AI News
All stories
Official

NVIDIA researchers show frequent context compaction can match full-history agent runs

NVIDIA researchers report that compacting agent context often matches full-history baselines on long tasks while cutting median peak context from 32K to 8K tokens.

Research · · Updated · 2 min read

NVIDIA researchers report that agents can run long tasks while keeping only a small slice of their history in context, and in one setup the results match full-history runs. In a paper posted to arXiv on November 3, 2026, the authors compare three regimes: accumulating the full history, compacting once the context reaches a threshold, and single-action compaction. In single-action compaction, the agent takes one action per context and carries forward only a short natural-language summary of its state. With one action per context, Claude Opus 4.5 scores 70.2 on SWE-bench Verified against 70.0 with full history, while median peak context falls from 32K to 8K tokens.

The motivation is cost. The paper says later tokens cost more to generate than earlier ones, and that on frontier open-weight models the gap may reach 7x. A context that keeps growing therefore makes every later step more expensive, and compaction targets that growth directly. On Terminal-Bench, Opus 4.5 scores 51.8 against a baseline of 47.2. Median peak context is 8K to 9K tokens against 27K to 32K for the baseline, so the reduction is a range across tasks rather than one figure. The best end-of-conversation compaction setting raises Opus 4.5 from 70.0 to 74.8 on SWE-bench Verified.

The long-input results rest on a different mechanism. Instead of reading an entire document, the agent reads the input selectively and falls back on compaction when it needs more. Across Terminal-Bench, FRAMES, BFCLv3 and multilingual SWE-bench, this approach matches or beats full-input baselines in almost every long-input setting the authors tested, with modest losses in some.

The authors use these results to outline semantic state space models, or SSSMs, which would carry a compact semantic state between steps instead of the full transcript. The paper does not claim that every task can be solved with short context, and its abstract questions the need for long-context models rather than declaring them unnecessary. The evidence covers software engineering and the other benchmarks named above, so it does not show how the approach behaves on tasks whose answers depend on details that a summary drops. Teams can test aggressive compaction on their own workloads before relying on it.

What changed

NVIDIA researchers report that single-action context compaction lets Claude Opus 4.5 score 70.2 on SWE-bench Verified against 70.0 with full history, while median peak context falls from 32K to 8K tokens.

Type: Research · Applies to: Claude Opus 4.5

Why it matters

The paper says later tokens can cost up to 7x more to generate on frontier open-weight models, so aggressive compaction could cut costs for teams running long agent sessions.

What you can do

Test aggressive context compaction on your own agent workloads before paying for million-token agent runs.

How the story developed

  1. Adds the paper's Qwen 3 Coder Next caveat, where the tightest periodic setting cut its SWE-bench Verified score from 53.0% to 3.6%, plus advice to check handoffs and actual cost.

Original source

Where we saw it

2 sources

AI Engineering, DAIR.AI (Elvis Saravia)

AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works