Teaching Language Models to Abstain Using Synthetic Negative Contexts

12 concepts11 papers1 proposal

Autoregressive language models are fundamentally eager to please. Because standard supervised fine-tuning pairs a prompt with a valid target response, models learn that their primary objective is always to generate an answer.

However, when deployed in real-world scenarios where the provided context is incomplete, irrelevant, or entirely missing the necessary facts, this behavior becomes a liability. Instead of admitting it does not know, a model will often force associations, drawing from its internal parametric memory to hallucinate a plausible-sounding but incorrect response. To build trustworthy systems, models must learn a critical alternative skill: knowing when not to answer.

The Solution: Context-Aware Negative Training

The paradigm of context-aware negative training solves this by explicitly teaching models to output a predefined "rejection token" or null response (such as "N/A", "none", or a specific abstention token) when the provided context is insufficient to answer a query.

Because human-annotated data for unanswerable queries is scarce and expensive to collect, researchers rely on synthetic negative augmentation. By deliberately breaking the relationship between the query and the context during training, we can explicitly map out the boundaries of what the model should refuse to answer.

Standard Training:
[Valid Context] + [Query] ----------> [Correct Answer]

Context-Aware Negative Training:
[Irrelevant Context] + [Query] ------> [Rejection Token]

Mechanics of Synthetic Negative Augmentation

The literature highlights three primary strategies for constructing these synthetic negative samples across different tasks:

1. Irrelevant Context Pairing

In grounded dialogue generation, Hongyuan Lu et al. in On Controlling Fallback Responses for Grounded Dialogue Generation intentionally pair dialogue contexts with irrelevant knowledge documents. This forces the model to learn the boundaries of the provided information rather than inventing facts.

Similarly, for zero-shot dialogue state tracking, Zhaojiang Lin et al. in Zero-Shot Dialogue State Tracking via Cross-Task Transfer pair text passages with questions about completely unrelated topics, training the model to emit an explicit "none" output for out-of-domain or unmentioned slots.

2. Information Ablation

Instead of swapping the entire context, we can surgically remove the answer-relevant information from an otherwise valid text. In Recall with Reasoning: Chain-of-Thought Distillation for Mamba’s Long-Context Memory and Extrapolation, Jun-Yu Ma et al. create synthetic training examples by dropping the specific text segments containing the answer. This ablation teaches the model to recognize when a text is missing the exact facts required to formulate a response.

3. Tool and Function Omission

This principle extends naturally to tool use and function calling. When an agent receives a user request but the correct tool is missing from its available API list, it should refuse to act.

Qiqiang Lin et al. in Robust Function-Calling for On-Device Language Model via Function Masking introduce "irrelevance-augmented training" where the correct tool is omitted, forcing a null response. This teaches the model to distinguish between valid tool requests and queries that fall outside the scope of available APIs.

Similarly, Yi-Chang Chen et al. in Enhancing Function-Calling Capabilities in LLMs: Strategies for Prompt Formats, Data Integration, and Multilingual Translation show that augmenting training sets with such negative samples significantly reduces unnecessary or erroneous tool calls.

Why It Works: Shifting the Optimization Boundary

To understand why this works, consider the model's optimization landscape. Without negative training, a model only learns the conditional probability distribution $P(\text{Answer} \mid \text{Context}, \text{Query})$ over valid pairs. When given an irrelevant context, the model's token probabilities are uncalibrated, causing it to fall back on internal biases or high-frequency vocabulary tokens.

By introducing synthetic negative pairs, the training objective explicitly shapes the model's behavior at the boundaries of its knowledge. The model learns a new conditional mapping:

$$P(\text{Rejection Token} \mid \text{Irrelevant Context}, \text{Query}) \to 1$$

This formalizes what Sheng Zhang et al. describe in Context-faithful Prompting for Large Language Models as context-constrained abstention. By providing an explicit path to acknowledge missing information, the model is prevented from defaulting to its internal knowledge base.

As Robert Wolfe et al. note in Laboratory-Scale AI: Open-Weight Models are Competitive with ChatGPT Even in Low-Resource Settings, prioritizing this context-contingent refusal over speculative generation directly enhances model trustworthiness in information-critical domains.

Scaling to Complex and High-Stakes Architectures

The power of explicit abstention becomes evident when scaling to complex system architectures like Retrieval-Augmented Generation (RAG). In Parallel Context Extension (PCE) frameworks—where multiple documents are processed concurrently across separate context windows—noisy or irrelevant documents can easily corrupt the aggregated output.

To solve this, Zexiong Ma et al. developed the DePaC framework, detailed in both Enhancing Retrieval-Augmented Generation with Dehallucinating Parallel Context Extension and Dehallucinating Parallel Context Extension for Retrieval-Augmented Generation. By utilizing context-aware negative training, they fine-tune individual context windows to emit a distinct rejection token when a retrieved document is irrelevant. This document-level abstention filters out high-confidence erroneous generations before they can be aggregated, drastically mitigating hallucinations.

[Query] + [Doc 1 (Relevant)]   ---> [Partial Answer]
[Query] + [Doc 2 (Irrelevant)] ---> [Rejection Token]  ---> Aggregator filters out Doc 2
[Query] + [Doc 3 (Relevant)]   ---> [Partial Answer]

A similar benefit is observed in zero-shot semantic parsing. In ZEROTOP: Zero-Shot Task-Oriented Semantic Parsing using Large Language Models, Dheeraj Mekala et al. show that augmenting question-answering datasets with synthetic negative samples ensures that models reliably abstain from filling slots when the required information is absent from the source text.

Finally, this capability is essential for high-stakes real-world applications. In long-term multimodal interactions, Haochen Xue et al. in MMRC: A Large-Scale Benchmark for Understanding Multimodal Large Language Model in Real-World Conversation identify a persistent reluctance to "say no" as a core failure mode when conversational context is insufficient.

In dense corporate compliance tasks, Siqi Sun et al. in ESG-Bench: Benchmarking Long-Context ESG Reports for Hallucination Mitigation demonstrate that training a model to explicitly evaluate whether a provided context contains sufficient information before generating a response provides a formal path to acknowledge missing facts, drastically reducing dangerous hallucinations in financial and sustainability audits.

Key Foundational Work

Zero-Shot Dialogue State Tracking via Cross-Task Transfer (2021) — contributed 1 concept elements to this direction.
On Controlling Fallback Responses for Grounded Dialogue Generation (2022) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_abstention_training_04361168,
  title = {abstention training},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/abstention-training/rdr_043611687f204c3b9491d40b04000e4e},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_abstention_training_04361168_v1,
  title = {abstention training},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/abstention-training/rdr_043611687f204c3b9491d40b04000e4e/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake