CoBa routing matches best-of-16 voting with 58.9% fewer tokens

Test-time scaling has a default shape: spend more inference compute and trust that the answer gets better. Sample more solutions, extend the reasoning trace, or bring in a stronger evaluator. A paper submitted to arXiv on 7 August 2026 argues those three levers compete, so the question is not how much compute a system spends but how it decides where the next unit goes. The proposed answer is CoBa, a compute-balanced routing policy, with a blunt headline result: matched accuracy at roughly half the token cost.

The paper, titled "CoBa: Cost-Effective Test-Time Scaling via Compute-Balanced Routing," reframes test-time reasoning as a compute-allocation problem. At each step, the system must choose between generation, verification, or stopping. Existing strategies tend to commit to one axis and pump compute into a single phase. CoBa starts from the opposite observation: under a fixed budget, spending on one axis starves the others.

Why single-axis scaling is hitting a wall

CoBa is not alone in questioning the spend-more logic. ThinkRetrieve, submitted to arXiv on 11 August 2026, reports that sequential test-time scaling often yields diminishing or even negative returns, because longer reasoning traces accumulate uncertainty, compound errors, and drift from the original problem. In the same window, "The Verification Horizon" argues that the old assumption for coding agents has inverted: reliably checking a candidate is now the harder problem, and every verifier is only a proxy for human intent. The pattern shows up outside reasoning too, in KlingTeam's Beacon study of when tool calls make easy questions harder.

Those two papers push toward CoBa's premise: the gains from smarter allocation may exceed the gains from more compute. PoTRE, a July 2026 framework, reaches improved reasoning performance with similar or fewer inference tokens than heavily scaled homogeneous baselines. Penelope, submitted 28 July 2026, localizes recurrent computation to avoid long visible reasoning traces. On the inference side, DSpark treats acceleration as a scheduling and resource allocation problem rather than pure model optimization. The wider industry is circling the same issue, with the gap between bought GPUs and actually keeping them busy becoming a procurement problem in its own right.

How CoBa routes compute

CoBa works in two tiers. It first generates a small set of candidates, applies cheap verification broadly across all of them, then routes the uncertain or high-value candidates to stronger verification. Weak verifiers filter obvious failures at low cost; strong verification is spent only where it can change the outcome.

The system was evaluated on 3,129 example-generator evaluations spanning MATH-500, AIME 2024 and 2025, AMC 2023, and procedural symbolic reasoning. Those benchmarks cover competition math and structured symbolic tasks, the terrain where test-time scaling claims are usually tested.

The efficiency numbers that matter

CoBa-Routed-Strong reaches 85.13% macro accuracy, statistically matching a self-evaluation weighted-voting proxy at 85.20% while using 49.1% fewer parameter-weighted tokens. Against brute force, the gap is wider: it matches best-of-16 majority voting within 0.01 macro-accuracy points while using 58.9% fewer parameter-weighted tokens, though paired tests keep a small edge for best-of-16 at substantially higher cost.

Parameter-weighted tokens count each token's work by the size of the model that produced it, so the comparison covers output length and model scale together. That works out to roughly 41% of best-of-16's cost, or about 2.4x cheaper. The self-evaluation proxy consumes almost twice CoBa's budget.

Paired bootstrap tests confirm significant gains over single-sample decoding. CoBa is not just saving tokens while holding quality; it beats the cheapest baseline and matches the expensive ones.

Where CoBa still falls short

The paper is explicit about the limits. Best-of-16 retains a small but significant edge in paired tests: when every accuracy point matters and the budget exists, brute-force voting still nudges ahead. And the pool oracle, which always picks the best candidate on the table, remains out of reach. The gap to the oracle is described as headroom for sharper routing, which is another way of saying the cheap-verification tier still misclassifies candidates that a stronger pass would have caught.

The results also rest on a specific benchmark family. Competition math and symbolic reasoning reward verifiable, well-scoped answers. Whether the same routing economics hold on open-ended agentic or creative tasks is not established by this paper. The broader evidence is cautious: a 957,253-record audit of 30 agent benchmarks finds gains concentrated in coding while enterprise tasks lag.

What it means for local and budget-limited systems

The paper closes with the implication that matters most for on-device and cost-constrained deployment: for local reasoning systems, test-time scaling becomes a question of where the next computation is most valuable. When every token carries a price, a policy that routes instead of sampling is not a convenience. It is the difference between a technique that fits the budget and one that does not.

Local models are already making that bet. Meta's Muse Glimmer runs a 30B agent offline on a single consumer GPU, and Liquid AI's 2.6B-parameter LFM2.5 fits under 2.5 GB of memory. Small footprints are the point: when the hardware is fixed, the only lever left is how the budget is spent.

That framing exposes the assumption CoBa leans on: cheap verification catches most failures. The Verification Horizon paper's warning applies here too. Routing policies reduce the cost of verification, but not the fact that a verifier is a proxy. The efficiency gain is real. The epistemic risk does not disappear.