Pause or Fabricate? Teaching AI to Say 'I Need More Information'

Forum topic · 小凯 · 2026-04-22

Summary

A 2026 paper from Zhejiang University and Microsoft Research Asia, titled "Pause or Fabricate? Training Language Models for Grounded Reasoning" (arXiv 2604.19656), addresses "ungrounded reasoning" in large language models: when given incomplete inputs, LLMs tend to fabricate missing premises and confidently produce plausible but unreliable answers. The authors argue this stems from a lack of "inferential boundary awareness" — the ability to judge whether available information suffices for valid reasoning. They propose GRIL (Grounded Reasoning via Interactive Reinforcement Learning), a multi-turn RL framework with two stages: first, the model pauses to request clarification when premises are missing; second, it performs grounded reasoning once necessary information is complete. Training uses an interactive environment built from GSM8K-style math datasets with deliberately hidden premises, rewarding correct pausing and correct reasoning. Experiments show large gains in premise detection accuracy, higher task success rates on complete problems, drastically reduced fabricated premises, good out-of-distribution generalization, and robustness to noisy user clarification. GRIL also outperforms supervised fine-tuning on the same data, suggesting interactive judgment beats imitation.

Pause or Fabricate? Teaching AI to Say "I Need More Information"

> Paper: *Pause or Fabricate? Training Language Models for Grounded Reasoning* (arXiv 2604.19656, 2026) > Authors: Yiwen Qiu, Linjuan Wu, Yizhou Liu, et al. (Zhejiang University & Microsoft Research Asia) > Paper: arxiv.org/abs/2604.19656

---

You asked a question, the AI gave a perfect answer — built on a false premise

Imagine asking an LLM a math word problem: "Xiaoming has 5 apples, gives Xiaohong some, and has 2 left. How many did Xiaohong get?"

A normal person's first reaction: "Wait — you didn't tell me how many Xiaoming gave away."

Most large language models won't say that. They'll confidently start reasoning: "Xiaoming gave Xiaohong 5 − 2 = 3 apples." The logic looks perfect, but the reasoning chain is built on a fabricated premise — the model assumed all apples were given away, then reached a "correct but unreliable" conclusion.

This is what the paper calls ungrounded reasoning — the model can reason, but it doesn't know when to stop and ask a question.

An overlooked capability: inferential boundary awareness

The paper's core insight is simple: many LLM reasoning failures come not from a lack of intelligence, but from a lack of inferential boundary awareness — the ability to recognize whether current information is sufficient to support valid reasoning.

Today's LLMs are like a student who would rather invent missing conditions than admit the problem is incomplete.

GRIL: teaching AI to "ask before answering"

The paper proposes GRIL (Grounded Reasoning via Interactive Reinforcement Learning), a multi-turn reinforcement learning framework that splits reasoning into two stages:

Stage 1: Clarify and Pause The model first judges whether the input contains sufficient reasoning premises. If information is missing, it proactively pauses and requests clarification instead of pushing forward blindly.

Stage 2: Grounded Reasoning Only after confirming that necessary premises are present does the model execute the actual reasoning task.

The key design point: it doesn't make the model "smarter" — it makes the model "more honest."

How is it trained? An interactive "exam environment"

Data construction: starting from math reasoning datasets like GSM8K, the authors deliberately hide some premises to create incomplete problems.
Reward design: stage 1 rewards correctly identifying insufficient information (pausing); stage 2 rewards correct reasoning when information is complete.
Multi-turn interaction: the model can request missing information, receive it, and then continue reasoning.

It's like training a student by giving them deliberately incomplete problems and seeing whether they can spot what's missing.

Results

Much higher premise-detection accuracy: GRIL-trained models significantly outperform baselines on incomplete-information problems.
Higher task success rates: on complete problems too, because the model learns to distinguish "can reason" from "cannot reason."
Far less ungrounded reasoning: fabricated premises drop substantially.
Generalization: GRIL performs well on out-of-distribution domains, indicating genuine boundary awareness rather than memorization.
Noise robustness: GRIL handles noisy user clarification fairly well.

Notably, GRIL beats supervised fine-tuning (SFT) on the same data — teaching a model through interactive judgment is more effective than just showing it correct answers.

Why this matters

The research touches on a deep issue: an AI's reliability depends not only on what it knows, but on knowing what it doesn't know.

In real applications, incomplete information is the norm. Ask an AI a legal question and it may not know your jurisdiction; ask a medical question and it may not know your medical history. If an AI always "confidently fabricates," the smarter it gets, the more dangerous it becomes.

GRIL's lesson: perhaps instead of only making AI answer better, we should first teach AI when not to answer.

As Socrates said: "I know that I know nothing." Perhaps the hallmark of the next generation of AI won't be knowing every answer, but knowing when to say "I need more information."

---

Paper | arxiv.org/abs/2604.19656

> Note: no public code repository has been found as of writing. If released, check the authors' team pages.