---
title: 'XtraGPT: Dual AI Systems'
url: https://www.emergentmind.com/topics/xtragpt
type: topic
---

# XtraGPT: Dual AI Systems

XtraGPT comprises two unrelated systems introduced under the same name in distinct domains: (1) a suite of open-source large language models specialized for controllable academic paper revision [2505.11336], and (2) an open-source Excel add-in that leverages ChatGPT for automated formula generation and verification [2309.00120]. Both are designed to augment human expertise by integrating LLM-based reasoning into complex, practice-driven workflows.

## 1. Human–AI Collaboration for Academic Paper Revision

XtraGPT [2505.11336] addresses the substantial limitations of general-purpose LLMs in high-fidelity scientific writing. Standard LLMs exhibit significant deficiencies, including a tendency to focus on surface-level linguistic improvements while neglecting logical structure, conceptual coherence, and cross-sectional argumentation. They also lack mechanisms for managing the multi-round, iterative feedback processes central to academic drafting, with each prompt treated in isolation and no memory of prior edits.

To overcome these constraints, XtraGPT implements a controllable, section-level paper revision paradigm. The system formulates each revision task as a 4-tuple $(q, T, p, \hat p)$, where $q$ is a high-level user instruction, $T$ the full paper context, $p$ the target paragraph, and $\hat p$ the revised paragraph. Revision control is operationalized via a canonical set of 20 section-level criteria $C$, covering all major paper sections (title, abstract, introduction, background, evaluation, conclusion) and encompassing priorities such as “Strength and Clarity of Motivation” and “Experimental Setup Clarity and Reproducibility.”

## 2. Dataset Construction and Quality Protocols

The XtraQA corpus comprises 7,040 ICLR 2024 submissions (filtered to 6,994 after deduplication and length capping), yielding 140,800 instruction–revision pairs. For each instance, a section, criterion, and paragraph are selected, an edit instruction $q$ is generated, and the paragraph is revised by GPT-4o-Mini to produce $\hat p$. A held-out test set (5% of papers; 7,000 pairs) supports benchmarking.

Quality control involved human annotation by three PhD-level evaluators across four axes: Instruction Following, Criteria Alignment, In-Context Reference, and Revision Acceptance (1–5 scale). Aggregated scores for GPT-4o-Mini edits were all $\geq 3.0$, indicating sufficient reliability for downstream model development.

## 3. Model Architecture, Objectives, and Inference

The XtraGPT model family consists of decoder-only transformers adapted from the following open-source LLMs:

| Model          | Backbone                  | Parameter Count |
|----------------|--------------------------|-----------------|
| XtraGPT-1.5B   | Qwen-2.5-1.5B-Instruct   | 1.5B            |
| XtraGPT-3.8B   | phi3.5-3.8B              | 3.8B            |
| XtraGPT-7B     | Qwen-2.5-7B-Instruct     | 7B              |
| XtraGPT-14B    | phi4-14B                 | 14B             |

All utilize a single-stream [q; T; p] input, with standard transformer block designs (rotary embeddings, multi-head attention) and up to 16K token contexts. There is no encoder; context and instruction are directly prepended.

Post-training—termed Controllable Post-Training (CPT)—employs a standard maximum-likelihood objective over $(q, T, p, \hat p)$ demonstration pairs:
$$
\mathcal{L}_{\mathrm{CPT}}(\theta) = -\,\mathbb{E}_{(q, T, p, \hat p) \sim \mathcal{D}_{CPT}}\left[\log P_\theta(\hat p \mid q, T, p)\right]
$$
Although $q$ and $\hat p$ are implicitly tied to criterion $c \in C$, no auxiliary losses (e.g., explicit coherence regularization) are used.

Inference supports both criterion-guided and free-form revision instructions, with system-level constraints enforcing scope (revise only the user-selected paragraph) and maximum output length (512 tokens). Iterative, user-driven submission cycles allow multi-round revision, but with each invocation contextually isolated (no persistent state).

