💬 NLP

Think Before you Write: QA-Guided Reasoning for Character Descriptions in Books

This paper proposes a training framework that decouples reasoning from generation by using a QA-guided reasoning model to produce structured traces that condition a separate generation model, thereby improving the faithfulness, informativeness, and grounding of character descriptions in long-form narratives compared to standard reasoning-enabled LLMs.

Original authors: Argyrios Papoudakis, Mirella Lapata, Frank Keller

Published 2026-04-14

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Argyrios Papoudakis, Mirella Lapata, Frank Keller

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to write a biography of a character from a 1,000-page novel. You have the whole book in front of you, but your brain (or the AI you're using) is like a student who gets overwhelmed when reading too much at once. They might start hallucinating facts, forgetting who the villain is, or mixing up the timeline.

This paper, "QA-Guided Reasoning for Character Descriptions in Books," proposes a clever solution to this problem. It suggests that instead of asking the AI to "think" and "write" at the same time, we should force it to take notes first and write later.

Here is the breakdown using simple analogies:

1. The Problem: The "Overconfident Student"

The authors discovered something surprising. When they asked modern AI models to read a whole book and describe a character, the models actually performed worse when they were allowed to use their built-in "thinking" mode (where the AI talks to itself before answering).

The Analogy: Imagine a student taking a final exam. If you tell them, "Just write down whatever comes to mind," they might write a coherent but factually wrong essay because they are guessing. But if you tell them, "Stop! Write down a list of facts you remember from the text before you write the essay," they do much better.
The Finding: The AI's default "thinking" process often leads to hallucinations (making things up) or getting lost in the details of the long text. An empty "thinking" trace (just jumping straight to the answer) was sometimes more accurate than a messy, self-generated thought process.

2. The Solution: The "Detective's Case File"

The authors created a two-step system called QA-Guided Reasoning. Instead of one AI doing everything, they split the job into two roles:

Step A: The Detective (The Reasoning Model)

First, a specialized AI acts like a detective reading the book page by page. Instead of writing a story, its only job is to fill out a Case File in the form of Questions and Answers (QA).

Question: "Who is the main character's best friend?"
Answer: "Character Y."
Question: "What is the character's biggest fear?"
Answer: "Being abandoned."

The detective doesn't write the biography yet. It just gathers the hard facts and organizes them into a structured list. This is the "QA Trace."

Step B: The Biographer (The Generation Model)

Once the Case File is complete, a second AI (the Biographer) picks it up. It doesn't need to read the whole 1,000-page book again. It just looks at the Case File (the Questions and Answers) and writes the final character description based only on those verified facts.

The Analogy: Think of the Detective as a researcher who gathers all the ingredients (facts) and puts them in a bowl. The Biographer is the chef who uses that bowl to cook the meal (the description). The chef doesn't need to go to the farm to find the vegetables; they just trust the researcher's list.

3. Why This Works Better

The paper tested this on two massive datasets of books (BookWorm and CroSS) and found that this "Detective then Biographer" approach was superior to standard methods for three reasons:

Faithfulness (Sticking to the Truth): Because the Detective has to prove its answers with evidence from the text before writing the final story, the Biographer can't make things up. It's like a lawyer who can't argue a point in court unless they have a witness to back it up.
Informativeness (Getting the Details Right): The structured Questions force the AI to look for specific things (like relationships, personality traits, and events) that it might otherwise skip over in a long text.
Grounding (Not Getting Lost): Long books are like a maze. The QA Trace acts as a map. The Biographer follows the map (the QA list) rather than wandering aimlessly through the maze of the full text.

4. The Secret Sauce: Training the Detective

How do you teach the Detective to be good at this? You can't just say "be accurate" because there are many ways to be accurate.

The authors used a technique called GRPO (Group Relative Policy Optimization).
The Analogy: Imagine a coach training a detective. Instead of grading the final biography (which is hard to grade objectively), the coach grades the Case File. If the detective's list of Questions and Answers matches the "Gold Standard" facts from the book, they get a reward. If the list is missing facts or includes made-up ones, they get a penalty.
Once the Detective is trained to make perfect Case Files, the Biographer can easily write perfect descriptions.

Summary

In the past, we asked AI to "read a book and tell me about the character." The AI would often get confused or make things up.

