Can one frozen model supply the reasoning variety I need — and save the cost of running several?

October 29, 2026 · 5 min read

SentX SI

The honest split: the variety half is supported, within a narrow scope, and the cost half is not answered by the evidence at all. If your problem is getting one model to produce genuinely different approaches to the same task, recent work shows it can — with one catch, that the more diverse the bundle of answers, the less reliable each individual path tends to be. If your question is whether one model replaces several for money, the paper offers no basis for that conclusion, and anyone quoting it otherwise is stretching.

Source: Agent-Conditioned Routing: Diverse Reasoning Paths Within One Frozen Model

The work in question is "Agent-Conditioned Routing: Diverse Reasoning Paths Within One Frozen Model," posted to arXiv on October 27, 2026 by Ananth Kalyanasundaram, Ankush Kadu, Athira Gopal and Aswanth Krishnan. It asks whether a frozen sparse mixture-of-experts model can generate diverse solutions from a single prompt — no user-assigned roles, no prescribed methods, no fine-tuning, no extra model copies. Their technique, Agent-Conditioned Routing or ACR, gives each of several parallel attempts a persistent, untrained mask over which experts it may use, while leaving the model's normal input-dependent routing intact. On a screened 126-question mathematics and physics benchmark, with model judges reading each derivation in full, they count the problems where at least two of four trajectories reach a correct solution — "soft-correct," in the authors' term — by distinct primary methods. At matched output budgets, ACR reaches about twice as many such problems as the strongest temperature setting on GLM 5.3 (37.3 versus 18.7) and 1.67 times as many on GPT-5.6 Sol (45.00 versus 27.00). Two caveats belong with those numbers: the study covers exactly those two backbones and that one metric, and the paper is an extended version of a non-archival NeurIPS 2026 workshop entry, so it carries preprint weight.

Source: Agent-Conditioned Routing: Diverse Reasoning Paths Within One Frozen Model

The tradeoff is stated plainly in the abstract. A blinded four-expert audit on a stratified sample of 30 problems found more multi-method solutions under ACR on both backbones — and lower correctness. In ablations, resampling the masks at the token level instead of holding them fixed attained higher correctness than ACR, leaving ACR with only modest multi-method advantages under matched eligibility. And the result that matters most for a practitioner: an informed sequential prompt — feeding earlier solutions back into the next attempt and explicitly asking for an alternative method — produced more diversity than ACR did. The new routing layer is not the ceiling for within-model variety; a few well-placed words in the prompt beat it on the paper's own metric.

Source: Agent-Conditioned Routing: Diverse Reasoning Paths Within One Frozen Model

Now the money question, where the evidence stops. ACR adds about 160 KiB of mask state, and in a three-problem microbenchmark capped at 1,024 tokens, its latency and peak memory stayed within 3 percent of batched temperature sampling on the same model. That is a resource-footprint observation: it tells you ACR is nearly free relative to temperature sampling inside one model. It is not a cost comparison against running several models. The abstract reports no pricing data for either backbone, no ensemble baseline, and no production workload — the comparison set it describes never leaves the single model. Our October 21 piece on model tiering covered the other axis, matching different tasks to different tiers; this one is the same task attempted repeatedly from one model, and it deliberately holds output budgets flat instead of asking what anything costs.

Source: Agent-Conditioned Routing: Diverse Reasoning Paths Within One Frozen Model

To see why the gap matters, here is some illustrative arithmetic — assumed prices, a made-up workload, not a measurement. Suppose a flagship model bills $3 per million input tokens and $15 per million output tokens, and a budget model bills $0.40 and $1.20. A problem statement of 1,500 tokens, attempted four times at 800 output tokens each, costs about $0.066 on the flagship and about $0.006 on the budget model — a factor of roughly ten. Whether that saving is real depends entirely on whether the cheaper model solves your problems correctly, which is precisely the question the paper's correctness finding warns you not to wave away. The arithmetic also understates the sequential-prompt method, because pasting earlier solutions back grows the input side with every attempt. Redo it with your own rate card before treating any number as a saving.

What you can actually run today, as our adaptation of the paper's findings rather than a protocol the authors measured: attempt the problem once with a plain prompt; for each further attempt, include the previous solution and ask explicitly for a different primary method; keep four trajectories, as the paper does; then count how many of your problems yield at least two distinct methods among the correct answers — that is the paper's metric applied to your workload. Read each derivation before trusting it. The audit's lesson is that the more varied the bundle, the more likely an individual path is wrong, so variety is coverage, not verification.

Source: Agent-Conditioned Routing: Diverse Reasoning Paths Within One Frozen Model

That last point lines up with the known-limits note on Victoria's model card: answers can contain incorrect facts, citations or calculations, so open the cited sources and check the supporting passage. A set of diverse reasoning paths is no exception to that discipline — each path gets checked on its own merits before it counts. One operational note for the sequential method: it works by pasting earlier derivations into the prompt, so if they contain work product you would not want shared, check the service's data terms first. SentX's own policy, for instance, states plainly that submissions are not confidential.

Source: Victoria AI Model Card; Privacy Policy | SentX SI

So the decision rule: if you need several approaches to one hard problem from a model you already run, within-model diversity is established to beat temperature tuning, and the sequential prompt is the most effective lever on the paper's own metric. If you need one model to replace several on cost, this paper cannot tell you yes — it measured the footprint of one technique inside one model and stopped there. Your rate card and your correctness requirements are where that answer actually lives.

Source: Agent-Conditioned Routing: Diverse Reasoning Paths Within One Frozen Model