💬 NLP

Co-FactChecker: A Framework for Human-AI Collaborative Claim Verification Using Large Reasoning Models

The paper introduces Co-FactChecker, a human-AI collaborative framework for claim verification that enhances Large Reasoning Models by translating expert feedback into direct modifications of the model's thinking trace, thereby outperforming traditional multi-turn dialogue approaches in reasoning quality and interpretability.

Original authors: Dhruv Sahnan, Subhabrata Dutta, Tanmoy Chakraborty, Preslav Nakov, Iryna Gurevych

Published 2026-04-16

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Dhruv Sahnan, Subhabrata Dutta, Tanmoy Chakraborty, Preslav Nakov, Iryna Gurevych

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 solve a very tricky puzzle, like figuring out if a rumor about a celebrity is true or false. You have a super-smart robot assistant (an AI) who is incredibly fast at reading books and searching the internet. However, this robot has a blind spot: it doesn't really "get" human context, nuance, or the messy reality of how news spreads. It often jumps to conclusions based on what it reads, missing the bigger picture.

This paper introduces CO-FACTCHECKER, a new way for humans and AI to work together to solve these puzzles. It's like moving from a game of "Telephone" to a game of "Shared Whiteboard."

Here is the breakdown using simple analogies:

1. The Problem: The "Telephone" Game

In the old way of doing things (called Multi-Turn Dialogue), you talk to the AI like you are chatting on a phone.

You say: "Hey, that's not right. They didn't ban all guns, just assault weapons."
The AI replies: "Okay, I see. Let me think again..."
The Problem: The AI often forgets what you said five minutes ago. It might get confused, repeat its mistakes, or give you a new answer that ignores your previous corrections. It's like playing a game of Telephone where the message gets garbled every time it passes from person to person. The AI also tends to "hallucinate" (make things up) because it's trying to guess what you want rather than looking at the actual facts.

2. The Solution: The "Shared Whiteboard" (Trace-Editing)

The authors propose CO-FACTCHECKER. Instead of chatting back and forth, imagine the AI is drawing its thought process on a shared whiteboard (called a "Thinking Trace").

The AI draws: It writes down its steps: "Step 1: Read the article. Step 2: Conclude it's a total ban."
You (the Expert) step in: You don't just say "No." You walk up to the whiteboard, take a red marker, and physically cross out the wrong step. You write, "Wait, they only mentioned assault weapons, not all guns. Change Step 2."
The Result: The AI doesn't have to guess what you meant. It sees your exact edit on the board. It then continues drawing from that corrected point.

The Metaphor:

Old Way (Dialogue): You are trying to fix a car engine by shouting instructions to a mechanic through a closed window. He can't see the engine, and he keeps guessing what you mean.
New Way (CO-FACTCHECKER): You are standing right next to the mechanic, pointing at the specific bolt that is loose and saying, "Tighten this one." You are editing the work directly.

3. Why This Works Better (The "Scratchpad" Magic)

The paper argues that editing the "thought process" directly is mathematically smarter than talking.

Precision: When you edit the text on the whiteboard, you are removing the exact wrong idea and replacing it with the right one. In a conversation, the AI has to "translate" your words back into its own internal logic, which often causes errors.
Focus: The AI stays focused on the specific claim. In a long conversation, it often gets "lost" and starts talking about things that don't matter. On the whiteboard, the context is always right there in front of it.

4. The Real-World Test: The Gun Ban Claim

The paper tested this with a real example: "Did Vice President Harris announce a 'gun ban'?"

The AI's First Guess: It read some headlines and said, "Yes, True." (Because it saw the words "ban" and "guns").
The Human Expert's Feedback: "No, look closer. They only said they would ban assault weapons, not all guns. Also, that source is biased."
The Result:
- Old Way: The AI got confused, forgot the nuance, and gave a messy answer.
- CO-FACTCHECKER: The human edited the AI's notes directly. The AI immediately corrected its reasoning to say, "Mostly False. They proposed an assault weapons ban, which is a specific type of restriction, not a total ban."

5. The Bottom Line

The researchers found that this "Shared Whiteboard" method is:

More Accurate: The final verdicts are closer to what human experts would decide.
Easier to Understand: You can see exactly how the AI changed its mind because the edits are visible.
Less Frustrating: Humans feel like they are actually guiding the AI, rather than shouting into a void.

In short: CO-FACTCHECKER stops treating the AI like a chatbot you have to nag, and starts treating it like a junior partner you can coach by directly fixing its notes. It turns "talking at" the AI into "working with" the AI.

1. Problem Statement

The paper addresses the limitations of fully automated fact-checking systems, particularly Large Language Models (LLMs) and Large Reasoning Models (LRMs), when applied to complex, real-world claims.

The Gap: Professional fact-checkers rely on deep domain knowledge, cultural context, and ethical judgment to verify claims. In contrast, LLMs/LRMs often lack this grounding, relying solely on parametric knowledge and retrieved evidence. This leads to errors such as overgeneralization, misinterpreting nuanced labels (e.g., confusing "assault weapons ban" with a "total gun ban"), and failing to distinguish between biased advocacy and neutral evidence.
The Collaboration Challenge: While human-in-the-loop approaches are proposed as a solution, existing interaction paradigms (specifically multi-turn dialogue) are ineffective. The authors observe that in dialogue setups, models often:
- Fail to integrate all expert feedback points.
- Lose track of global instructions or the original task context over multiple turns.
- Produce fragmented reasoning traces that require users to stitch together multiple interactions to understand the full logic.
- Struggle to correct specific reasoning errors without regenerating the entire response.

