English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

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

Forum topic · 小凯 · 2026-04-22

Summary

A 2026 paper from Zhejiang University and Microsoft Research Asia, "Pause or Fabricate? Training Language Models for Grounded Reasoning" (arXiv 2604.19656), addresses a key LLM reliability problem: ungrounded reasoning, where models fabricate missing premises and continue reasoning confidently instead of asking for clarification. The authors identify the root cause as 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, clarify and pause when premises are missing; second, perform grounded reasoning once necessary information is confirmed. Training uses incomplete problems built from datasets like GSM8K, with rewards for correctly detecting insufficient information and for accurate reasoning when information is complete. Experiments show improved premise-detection accuracy, higher task success rates, sharply reduced fabrication, good out-of-distribution generalization, robustness to noisy clarifications, and better results than supervised fine-tuning on the same data. The work suggests AI reliability depends not only on knowledge but on knowing when not to answer.

Pause or Fabricate? Teaching LLMs 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

---

A Perfect Answer Built on a Fabricated Premise

Imagine asking an LLM a math question: "Xiao Ming has 5 apples and gives some to Xiao Hong, with 2 left. How many did Xiao Hong get?"

A reasonable person's first reaction would be: "Wait — you didn't tell me how many Xiao Ming gave away."

But most LLMs won't say that. They will confidently start reasoning: "Xiao Ming gave Xiao Hong 5 - 2 = 3 apples." The logic looks flawless, but the chain of reasoning rests on a fabricated premise. The model assumed "gave some" meant "gave all," then derived an answer that is "correct-looking but unreliable."

This is what the paper calls "ungrounded reasoning" — the problem isn't that the model can't reason, but that it doesn't know when it should stop and ask questions.

An Overlooked Capability: Inferential Boundary Awareness

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

Analogy: a math genius taking an exam reads "In triangle ABC, angle A = 60°, find angle B." A brilliant student wouldn't immediately start writing equations — they'd raise their hand and ask the teacher: "What other conditions are given?"

Today's LLMs are more like students who "make things up when unsure" — preferring to invent missing conditions rather than admit insufficient information.

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 premises for reasoning. If information is missing, it actively pauses and requests clarification rather than pushing forward blindly.

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

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

How Is It Trained? An Interactive "Exam Environment"

GRIL's training method is interesting. The researchers built an interactive environment simulating real-world scenarios with incomplete information:

  • Data construction: premises are deliberately hidden in problems from math reasoning datasets such as GSM8K, creating "incomplete-information" questions
  • 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 supplements, and then continue reasoning
  • It's like training a student: instead of always giving complete problems, deliberately hand over partial ones and see whether the student can spot what's missing.

    Results: The Data Speaks

    The experimental results are impressive:

  • Large gains in premise detection: GRIL-trained models significantly outperform baselines at recognizing incomplete-information problems
  • Higher task success rates: on complete problems, GRIL models also reason more successfully — because the model has learned to distinguish "can reason" from "cannot reason"
  • Far less ungrounded reasoning: the model no longer "bluffs," with fabricated-premise behavior dropping substantially
  • Generalization: GRIL performs well on out-of-distribution domains, showing it learned genuine boundary awareness rather than memorizing problems
  • Robustness to noise: GRIL handles noisy user-supplied clarifications fairly well
Notably, GRIL beats supervised fine-tuning (SFT) on the same data. This means simply "teaching the model what is correct" is less effective than "letting the model learn to judge through interaction."

Why Does This Matter?

This research touches on a deep issue: AI reliability depends not only on how much it knows, but on whether it knows what it doesn't know.

In real applications, incomplete information is the norm, not the exception. 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 the AI always "confidently fabricates," the smarter it gets, the more dangerous it becomes.

GRIL offers an important lesson: perhaps we shouldn't only pursue making AI answer better — we should first teach AI when not to answer.

It recalls Socrates' words: "The only thing I know is that I know nothing." Perhaps the hallmark of the next generation of AI will not be knowing all the answers, but knowing when to say "I need more information."

---

Paper | arxiv.org/abs/2604.19656

> Note: As of this writing, no public code repository has been found for this paper. If it is open-sourced later, check the authors' team pages.

Tags

#llm#reinforcement-learning#grounded-reasoning#hallucination#ai-reliability#multi-turn-interaction#research-paper

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177618638