LLM Summarizers Fail Without Prior Identification

LLM Summarizers and the Importance of Identification

Meeting summarizers using language models (LLMs) have a major flaw: the absence of the identification step. One argument made is that this omission is comparable to the failures of regressions that do not check whether the data can actually support the conclusions drawn. During an exchange, a summary produces eight distinct sections, such as decisions, actions to be taken, risks, and open questions. However, upon analyzing the original transcript, it appears that some sections are derived from ambiguous phrases or even invented, without the reader being able to easily verify this information.

This problem is not merely a hallucination, where the model invents facts about the outside world. Here, the model invents facts concerning the meeting itself. The mode of failure is invisible in the final result, as the text appears credible, but it is impossible for the reader to easily verify its validity against the original transcript. There is a name for this mode of failure in another field, older than language models: it occurs when you make an estimate without identification.

The Crucial Step of Identification

Causal inference is an analytical method that distinguishes identification of a quantity from its estimation. Identification involves demonstrating that the available data can support the desired claim. Estimation, on the other hand, is the procedure that generates a figure once identification is established. This order is essential and non-negotiable. One cannot estimate an effect without first proving that it is identifiable from the observed data; otherwise, the resulting figure has no meaning. It may resemble an effect, but it is not one.

Practitioners working with observational data spend a significant portion of their time on identification. They develop causal graphs, debate confounding factors, and distinguish what the data can support from what it cannot. The estimation step, when it finally arrives, is often the simplest.

In the case of LLM summarizers, the process is similar to observational analysis, but it is often deployed without an adequate identification step. The model receives a transcript and produces structured claims about its content: decisions made, commitments accepted, risks raised, next steps assigned. Each claim is, in reality, an estimate of a latent quantity. The decision was made or not, the commitment was accepted or not. The summary asserts a value for each of these quantities, without questioning whether the transcript contains sufficient evidence to support these claims.

Identification and Transcript Data

Identification in observational data raises the question of what the data can support. For a transcript, it is the same question, but applied to a specific source. What can be directly observed, what can be inferred with stated assumptions, and what cannot be supported at all?

Each claim produced by a summarizer should indicate to which category it belongs. Observed claims must point to a specific part of the transcript and assert nothing beyond what that part states. Inferred claims must declare the hypothesis made and the evidence that supports the inference. Recommendations must indicate that they are the model's suggestion, not the decision of the participants.

A summarizer that cannot classify a claim into one of these categories should not produce that claim. The correct output in this case is not a smoother claim, but the absence of a claim.

An Architecture to Enforce Discipline

The proposed architecture relies on a three-step LLM framework and a deterministic output. The first step involves extracting structured facts from the transcript: speech turns, explicit commitments, explicit decisions, explicit quantities. This step is deliberately conservative; it may miss elements, but it is not allowed to invent them.

The second step synthesizes these facts into claim objects across eight sections. Each claim carries a label: observed, inferred, or recommendation. Each claim is associated with a pointer to the evidence in the extracted facts. It is at this stage that the analytical work occurs, and it is also where the model is most likely to diverge.

The third step is the audit, which performs the identification work. The constraint applied at this stage is crucial for the design. The audit cannot rewrite the analysis to make it smoother, nor add a better-formulated recommendation, nor invent missing context. It has a limited set of operations and can do nothing else. It can delete a claim, downgrade a claim from observed to inferred, or from inferred to recommendation, move a claim to a more appropriate section, or replace a claim with a placeholder in case of insufficient evidence. It can reduce an entire section when nothing survives scrutiny.

Results of the Architecture and Implications

This architecture is not a benchmark but a small stress test based on fixtures, designed to verify whether it produces the expected behavior. Three transcripts were used: a decision meeting where a pricing model was selected from three real alternatives, a working session that highlighted a measurement issue without resolving it, and a thin sync between two people with almost no decision-making content.

The results show that the pipeline produced no fabricated commitments or unfounded quantities. This is what the architecture is designed to make more difficult. A claim cannot survive the pipeline without a pointer to evidence, and the audit step cannot fabricate evidence to keep a claim alive.

The most interesting result is the abstention rate. This rate increases with the thinness of the input signal. Across the three fixture transcripts, the share of empty sections increased from 17% to 58%. For the rich decision meeting, the pipeline left 17% of the sections empty or replaced with a placeholder for insufficient evidence. During the working session, this figure reached 25%. In the thin sync, it reached 58%. The system produced about three and a half times more empty sections when the input signal was thin compared to when it was rich.

This behavior is what the design seeks to produce. A summarizer that fills the same eight sections regardless of the input is not the desired outcome.