2. Methodology: CO-FACTCHECKER

The authors propose CO-FACTCHECKER, a framework that treats the model's thinking trace (the chain of thought) as a shared scratchpad between the human expert and the AI. Instead of conversing, the human edits the reasoning process directly.

Core Architecture

The framework consists of three main modules and an interaction loop:

Retriever ( $R$ ): Generates search queries based on the claim and retrieves relevant evidence documents.
Verifier ( $V$ ): A base LRM that processes the evidence to draft a candidate solution ( $S_c$ ), comprising a veracity label, an explanation, and a thinking trace ( $t_c$ ) containing ordered reasoning steps.
Editor ( $E$ ): A critical component that translates natural language feedback from the expert into trace-edits.
- Trace-Editing Mechanism: The editor maps feedback to three specific actions:
  - Remove: Delete a flawed reasoning step.
  - Modify: Correct a specific reasoning step.
  - Append Guidance: Add guidance tokens to steer the model's future generation path.
Oracle ( $O$ ): For automatic evaluation, an LLM-based oracle acts as a proxy for the human expert, using ground-truth fact-checks and rubrics to generate synthetic feedback.

Workflow

Proposal: The Verifier proposes a solution with a thinking trace.
Review & Feedback: The expert reviews the trace and provides feedback (e.g., "The claim is 'gun ban' but they only said 'assault weapons ban'").
Trace-Editing: The Editor converts this feedback into structural edits on the trace (e.g., modifying the step that equates the two concepts).
Regeneration: The Verifier continues generation from the edited trace (with guidance tokens), producing a refined solution ( $S^*_c$ ).
Iteration: This loop continues until the expert is satisfied (max 3 iterations in experiments).

3. Key Contributions

New Interaction Paradigm: The introduction of trace-editing as a mechanism for human-AI collaboration, moving beyond multi-turn dialogue to direct manipulation of the reasoning process.
Theoretical Guarantees: The authors provide theoretical proofs demonstrating that trace-editing has higher information capacity and optimizability than dialogue-based feedback.
- Information Capacity: Under the assumption of an information bottleneck (where models compress feedback into internal representations), trace-editing allows for a larger set of reachable states ( $\tau(t)$ ) compared to the constrained mapping of dialogue feedback.
- Optimization: Trace-editing allows for a strictly better local optimum because it avoids the "many-to-one" mapping problem inherent in interpreting natural language feedback.
Comprehensive Evaluation: The framework is evaluated on two datasets (ExClaim and AmbiguousSnopes) using both automatic metrics and human evaluations.
Human-Centric Validation: Extensive user studies with professional fact-checkers and NLP researchers validate the practical utility of the system.

4. Results

Automatic Evaluation

Performance: CO-FACTCHECKER outperformed all baselines, including autonomous systems (FIRE, FactCheck-GPT, SAFE) and other human-AI collaboration methods (Multi-turn Dialogue, Best-of-N, MCTS).
Metrics:
- Veracity Prediction: Improved Precision and F1 scores by 3–4 points over the best autonomous baseline.
- Reasoning Quality: The Entailment Score (measuring alignment between the model's trace and expert-written fact-checking articles) improved significantly (e.g., +6 points on ExClaim).
- Explanation Quality: Generated explanations showed higher ROUGE-L and BERTScore similarity to ground truth.
Ablation Studies: Removing the reward model, restricting trace edits to only appending guidance, or removing guidance tokens all resulted in performance degradation, confirming the necessity of the full trace-editing pipeline.

Human Evaluation

Preference: Human experts preferred CO-FACTCHECKER over multi-turn dialogue in 78% of cases for overall preference.
Key Advantages:
- Instruction Following: Experts noted that CO-FACTCHECKer retained task context better, whereas dialogue models often "got lost" or forgot the original claim.
- Trace Usefulness: The edited traces were rated as more coherent, easier to interpret, and more useful as reusable artifacts compared to the fragmented traces of dialogue.
- Effort: Collaboration required less cognitive effort in the CO-FACTCHECKER setup.
Limitations Observed: Experts noted that while the system improved, it still occasionally made fundamental errors (e.g., conflating direct speech with journalistic interpretation) that required human correction, highlighting that current LRMs still lack perfect alignment with fact-checking practices.

5. Significance and Future Work

Paradigm Shift: The paper argues that for high-stakes, expert-domain tasks like fact-checking, co-constructing a solution via trace-editing is superior to iterative dialogue. It bridges the gap between the "black box" nature of LLMs and the need for transparent, controllable reasoning.
Generalizability: While focused on fact-checking, the authors suggest that trace-editing is a generalizable interaction paradigm applicable to other reasoning-intensive fields like legal analysis and scientific review.
Future Directions: The authors plan to integrate feedback-driven evidence retrieval (where human feedback triggers new searches) and improve the underlying reasoning models to reduce fundamental errors that currently necessitate heavy human intervention.

In summary, CO-FACTCHECKER demonstrates that by treating the model's reasoning trace as a malleable, shared workspace rather than a static output, human experts can effectively guide AI to produce more accurate, context-aware, and logically sound fact-checks.