Grafy News
Try Grafy free

MIRAGE Introduces Multi‑hop Reasoning with Dynamic Knowledge Injection

Researchers unveil MIRAGE, a framework that lets large language models update and reason with conflicting and new knowledge across multiple hops.

Illustration for the news story: MIRAGE Introduces Multi‑hop Reasoning with Dynamic Knowledge InjectionSource: Grafy Research Lab

MIRAGE, a new framework for multi‑hop reasoning under dynamic knowledge injection, was presented by Heyang Xu, HanWen Hao, Haifeng Wang, Haoyi Zhou, and Jianxin Li in a paper submitted to COLING 2027. The work tackles a gap in current multi‑hop question answering (MQA) research: while many systems can edit isolated facts, they struggle when a model must simultaneously integrate conflicting updates and entirely novel information. MIRAGE proposes a systematic way to inject such knowledge at the level of individual reasoning hops, improving both accuracy and robustness.

What problem does MIRAGE address?

Traditional MQA pipelines rely on a static knowledge base that the language model queries repeatedly to build a reasoning chain. When the underlying facts change—whether due to corrections, new discoveries, or contradictory sources—most systems either ignore the updates or require costly full‑model retraining. The authors point out that real‑world applications, such as legal research or medical diagnostics, often involve dynamic knowledge where new evidence can overturn earlier conclusions. MIRAGE targets this scenario by allowing the model to receive knowledge injections on the fly, and to decide which injected piece to trust for each hop of the reasoning process.

How does the framework perform dynamic knowledge injection?

MIRAGE introduces three core components:

  1. Hop‑level Knowledge Buffers – a lightweight storage that holds newly injected facts, each tagged with provenance metadata (source, timestamp, reliability score).
  2. Selective Trust Module – a trainable classifier that, given the current query, the partially built reasoning path, and the contents of the buffers, predicts which buffer entry should be consulted for the next hop.
  3. Dynamic Retrieval‑Augmented Generation – the language model conditions its next generation step on the selected buffer entry in addition to the usual retrieved documents.

During inference, the system first processes the question, then iteratively performs: (a) retrieve external documents, (b) consult the knowledge buffers via the trust module, and (c) generate the next reasoning statement. This loop continues until the answer is produced. Crucially, the buffers can contain contradictory statements; the trust module learns to weigh them based on context, enabling the model to resolve conflicts without external supervision.

What experimental evidence supports MIRAGE’s claims?

The authors evaluated MIRAGE on two newly constructed benchmarks that extend existing MQA datasets with dynamic updates:

  • DynamicHotpotQA – a variant of HotpotQA where a subset of supporting facts is altered or replaced after the initial dataset release.
  • ConflictingWikiMulti – a collection of multi‑hop questions that intentionally embed contradictory statements from different Wikipedia revisions.

Across both benchmarks, MIRAGE achieved an average improvement of 7.4 % in exact‑match accuracy over strong baselines that use static retrieval or naïve fact‑editing. The framework also reduced hallucination rates by 12 % when faced with conflicting injections, demonstrating more reliable source selection. Ablation studies showed that removing the selective trust module degraded performance by over 5 %, confirming its central role.

How can developers and researchers adopt MIRAGE’s techniques?

The paper releases code and a lightweight Python library that implements the three components. Developers can integrate the library into existing retrieval‑augmented generation pipelines by:

  1. Defining a knowledge buffer JSON schema that matches the expected metadata fields.
  2. Training the trust classifier on a small set of annotated hop‑level decisions (the authors provide a starter dataset of 5 k examples).
  3. Plugging the dynamic generation loop into any LLM API (e.g., OpenAI, Vertex AI) using the provided wrapper.

Because the buffers are model‑agnostic, MIRAGE can be paired with both encoder‑decoder and decoder‑only architectures. Researchers interested in probing the limits of dynamic reasoning can extend the benchmarks or create new injection scenarios, such as time‑sensitive news streams or user‑provided corrections.

Applying MIRAGE concepts with Grafy tools

While Grafy does not yet host a dedicated MIRAGE module, the framework’s workflow aligns closely with several Grafy apps:

  • Grafy Reader (/reader) lets you upload the MIRAGE paper, search it, and ask follow‑up questions, making the research instantly searchable for your team.
  • Grafy Graph (/graph) can visualise the hop‑level reasoning chains produced by MIRAGE, letting you branch, edit, or replay each step of the answer generation.
  • Grafy Chat (/chat) enables you to experiment with dynamic knowledge injection by feeding custom buffer entries to a side‑by‑side model pair—one acting as the base LLM, the other as the trust module.

By combining these apps, creators can prototype a MIRAGE‑style pipeline without writing code: upload a set of knowledge updates, use Chat to run a multi‑hop query, and capture the reasoning graph for analysis. This low‑code approach helps developers evaluate the impact of dynamic knowledge on their own use cases before committing to a full implementation.

Frequently asked questions

How does MIRAGE differ from simple fact‑editing methods?

MIRAGE operates at the hop level, allowing the model to choose different injected facts for each reasoning step. Simple fact‑editing replaces a single static entry, which cannot handle contradictory or context‑dependent updates.

Can MIRAGE be used with any large language model?

Yes. The framework’s buffers and trust classifier are model‑agnostic, and the authors provide adapters for popular APIs. Performance gains have been demonstrated with both GPT‑4‑style and open‑source LLaMA variants.

What resources are needed to train the selective trust module?

The authors release a pre‑trained checkpoint that works out‑of‑the‑box. Fine‑tuning on a domain‑specific dataset typically requires a few thousand annotated hop decisions and modest GPU resources (e.g., a single RTX 3090 for a few hours).

Source: https://arxiv.org/abs/2611.24058


Source: Grafy Research Lab — MIRAGE: A Framework for Multi-hop Reasoning under Dynamic Knowledge Injection