## 4. Evaluation Strategies and Benchmarks

Automated evaluation relies on the Length-Controlled Win Rate (LC-Win Rate), with a modified LLM-as-Judge protocol (alpaca_eval_gpt4_turbo_fn) to debias for response verbosity. The winrate is formalized as:
$$
q_{\theta, \phi, \psi}(m \succ M) = \operatorname{logistic} \left[(\theta_m - \theta_M) + \phi_{M, b} \tanh\left(\frac{\operatorname{len}(z_m) - \operatorname{len}(z_M)}{\operatorname{std}(\ldots)}\right)\right]
$$
$$
\operatorname{winrate}^{LC}(m, M) = 100 \cdot \mathbb{E}_x\left[q_{\theta, \phi, \psi}(m \succ M \mid x)\right]
$$
Comparisons encompass GPT-4o-Mini, GPT-3.5-Turbo, Qwen-2.5-*xB*, Phi-4, DeepSeek-v3, and Llama-3.

Section-wise LC-Win Rates demonstrate that XtraGPT-7B outperforms all same-scale baselines and surpasses GPT-4o-Mini in the Abstract, Evaluation, and Conclusion sections. 

Human evaluation on a held-out set (300 paragraphs) via pairwise expert rating yields mean scores of $\sim 3.7/5$ for Instruction/Criteria, $\sim 3.4/5$ In-Context Reference, and $\sim 3.2/5$ for Revision Acceptance, all above neutral threshold. Case studies establish XtraGPT’s superiority in targeted, context-consistent edits.

## 5. Quantitative Results and Impact

XtraGPT-7B achieves a 55.5% overall LC-Win Rate, compared to GPT-4o-Mini’s 51.8% and Qwen-2.5-7B’s 40.8%. At the 14B-parameter scale, XtraGPT maintains this lead, with similar advantages over proprietary systems. For each backbone scale (1.5B to 14B), XtraGPT outperforms its base model by 10–30 points on LC-Win Rate.

“AI-Scientist” full-paper quality reassessment on 54 real ICLR submissions confirms that post-revision drafts exhibit marked improvements: Contribution (+0.23/4, +7.9%), Presentation (+0.28/4, +12.5%), Soundness (+0.19/4, +6.4%), and Overall rating (+0.65/10, +10.8%), all with $p < 0.01$.

## 6. Limitations and Prospects

The primary limitations of XtraGPT [2505.11336] are domain specificity (exclusive training on AI/ML papers), generator bias due to the use of GPT-4o-Mini for revision pair synthesis, and contextual isolation—no persistent revision state or full-paper consistency enforcement. Evaluation protocols depend on LLM-based judges and AI-Scientist scoring, which remain only partial proxies for actual peer review.

Future work targets multi-round dialogue interfaces with explicit change tracking, coverage expansion to further scientific domains, introduction of auxiliary objectives (e.g., global coherence, citation consistency), and improved automated metrics for holistic paper quality.

## 7. XtraGPT for Excel: Formula Generation Add-in

A separate instantiation of “XtraGPT” is documented as an open-source Excel extension enabling users to prompt ChatGPT for formula construction, explanation, and test-case suggestion directly within the worksheet environment [2309.00120]. The architecture integrates a ribbon button, an Office.js-based taskpane interface, and interaction with the OpenAI API. The add-in leverages system and user prompt engineering to elicit precise, testable responses, automatically injects formulas and explanations, and logs/catches all errors.

Workflow best practices explicitly adopt O’Beirne’s “trust but verify” paradigm: every ChatGPT-offered formula is unit-tested, checked for edge cases, and logged for reproducibility. The tool targets rapid prototyping but enforces systematic error-handling and human-in-the-loop verification, particularly for multi-criteria or nontrivial calculations.

Deployment requires Office 365 with support for sideloaded add-ins and Node.js for local hosting. The system is distributable via manifest.xml and can be centrally managed through administrative catalogs. Maintenance recommendations emphasize version-pinning of system prompts and thorough testing in controlled environments before production use.

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