Found first: a primary source the press has not covered yet.
Researchers from KAIST present TAC-CoT, a framework that keeps tool calls and answers as explicit text while converting intermediate reasoning steps into latent representations. In a paper accepted to EMNLP 2026 Main, they report up to 8.8-point accuracy gains and up to 6.8x fewer model forward passes compared to explicit chain-of-thought baselines on agentic math and QA benchmarks.
What the source says
The work comes from the Data Intelligence System Lab and the Optimization and Statistical Inference Lab at KAIST, with authors Juseon Do, Sangmin Bae, Jeesu Jung, and Hwanjun Song. Standard chain-of-thought forces models to verbalize every reasoning step as tokens; TAC-CoT interleaves explicit token segments only where agentic communication is required, at tool calls, observations, and final answers, while the remaining reasoning happens in latent space. Evaluations cover agentic math benchmarks including GSM8K and SVAMP, multi-hop QA benchmarks including HotpotQA and 2WikiMultihopQA, and ToolBench for single- and multi-tool scenarios. The 6.8x forward-pass reduction was observed on QA benchmarks; math tasks showed up to 2.6x reductions.
Why it matters
Fully explicit CoT generates a large number of tokens, raising latency and cost; fully implicit CoT is faster but drops the interface signals that agentic pipelines depend on for tool dispatch and answer extraction. TAC-CoT keeps those signals textual while making internal reasoning steps latent, so orchestration layers keep working without paying the full token overhead. For teams running agentic workloads at scale, up to 6.8x fewer forward passes on QA tasks is a direct throughput and infrastructure cost lever. The accuracy improvement and efficiency gain move in the same direction, with no reported trade-off between them.