Absence-Aware Training to Suppress Hallucinations

581 concepts376 papers5 proposals

State-of-the-art vision-language and multimodal models suffer from a fundamental flaw: they are "eager to please" (Menghao Li et al. in Iterative Robust Visual Grounding with Masked Reference based Centerpoint Supervision). Because standard datasets primarily pair inputs with descriptions of entities that are present, models develop a severe affirmation bias. They learn to assume that if a noun or query is mentioned, its corresponding target must exist somewhere in the context.

As a result, models often become "lazy," relying on simple token-matching heuristics rather than deeply verifying the context (Liunian Harold Li et al. in DesCo: Learning Object Recognition with Rich Language Descriptions). When presented with a query about an entity that is missing, they frequently hallucinate its presence.

The Mechanism: Learning the Null Response

To break this bias, a growing body of research shifts the training paradigm from purely positive-pair alignment to absence-aware training. The core idea is to explicitly feed the model negative samples—scenarios where the requested entity, attribute, or relation is entirely absent—and train it to output a designated "null" state.

Depending on the task, this null response takes different forms:

By treating absence as a valid, learnable state rather than an unresolvable error, the model establishes a precise boundary for semantic consensus.

Strategies for Generating Hard Negatives

The primary challenge of absence-aware training is acquiring high-quality negative samples without relying on expensive manual annotation. Researchers have developed several automated pipelines to generate these "hard negatives":

Why It Works: Shifting from Matching to Verification

Why does adding negative samples fundamentally change model behavior?

In standard training, a model minimizes loss by finding any correlation between the text tokens and the visual features. If a query asks for a "white cup," and no white cup exists, the model might attend to a white plate or a clear glass because they share partial feature embeddings.

Absence-aware training forces the model to shift from a matching heuristic to a verification process. For instance, Ayush Jain et al. in Bottom Up Top Down Detection Transformers for Language Grounding in Images and Point Clouds introduced a dedicated "not-mentioned" text label to represent objects present in a scene but absent from the text, forcing the model to explicitly decouple unaligned features. Similarly, Aisha Urooj Khan et al. in Found a Reason for me? Weakly-supervised Grounded Visual Question Answering using Capsules utilized dedicated representational units trained to capture background or non-object information, allowing the model to explicitly attend to empty regions when a query refers to an absent entity.

When a model is penalized for false positives on verified negative samples, it learns that predicting "not present" yields a lower loss than guessing a partial match. This mechanism is critical for real-world deployment, such as robotics, where a target object may be absent from most sensor frames (Thao Nguyen in Language-Conditioned Observation Models for Visual Object Search), or document analysis, where models must rely strictly on visual grounding rather than memorized associations (Francesco Pinto et al. in Extracting Training Data From Document-Based VQA Models).

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 363 papers
  2. Version 1 · 2026-10-04 · 391 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_negative_sample_training_bc839fa2,
  title = {negative sample training},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/negative-sample-training/rdr_bc839fa2af004f77879fb57f70f271f9},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 2, frozen text):

@misc{lacuna_negative_sample_training_bc839fa2_v2,
  title = {negative sample training},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/negative-sample-training/rdr_bc839fa2af004f77879fb57f70f271f9/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake