent the collapse of reasoning quality under answer-only rewards [2510.09278]. **CLARiTy** in radiology is a **pure ViT** augmented with **multiple class tokens**, **SegmentCAM**, **anatomical priors**, and **distillation** for classification and weakly supervised localization [2512.16700].

There are, however, recurring design motifs. One is the use of **structured constraints around a flexible model**. In CLARITY, the FSM bounds LLM behavior [2510.02463]. In software-engineering CLARITI, the four-stage reward pipeline with rejection filtering constrains question generation before optimizing task relevance [2604.14624]. In CLARity, the refine-then-monitor pipeline constrains output structure before applying consistency-based supervision [2510.09278]. In CLARiTy, anatomical priors and foreground-background decomposition constrain what kinds of activations are accepted as plausible localization [2512.16700]. This suggests a shared design philosophy: expressive models are embedded within explicit control structures rather than left fully unconstrained.

A second recurring motif is **efficiency under scarce or weak supervision**. CLARITY relies on 2,500 expert-annotated dialogues yet is deployed at national scale [2510.02463]. Software-engineering CLARITI uses intrinsic proxies instead of downstream outcome reward during RL because full coding-agent rollouts are too expensive [2604.14624]. CLARity avoids full process reward models and instead supervises only reasoning consistency with a small reward model [2510.09278]. CLARiTy trains localization without pathology bounding boxes, using only image-level labels plus automatically generated anatomical masks [2512.16700]. A plausible implication is that the shared naming convergence is accompanied by a shared engineering bias toward extracting high-value supervision from limited annotation budgets.

A common misconception would be to treat “CLARITI” as one research platform spanning telehealth, software engineering, LLM reasoning, and radiology. The literature does not support that interpretation. These are separate artifacts with distinct problem formulations, datasets, and evaluation protocols [2510.02463] [2604.14624] [2510.09278] [2512.16700]. The only safe use of the label is therefore context-specific: in a telehealth setting it most likely refers to **CLARITY**, in a software-engineering clarification setting to **CLARITI**, in reasoning-focused RL to **CLARity**, and in chest radiography to **CLARiTy**.

Source: https://www.emergentmind.com/topics/clariti