A new transformer architecture skips the token-heavy chain-of-thought approach and still reasons like a much larger model.
Researchers built a depth-recurrent transformer that reasons by looping a single shared-weight block rather than generating extra tokens, so each added reasoning step costs flat memory instead of growing the key-value cache the way chain-of-thought does. They tested it on three tasks with decreasing structural cues: graph reachability, nested boolean logic, and unstructured relational text. On the two sequence tasks, nested boolean logic and unstructured text, the recurrent model landed within two points of fixed-depth transformers using 4 to 6.4 times more parameters. On the graph task it extrapolated well past its training range, something fixed-depth models could barely manage, though the paper does not report a matching parameter-count comparison for that result.
Chain-of-thought reasoning gets expensive to serve at scale because every reasoning token bloats the key-value cache; this approach trades token generation for looping a small block, which could meaningfully cut inference memory for reasoning-heavy workloads. It also backs a broader argument that reasoning ability comes from computation depth, not raw parameter count, a distinction that matters for both training and serving budgets.
These are still narrow, synthetic compositional tasks, not open-ended reasoning, so the real test is whether the trick survives contact with messier problems like actual math or code.