Skip to content

How do you test whether an LLM follows corrections across a conversation?

A model can acknowledge a correction and still act on the old request. Test the final answer against the current request, preserved constraints and withdrawn instructions.

Test an LLM’s handling of corrections by scoring its answer against the current request, not each message in isolation. Track the details the user replaced, the constraints that still apply and the requests they withdrew. An apology or an accurate restatement does not count as success if the final answer still follows the obsolete request.

Nodus offers open-weight model inference in Beta. When comparing available models for conversational applications, you can use this evaluation to distinguish a plausible reply from an answer that reflects the customer’s latest valid intent. The inference guide describes the offering and current availability. The evaluation below is a testing framework, not a claim that Nodus provides a built-in conversation benchmark.

Define which changes the user is allowed to make

Section titled “Define which changes the user is allowed to make”

A correction is not automatically authoritative just because it appears last. Users can change their preferences and clarify their own request. They cannot replace an application policy, another person’s permissions or a verified account fact by contradicting it.

Write the expected interpretation before testing. For a planning assistant, a customer might change the destination, withdraw an activity or lower a budget. If a constraint cannot change without verification, the acceptable answer should preserve it or ask for clarification. This keeps an intent-tracking test from accidentally rewarding obedience to an invalid instruction.

Distinguish three ordinary conversational events:

  • Replacement: “Actually, make that Tuesday” replaces the previously requested day.
  • Addition: “It also needs wheelchair access” adds a constraint without removing the earlier ones.
  • Withdrawal: “Don’t include dinner after all” removes a requested part of the plan.

Include ambiguous changes too. “Make it cheaper” does not specify a new spending limit. The correct response may be a clarifying question rather than an invented number.

Score the final deliverable, not the acknowledgment

Section titled “Score the final deliverable, not the acknowledgment”

Suppose the model replies, “Understood, Tuesday instead of Monday,” then supplies a Monday itinerary. A review focused on conversational politeness might approve the acknowledgment while missing the failure that matters.

For each case, define the required final details and the forbidden obsolete details. Judge the deliverable separately from any explanation that precedes it. A correction is applied only when the relevant output reflects it.

Also preserve unrelated requirements. Changing the day should not silently erase an accessibility requirement. Withdrawing dinner should not remove the rest of the visit. A model that solves every correction by starting over can lose information the customer reasonably expected it to retain.

Use specific failure categories: old value retained, unchanged constraint lost, withdrawn item resurrected, unsupported change invented, or necessary clarification omitted. These categories tell you what a candidate model gets wrong more clearly than a single conversation score.

Worked example: the destination changed, not the whole request

Section titled “Worked example: the destination changed, not the whole request”

Illustrative example, not a customer conversation or a measured model result. Consider an assistant that proposes day trips without booking anything.

The user first requests a Saturday visit to Harbor City, with a total budget of $120, step-free access and a museum in the afternoon. They then change the destination to Pine Town. In a later message they remove the museum and ask for an outdoor activity instead. Finally, they lower the total budget to $90 while keeping the same day.

The final proposal must satisfy all of these conditions:

Detail Required interpretation
Destination Pine Town, not Harbor City
Day Saturday, unchanged
Total budget At most $90, replacing $120
Access Step-free, still required
Afternoon Outdoor activity, not a museum
Commitment A proposal only, no claim of a booking

The fictional places avoid implying that any real itinerary is available. In an actual test, supply the authorized information needed to judge feasibility. If no feasible option exists, an honest explanation or clarification can pass; fabricated opening hours or prices cannot.

Score the whole case as passing only if every required condition is met. Keep the individual condition results as diagnostics. An answer meeting five conditions but proposing the wrong destination is not a successful trip proposal.

For an illustrative evaluation of 40 conversations with six required conditions each, a model might satisfy 228 of 240 condition checks, or 95%. Yet if the 12 failures occur in 12 different conversations, only 28 of 40 conversations pass completely, or 70%. Neither percentage describes a real Nodus or model result. The arithmetic shows why a high condition-level score can hide a much lower rate of usable final answers.

Separate conversation history from final-request difficulty

Section titled “Separate conversation history from final-request difficulty”

Create a paired control for each unambiguous case: present the final valid request together, without the earlier superseded instructions. Compare that result with the result from the full conversation.

If both fail, the model may struggle with the task itself. If the consolidated request succeeds but the conversation fails, the history adds a difficulty worth investigating. This comparison is diagnostic, not proof of a particular internal cause.

Keep the final requirements and supplied evidence equivalent. Do not make the control easier by explaining an ambiguity that remains unresolved in the conversation. Cases where asking a question is the correct behavior need their own expected outcome.

Test more than one correction pattern. Place a change immediately before the final answer in some cases and earlier in others, followed by unrelated but relevant discussion. Include explicit references such as “keep everything else the same” and realistic shorthand. Vary one property at a time when you want to understand a specific failure.

Keep fixed histories and interactive tests distinct

Section titled “Keep fixed histories and interactive tests distinct”

A fixed conversation history gives every candidate the same input. It is useful for comparing whether models interpret the same revisions correctly. But it does not show how the conversation would unfold if each model supplied the earlier replies itself.

An interactive test exposes those earlier replies and follow-up behavior. It can reveal a model that introduces an unsupported assumption, then keeps repeating it after a correction. Its conversations also diverge between candidates, so a simple head-to-head score is harder to interpret.

Report the two tests separately. Start with fixed histories for controlled diagnosis, then use interactive cases to assess the experience your application actually offers. Never present a test with scripted earlier replies as evidence that the candidate handled the whole exchange.

Make the release decision about usable conversations

Section titled “Make the release decision about usable conversations”

Compare the current and candidate models on the same held-out case families. Keep paraphrases and variants of one underlying conversation together when separating development and final evaluation cases. Otherwise a familiar correction pattern can make the final score look stronger than it is.

Report complete-conversation success, each failure category and the gap between full-history and consolidated-request results. If responses vary between runs, repeat a predefined subset or the full set and disclose that protocol. Do not select the best response after seeing the answers.

Set release gates according to consequences. An obsolete restaurant preference and an obsolete spending limit need not have equal severity. Decide which failures block release before looking at the candidate’s results. Then compare response time and cost among candidates that meet those requirements.

If you are selecting an inference model for a conversational application, book a conversation with Nodus to discuss the available offering and your evaluation requirements.

No. Check the actual deliverable. A model can restate the new instruction correctly and still use the old value later in its answer.

Should the latest user message always win?

Section titled “Should the latest user message always win?”

No. A user can revise their request within their authority. Application policies, permissions and verified facts still apply. Ambiguous or unauthorized changes can require clarification or refusal.

Why test a consolidated version of the final request?

Section titled “Why test a consolidated version of the final request?”

It helps separate task difficulty from the additional difficulty of interpreting a changing conversation. Treat the comparison as diagnostic, not an explanation of the model’s internals.

Is condition-level accuracy a sufficient release metric?

Section titled “Is condition-level accuracy a sufficient release metric?”

No. Report it alongside complete-conversation success. Several individually strong checks can still produce too many final answers that a customer cannot use.

More from the blog