Single-Agent LLMs Outperform Multi-Agent Systems on Multi-Hop Reasoning Under Equal Thinking Token Budgets
Dat Tran, Douwe Kiela
cs.CL, cs.MA
2026-04-03
Matched thinking tokens: single agents match or beat five MAS designs on FRAMES and 4-hop MuSiQue. MAS wins only after heavy context corruption.
Multi-agent papers often post higher scores while quietly spending more test-time compute: longer traces, more rounds, more roles. Match the budget, and many of those tricks lose to a strong single agent. Two pieces were still thin: an information-theoretic reason a single agent should not be worse, and a concrete regime where multi-agent structure should actually help.
Dat Tran and Douwe Kiela at Stanford define thinking tokens as tokens spent on intermediate reasoning, excluding prompts and the final answer, and run the comparison on multi-hop QA.
The theory is short. Let Y be the answer, C the full context a single agent sees, and M = g(C) the messages a multi-agent system passes. Y ↔ C ↔ M is a Markov chain. The data processing inequality gives I(Y; C) ≥ I(Y; M), so the ideal single agent cannot have a worse minimum error rate. If the single agent uses C poorly, the effective context is a degraded C̃α, and the inequality no longer says it must beat a structured M. The prediction: SAS should win on clean context; MAS can catch up when context is deleted, masked, substituted, or stuffed with distractors.
Empirically they run FRAMES and the 4-hop slice of MuSiQue at a global thinking budget B ∈ {100, 500, 1k, 2k, 5k, 10k}. Models: Qwen3-30B-A3B, DeepSeek-R1-Distill-Llama-70B, Gemini-2.5-Flash and Pro. SAS spends B in one call. SAS-L is still one call, with a short "list ambiguities, then answer" scaffold meant to fill Gemini's thinking channel. MAS covers Sequential (the main foil: planner, serial workers, aggregator), subtask-parallel, parallel roles, debate, and ensemble. Planner and aggregator stay close to budget-neutral. A separate judge model scores semantic match to the gold answer under a fixed rubric.
Except at 100 tokens, where neither side writes useful reasoning, SAS is best or tied with the best 95% interval across models and both datasets. Table-wide averages at 2,000 tokens: SAS 0.421, SAS-L 0.420, Sequential 0.389, Debate 0.403, Ensemble 0.372. At 10,000: SAS 0.426, Sequential 0.387, Debate 0.420. Gains flatten from 1k/2k to 5k/10k; extra budget starts to spin. Debate is the steadiest MAS variant, parallel-roles next. Ensemble is weak at low and mid budgets and only leads on Gemini-2.5-Pro FRAMES at 5k and 10k. SAS-L is mostly a Gemini effect: it often beats plain SAS on Flash MuSiQue, and is often worse on open-source models at mid-to-high budgets.
Gemini's API accounting warps the comparison. At a requested 10k budget on Flash, SAS reports 1,687 thinking tokens while visible thought is about 251 words, a 359-token proxy: 4.7× inflation. Visible thought also caps: Flash SAS sits near 354 tokens at 1k requested and 359 at 10k. Sequential concatenates several calls, so at 1k on Pro the visible proxy is about 693 versus SAS 390. Part of "MAS thinks more" is bookkeeping.
Context degradation (Qwen3, MuSiQue, 1,000 tokens) lands the theory. Under masking and substitution SAS leads at α = 0.3, the two systems meet in the middle, and Sequential wins at α = 0.7. Deletion is a weaker version of the same curve. Distractor sentences hurt both; SAS stays ahead with a thinner margin. MAS helps when a single trajectory cannot tell relevant from misleading content, not merely when context is long.
A paraphrase ablation adds a warning: light rewrites hurt, deep rewrites help, which is a memorization-shaped pattern. A standard benchmark can book "already seen this item" as an architectural win.
If the product is multi-hop factual QA or retrieve-then-reason, give the thinking budget to one agent before standing up a debate club. The paper does not claim MAS always loses. When context is noisy and one-pass attention cannot sort signal from bait, a structured pipeline can act as a filter. What it does claim is that many published MAS edges are unaccounted compute plus longer visible traces.
It sits next to the Nature result that extra agents stop helping once the single-agent baseline is high. That study has tools and environments and hinges on decomposability. This one strips the environment and equalizes thinking tokens; the SAS default is cleaner. Do not paste these numbers onto coding agents.
Only two multi-hop QA sets. No tools, repos, or browsers. The judge is an LLM with a fixed rubric and can still prefer a phrasing family. Gemini's thinkingBudget is documented as a guide, not a hard cap, so "matched budget" on that API is approximate. SAS-L shows that prompting changes visible thought length, which means MAS prompt craft could be under- or over-credited. α = 0.7 masking/substitution is a stress test, not a product document. DPI assumes M is a function of C; if workers can call tools and fetch evidence that was never in C, the bound needs a rewrite, and this paper does not run that case.