TAC-CoT cuts agent forward passes 6.8x, lifts accuracy 8.8 points

TAC-CoT, a framework introduced in an EMNLP 2026 paper on arXiv from Juseon Do, Sangmin Bae, Jeesu Jung and Hwanjun Song, reports agent accuracy gains of up to 8.8 points while reducing model forward passes by up to 6.8x on agentic math and QA benchmarks.

The approach "replaces explicit reasoning traces with compact latent reasoning segments while retaining textual tokens only for tool calls, observations, and final answers," the authors write. They frame it as a response to a known tradeoff: explicit chain-of-thought "incurs substantial generation overhead," while existing implicit CoT schemes compress rationales into continuous tokens that break the text interface a tool-calling agent needs.

The abstract also claims gains in "both single- and multi-tool settings."

Both headline figures are stated as "up to," and the abstract publishes no per-benchmark numbers, no base-model names or sizes, and no named baselines against which the 8.8-point lift was measured.