temperature & think: unintended consequences

[Mr-Ruben]

Reasoning models behavior (high-level)

Considering models with reasoning_model (or similar):

• claude
• gemini
• chatgpt

Key difference:

• claude → reasoning is strictly opt-in
• gemini / chatgpt → reasoning cannot be turned off

---

Temperature handling in Lisette

In core.py:

• Chat.__init__ hardcodes:

  temp = 0

• Later, in _call / _call_async:

  temperature = None if think else ifnone(temp, self.temp)

So the actual behavior is:

• If you do:

  chat = lisette.Chat(...)

  → self.temp = 0

• Then:

  • if think is active → temperature = None
  • otherwise → temperature = temp or self.temp → 0

So default = temperature 0 unless think is enabled

---

Why this matters (Gemini warning)

Logs showed:

21:57:20 - LiteLLM:INFO: vertex_and_google_ai_studio_gemini.py:886 - Warning: Setting temperature < 1.0 for Gemini 3 models (gemini-3.1-flash-lite-preview) can cause infinite loops, degraded reasoning performance, and failure on complex tasks. Strongly recommended to use temperature = 1.0 (default).

2026-04-24 21:57:20,803 - LiteLLM - INFO - Warning: Setting temperature < 1.0 for Gemini 3 models (gemini-3.1-flash-lite-preview) can cause infinite loops, degraded reasoning performance, and failure on complex tasks. Strongly recommended to use temperature = 1.0 (default).

Warning: Setting temperature < 1.0 for Gemini 3 models (gemini-3.1-flash-lite-preview) can cause infinite loops, degraded reasoning performance, and failure on complex tasks. Strongly recommended to use temperature = 1.0 (default).

Triggered by:

model = "gemini/gemini-3.1-flash-lite-preview"

chat = lisette.Chat(model=model)
response = chat("Just respond with OK.")

So:

• No think
• Default temp = 0
• Gemini complains

Also: realistically, most people won’t read these logs.

---

Effort / thinking levels mismatch

Providers:

• Anthropic → low, medium, high, max
• Gemini / OpenAI → minimal, low, medium, high

Problem:

Lisette seems to expect only:

low / medium / high

---

Case: think="minimal"

response = chat("Just respond with OK.", think="minimal")

Raw log:

23:18:24 - LiteLLM:DEBUG: litellm_logging.py:1068 - 

POST Request Sent from LiteLLM:
curl -X POST \
https://generativelanguage.googleapis.com/v1alpha/models/gemini-3.1-flash-lite-preview:generateContent?key=*****M_uw \
-H 'Content-Type: application/json' \
-d '{'contents': [{'role': 'user', 'parts': [{'text': 'Just respond with OK.'}]}], 'generationConfig': {'temperature': 1.0, 'thinkingConfig': {'thinkingLevel': 'low'}}}'

Issue

• minimal is not recognized
• It is treated as None
• Gemini default kicks in → thinkingLevel = low

So effectively:

minimal → ends up as default (low)

---

Control case: think="low"

response = chat("Just respond with OK (low).", think="low")

Raw log:

POST Request Sent from LiteLLM:
curl -X POST \
https://generativelanguage.googleapis.com/v1alpha/models/gemini-3.1-flash-lite-preview:generateContent?key=*****M_uw \
-H 'Content-Type: application/json' \
-d '{'contents': [{'role': 'user', 'parts': [{'text': 'Just respond with OK (low).'}]}], 'generationConfig': {'temperature': 1.0, 'thinkingConfig': {'thinkingLevel': 'low'}}}'

Issue

No issue — behaves as expected.

---

No think specified

response = chat("Just respond with OK. (no think)")

Raw log:

23:33:37 - LiteLLM:DEBUG: litellm_logging.py:1068 - 

POST Request Sent from LiteLLM:
curl -X POST \
https://generativelanguage.googleapis.com/v1alpha/models/gemini-3.1-flash-lite-preview:generateContent?key=*****M_uw \
-H 'Content-Type: application/json' \
-d '{'contents': [{'role': 'user', 'parts': [{'text': 'Just respond with OK. (no think)'}]}], 'generationConfig': {'temperature': 0, 'thinkingConfig': {'thinkingLevel': 'low'}}}'

Important clarification

• thinkingLevel: low here is Gemini’s default

• It happens because:

  • Lisette does not set a valid reasoning level
  • So reasoning_effort=None
  • Gemini defaults to "low"

Issue

• Not setting think → temperature = 0
• This is the real problem (not the thinking level)

---

Summary of issues

1. Default temperature = 0

   • Problematic for Gemini 3 models
   • Hidden unless logs are checked

2. think affects temperature indirectly

   • think=True → temperature=None
   • think=False → temperature=0

3. minimal is not supported

   • Treated as None
   • Falls back to Gemini default (low)

4. Gemini default reasoning can be misleading

   • thinkingLevel: low may appear even if you didn’t request it
   • This is not Lisette explicitly setting it

---

[Mr-Ruben]

I guess the final 'issue' is:

Each provider/model has its own intricacies.

LiteLLM does try to adjust to that by gathering & using the per-model defaults.

Hardcoding a conversion table (think == reasoning_effort) and adjusting other parameters (temperature) based on that, may work today, for a few models, but anything can change [1], anytime. I think forwarding whatever the user passes, without 'black box' magic, is more future-proof, avoids unintended-consequences and becomes more maintainable.

[1] some models have low/medium/high, others minimal/low/medium/high, others none/low/medium/high/xhigh, ...