Let's (not) just put things in Context: Test-Time Training for Long-Context LLMs
This addresses a critical issue for users of long-context LLMs by showing that inference-time compute is better spent on context-specific training rather than current scaling strategies, offering a practical solution.
The paper tackles the problem of long-context LLMs failing to reliably use extended text due to limitations like score dilution in static self-attention, and finds that a method using targeted gradient updates on context during inference leads to large performance improvements, such as 12.6 and 14.1 percentage point gains for Qwen3-4B on benchmarks.
Progress on training and architecture strategies has enabled LLMs with millions of tokens in context length. However, empirical evidence suggests that such long-context LLMs can consume far more text than they can reliably use. On the other hand, it has been shown that inference-time compute can be used to scale performance of LLMs, often by generating thinking tokens, on challenging tasks involving multi-step reasoning. Through controlled experiments on sandbox long-context tasks, we find that such inference-time strategies show rapidly diminishing returns and fail at long context. We attribute these failures to score dilution, a phenomenon inherent to static self-attention. Further, we show that current inference-time strategies cannot retrieve relevant long-context signals under certain conditions. We propose a simple method that, through targeted gradient updates on the given context, provably overcomes limitations of static self-attention. We find that this shift in how inference-time compute is spent leads to consistently large performance improvements across models and long-context benchmarks. Our method leads to large 12.6 and 14.1 percentage point improvements for Qwen3-4B on average across subsets of LongBench-v2 and ZeroScrolls benchmarks. The takeaway is practical: for long context, a small amount of context-specific training is a better use of inference compute than current inference-time scaling strategies like producing more thinking tokens.