Codex sometimes re-answers the previous user question before answering the current follow-up

Open 💬 1 comment Opened Jun 16, 2026 by chenkang404

Codex sometimes re-answers the previous user question before answering the current follow-up

Summary

In multi-turn conversations, Codex sometimes appears to over-anchor on the immediately previous user question. When the user asks a follow-up question, Codex may re-answer the earlier question before answering the new one.

This is disruptive because the user expects the new turn to be the primary instruction, with previous turns used only as context.

Observed Behavior

The user reported the following recurring pattern:

User asks a first question:

``text query 理解包括哪些操作，涉及什么算法``

User then asks a second, related but distinct question:

``text codex 你是这样做的吗，你都做了什么，怎么做的，是基于llm吗``

Codex may answer by first explaining the first topic again, for example by restating what query understanding includes, and only then answer whether Codex itself does something similar.

The user explicitly pointed this out:

你有没有发现你现在回答我的问题会把之前的问题再回答一次，你发现自己这个问题了吗

And clarified the expected turn separation:

‘query理解包括什么“ 这个问题 和”codex 是不是也这样做“ 这个问题我是分两次问的，你在回答第二个问题的时候会再次回答一下上一个问题

The user also noted that this has happened frequently recently:

这个问题最近经常出现，是codex 更新的原因吗

Expected Behavior

When a follow-up question is asked, Codex should:

Treat the newest user message as the primary task.
Use the previous question only as context.
Avoid re-answering the previous question unless the user explicitly asks for a recap, comparison, or synthesis.
Start the answer with the current question's requested focus.

For the example above, a better answer to the second question would start with:

是的，Codex 也会做类似的 query 理解，但它不是一个单独的传统 NLP 模块，而是主要由 LLM 在当前上下文中完成意图识别、任务边界判断、工具选择和执行计划生成。

It should not begin by re-explaining the full list of generic query-understanding operations unless that is directly needed.

Impact

The issue affects multi-turn technical discussions where the user intentionally asks questions step by step. It causes:

unnecessary repetition,
weaker turn-level instruction following,
more user interruptions,
loss of conversational efficiency,
confusion about whether Codex understood the newest question.

This is especially visible in Chinese conversations with elliptical follow-ups such as:

"你也是这样吗?"
"那 Codex 是怎么做的?"
"这个和刚才那个有什么区别?"
"你刚才说的在你身上适用吗?"

My Analysis

This looks less like a knowledge problem and more like a turn-prioritization / discourse-tracking issue.

Possible causes:

Over-aggressive context synthesis

Codex may merge the previous and current user turns into one combined task, producing a broad answer that includes the prior question again.

Weak recency prioritization

The newest user message may not be weighted strongly enough when it references a prior concept with pronouns such as "这样" or "这个".

Follow-up resolution expands too much

Resolving "这样" may cause Codex to restate the full antecedent rather than answer the narrower follow-up.

Helpfulness bias toward completeness

Codex may assume that repeating the earlier explanation improves completeness, even though the user wants forward progress.

Conversation batching / interruption edge case

If multiple user messages arrive before the assistant responds, the assistant must answer the newest request and avoid responding to stale intermediate questions as if they were all still active tasks.

Suggested Improvement

Add or strengthen a behavior rule similar to:

For follow-up turns, answer the newest user request first. Use previous turns only to resolve references. Do not re-answer prior questions unless the user asks for a recap or explicitly combines the questions.

For implementation/evaluation, a regression test could use conversations like:

User: query 理解包括哪些操作，涉及什么算法？
Assistant: ...
User: Codex 也是这样做的吗？你做了什么，怎么做的，是基于 LLM 吗？
Expected: Explain Codex's own process directly. Do not repeat the full generic query-understanding answer.

Another test:

User: RAG 的价值是什么？
Assistant: ...
User: 那在我的医疗项目里怎么体现？
Expected: Discuss the user's medical project specifically, not re-answer the general value of RAG.

Codex sometimes re-answers the previous user question before answering the current follow-up