This paper says: "Don't ask the AI to think and write at the same time."
Instead, tell it: "First, act like a detective and list the facts in a Q&A format. Then, act like a writer and use that list to tell the story."

This simple separation of duties makes the AI much smarter, more accurate, and less likely to lie about the characters in the books it reads.

`).
* Produces the final free-form character description.

B. Training Strategy

The framework uses a unique training approach to avoid the difficulty of rewarding open-ended generation:

Reasoning Model Training (GRPO):
- The authors do not have gold-standard reasoning traces. Instead, they derive silver-standard QA traces from gold-standard character descriptions.
- They use Group Relative Policy Optimization (GRPO) to optimize $R_\phi$ .
- Reward Signal: An LLM-as-a-judge computes Precision and Recall by comparing generated QA pairs against the gold-standard description. The F1 score serves as the reward. This allows direct optimization of the reasoner without needing a reward model for the final generated text.
Generation Model Training (SFT or Zero-Shot):
- The generator can be fine-tuned via Supervised Fine-Tuning (SFT) on tuples of (Character, Context, QA Trace, Gold Description).
- Alternatively, it can operate in zero-shot mode, conditioned on the trace produced by the trained reasoner.

C. Integration with Long-Context Strategies

The approach is model-agnostic and can be layered over various long-context processing methods:

Retrieval-based: Run reasoning on retrieved chunks.
Hierarchical: Generate traces for chunks, then merge intermediate descriptions.
Incremental: Update descriptions chunk-by-chunk using the new trace.
Direct Long-Context: Inject the concatenated trace into a long-context LLM.

3. Key Contributions

Empirical Finding: Demonstrated that enabling built-in reasoning in current LLMs often reduces faithfulness in character description generation, whereas empty traces yield more accurate outputs.
QA-Guided Framework: Introduced a modular approach that decouples reasoning (via structured QA traces) from generation, forcing the model to consolidate evidence before writing.
Training Innovation: Proposed a GRPO-based training strategy for the reasoning model using automatically derived QA traces and gold descriptions, bypassing the need for complex reward design on open-ended outputs.
Generalizability: Showed that the method improves performance across different long-context strategies (retrieval, hierarchical, incremental) and transfers effectively to out-of-distribution datasets.

4. Experimental Results

The method was evaluated on two datasets: BookWorm (Project Gutenberg novels/plays) and CroSS (recently published novels).

Faithfulness and Grounding: QA-guided reasoning consistently improved PRISMA (factual accuracy), NLI (grounding in the text), and Entity-Mention F1 (coverage) compared to baselines.
- Example: On BookWorm with Coref-32k, guided-QA improved QA F1 from 13.96 to 15.01 and PRISMA from 19.11 to 19.49 (with GRPO).
Built-in Reasoning vs. Guided Reasoning:
- Built-in reasoning consistently lowered metrics (e.g., NLI dropped from 52.58 to 49.94 in Coref-32k).
- Guided-QA (especially with GRPO-trained traces) reversed this trend, providing significant gains.
Trace Quality Matters: GRPO-optimized traces yielded better downstream descriptions than zero-shot traces, confirming that optimizing the reasoning step directly improves the final output.
Comparison with Proprietary Models: A smaller open-source model (Qwen-3-8b) equipped with guided reasoning achieved better entity coverage (EntMent) than a much larger proprietary model (GPT-4.1-mini) with full context access, highlighting the efficiency of structured reasoning.
Ablation Studies: Removing components like explanations or question types degraded performance, proving that the structured format (Question + Explanation + Answer + Type) is essential for grounding.

5. Significance

Paradigm Shift: Challenges the assumption that "more reasoning" (via built-in CoT) is always better for generative tasks. It suggests that for open-ended generation, structured, explicit reasoning is superior to unstructured internal deliberation.
Scalable Training: Offers a practical solution for training reasoning capabilities in long-context scenarios where reward signals are hard to define, by leveraging intermediate structured outputs (QA pairs) as a proxy.
Practical Application: Provides a robust method for improving character understanding in AI applications, enabling smaller models to compete with larger ones by better utilizing context and evidence.

In conclusion, the paper establishes that decoupling reasoning into a structured QA trace is a superior strategy for long-form character description, outperforming both standard generation and built-in reasoning modes by ensuring the model explicitly grounds its output in verifiable evidence before generating the final narrative.