Token generation loop: model repeats same word dozens of times when listing proper nouns
Description
During a long interactive session (~100k+ tokens of context), Claude Opus 4.6 (1M context) was asked to describe a demo scenario involving mobile SIP client applications. When attempting to list popular SIP/VoIP softphone app names for iOS/Android, the model entered a degenerate token generation loop, repeatedly outputting the same app name dozens of times with self-interruptions and false restarts.
Observed Behavior
The model produced output like:
phone uses a SIP softphone app (Odin, Odin, Odin, Odin, Odin, Odin — or Odin, or more
practically: Odin/Odin/Odin — or simply Odin, or more practically: Odin SIP client,
Odin (Odin) 2024, or really any SIP client like Odin, Odin, Odin — let me be practical:
Odin's native SIP client, Odin (lol, sorry — Bria, Odin, or Odin). Realistically:
Odin SIP/VoIP client app on your phone.
When the user pointed out the problem and the model attempted to correct itself, the loop recurred across three additional attempts:
- Attempt 2: listed same name repeated with dashes
- Attempt 3: listed same name repeated again
- Attempt 4: The model recognized it could not produce the list and gave up
Expected Behavior
The model should have produced a simple list of 3-5 distinct SIP softphone app names without repetition or looping.
Context
- Session was deep into a technical conversation (~100k+ tokens accumulated)
- The model had been performing well on complex tasks (C code generation, Docker builds, benchmarking, architecture analysis) throughout the session
- The loop occurred specifically when generating a parenthetical list of proper nouns (product names) in an informal/conversational tone
- The model appeared aware something was wrong (inserting "lol, sorry" mid-loop) but could not break the pattern
- Even after explicit user feedback and a fresh attempt, the loop recurred on the same topic
- Other content generation (code, technical analysis, markdown) worked perfectly in the same session before and after the incident
Reproduction Steps
- Start a Claude Code session with Opus 4.6 (1M context)
- Build up significant context (~100k tokens) through a technical conversation
- Ask the model to list popular mobile SIP softphone applications for iOS/Android as part of a longer response
- Observe: the model may loop on one app name, repeating it with variations
Environment
- Claude Code CLI v2.1.86
- Model: claude-opus-4-6 (1M context)
- Platform: Linux aarch64
- Shell: bash
Additional Notes
- The looping behavior is reminiscent of known LLM degenerate repetition issues, but unusual at this model scale
- The model's self-awareness of the problem ("lol, sorry") without ability to correct it is notable
- The problem persisted across multiple generation attempts within the same conversation turn
- Other content generation (code, technical analysis, markdown) worked perfectly in the same session
- The loop specifically targeted proper nouns (product names) in an enumeration context
This issue has 3 comments on GitHub. Read the full discussion on GitHub ↗