In plain words
When a language model reads a long conversation, it can update temporary memory as it goes. Select-TTT is a method that gives stronger updates to parts of the conversation that seem related to the user’s question, and weaker updates to other parts.
In tests of long conversations, Select-TTT had higher average scores than giving every part the same update strength, for both models tested. The results cover those models and tests, not every model or conversation.
The research
When a model reads a long conversation, updating temporary memory from every chunk equally may give irrelevant details as much influence as information that answers the user’s question. Can it direct stronger updates toward useful chunks without already knowing which chunks contain the answer?
Select-TTT estimates which context chunks relate to a question and scales each chunk’s memory update before the model answers.
LongMemEval-S tests whether a chat model can answer questions using information spread across past conversations; its reported accuracy is averaged across questions. Researchers compared Select-TTT with uniform In-Place TTT on the same TTT-enhanced checkpoint for each model.
Select-TTT averaged 37.4% versus 33.6% for Llama-3.1-8B-Instruct, and 28.8% versus 25.4% for Qwen3-4B.
Because the checkpoints and test were shared within each comparison, the results support the value of query-conditioned updates in this inference setting. A second test of reasoning over long conversations also showed higher overall scores for Select-TTT on both models.
The earliest reported continual-training checkpoint is an exception: Select-TTT scored below uniform In-Place TTT there, then matched or exceeded it at later checkpoints. The relevance score is a heuristic, and the evaluation covers only two backbones and two English tests.
What the evidence shows
- Why equal updates matter
Select-TTT is a query-conditioned update method for In-Place test-time training: it changes how strongly each context chunk updates temporary model parameters. The question is whether this selective allocation helps when a model answers from long conversation histories.
LongMemEval-S tests chat models’ ability to answer questions about information across past conversations, and reports accuracy averaged over questions. In one tested setting, uniform updates barely changed the average, which prompted a closer look at how chunks were written.
- A diagnostic with answer labels
The researchers used answer-bearing chunk annotations from LongMemEval-S sessions to vary write strength for answer-bearing and other chunks separately. With the other-chunk update strength fixed, stronger writes to answer-bearing chunks raised accuracy from 33.6% to 37.8%.
In a separate sweep, stronger writes to irrelevant chunks lowered accuracy from 34.8% to 29.6% while answer-chunk strength stayed fixed. Because the answer labels were supplied from the sessions, these results diagnose the effects of selective writes rather than showing that the method can find answer chunks itself.
- Turning relevance into updates
Select-TTT estimates chunk relevance from how well the model’s internal representations predict the words in the user’s question. It then uses those scores to strengthen or weaken each chunk’s temporary-memory update.
The score is a heuristic estimate, not a direct test for answer-bearing content. This gives the method a way to distribute updates by question relevance without relying on the answer labels used in the diagnostic experiment.
- Results across two tests
For inference-only evaluation, the researchers compared Select-TTT with uniform In-Place TTT on the same TTT-enhanced checkpoint for each backbone. On LongMemEval-S, Select-TTT had higher average accuracy across questions for both Llama-3.1-8B-Instruct and Qwen3-4B.
On LoCoMo, a test of reasoning over very long conversations, it also had a higher overall score for both models. The results support the query-conditioned update rule across these two tests and backbones.
- Training results take time
The researchers also trained Qwen3-4B with Select-TTT active during continual training and evaluation. At 5,000 steps, its average LongMemEval-S accuracy was 37.4%, compared with 35.6% for uniform In-Place TTT.
At the 1,000-step checkpoint, however, Select-TTT scored 32.0%, below uniform In-Place TTT at 33.2% and standard continual pretraining at 32.4%. The later advantage therefore does not describe the earliest reported checkpoint.
Deep
- Original paper authors
- Yue Wu, Yang Zhang, Weinan Gan, Chongming Gao, Xinsen Guo
- Submitted to arXiv
- License
- arXiv non-exclusive