Single-Agent LLMs Outperform Multi-Agent Systems on Multi-Hop Reasoning Under Equal Thinking Token Budgets

Dat Tran; Douwe Kiela

cs.CLApr 2, 2026

Single-Agent LLMs Outperform Multi-Agent Systems on Multi-Hop Reasoning Under Equal Thinking Token Budgets

Authors: Dat Tran, Douwe Kiela

Organizations: Stanford University

Abstract

Recent work reports strong performance from multi-agent LLM systems (MAS), but these gains are often confounded by increased test-time computation. When computation is normalized, single-agent systems (SAS) can match or outperform MAS, yet the theoretical basis and evaluation methodology behind this comparison remain unclear. We present an information-theoretic argument, grounded in the Data Processing Inequality, suggesting that under a fixed reasoning-token budget and with perfect context utilization, single-agent systems are more information-efficient. This perspective further predicts that multi-agent systems become competitive when a single agent's effective context utilization is degraded, or when more compute is expended. We test these predictions in a controlled empirical study across three model families (Qwen3, DeepSeek-R1-Distill-Llama, and Gemini 2.5), comparing SAS with multiple MAS architectures under matched budgets. We find that SAS consistently match or outperform MAS on multi-hop reasoning tasks when reasoning tokens are held constant. Beyond aggregate performance, we conduct a detailed diagnostic analysis of system behavior and evaluation methodology. We identify significant artifacts in API-based budget control (particularly in Gemini 2.5) and in standard benchmarks, both of which can inflate apparent gains from MAS. Overall, our results suggest that, for multi-hop reasoning tasks, many reported advantages of multi-agent systems are better explained by unaccounted computation and context effects rather than inherent architectural benefits, and highlight the importance of understanding and explicitly controlling the trade-offs between compute, context, and coordination in agentic systems.

Figures & tables

		100 thinking tokens							500 thinking tokens
Data	Model	SAS	SAS-L	Seq	Sub	Roles	Deb	Ens	SAS	SAS-L	Seq	Sub	Roles	Deb	Ens
FRAMES	Qwen3-30B	0.191	0.195	0.198	0.155	0.207	0.204	0.146	0.240	0.220	0.223	0.187	0.223	0.206	0.193
MuSiQue	Qwen3-30B	0.200	0.210	0.220	0.174	0.202	0.225	0.149	0.250	0.213	0.226	0.187	0.204	0.229	0.183
FRAMES	DeepSeek-R1-70B	0.365	0.374	0.387	0.385	0.427	0.400	0.365	0.427	0.432	0.396	0.402	0.407	0.392	0.400
MuSiQue	DeepSeek-R1-70B	0.294	0.316	0.328	0.309	0.316	0.308	0.316	0.383	0.375	0.332	0.303	0.329	0.320	0.319
FRAMES	G2.5-F	0.333	0.427	0.462	0.365	0.441	0.445	0.284	0.487	0.459	0.494	0.442	0.451	0.470	0.384

Table 1: Single-agent (SAS) vs. multi-agent architectures under matched thinking-token budgets on FRAMES and MuSiQue. SAS-L is the “longer thinking” single-agent variant. “Seq” denotes our primary Sequential multi-agent baseline. “Sub”, “Roles”, “Deb”, and “Ens” denote subtask-parallel, parallel roles, debate, and ensemble architectures, respectively. Bold indicates the highest-accuracy systems and every other system whose 95% bootstrap confidence interval overlaps with that highest system’s confidence interval (see Appendix F for bootstrap confidence interval values).

Figure 2: MuSiQue 4-hop accuracy across Gemini model versions with unlimited thinking tokens.

Panel of Figure 3 — Figure 3: Context degradation results on MuSiQue 4-hop with Qwen3-30B-A3B under a fixed 1000-token thinking budget. The x-axis is the degradation level ( $\alpha$ for deletion, masking, and substitution; $k$ for distractors), and the y-axis is answer accuracy.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

	Gemini-2.5-Flash						Qwen3-30B-A3B
Bucket	$N$	Avg Tokens		Entity Spans		% gold in thoughts	$N$	Avg Tokens		Entity Spans		% gold in thoughts
		SAS	Seq	SAS	Seq	SAS / Seq / Both		SAS	Seq	SAS	Seq	SAS / Seq / Both
MR/SW	72	260	516	4.8	9.5	12.5 / 41.7 / —	60	788	769	9.3	9.9	18.3 / 56.7 / —
SR/MW	124	286	565	5.9	10.3	42.7 / 18.6 / —	96	733	791	8.1	9.2	63.5 / 28.1 / —
BR	265	217	483	5.0	10.0	61.9 / 64.5 / 50.9	209	666	761	7.8	9.4	73.7 / 78.5 / 69.9
BW	714	274	542	5.2	10.1	3.36 / 2.66 / —	810	787	809	7.8	8.9	3.33 / 3.58 / —

Appendix

Table 2: Error analysis on MuSiQue 4-hop at 1k thinking tokens, comparing Gemini-2.5-Flash and Qwen3-30B-A3B. The MAS being used in here is Sequential MAS. Buckets are: MAS right & SAS wrong (MR/SW), SAS right & MAS wrong (SR/MW), both right (BR), and both wrong (BW). “Avg Tokens” gives the average thinking-token count; “Entity Spans” is the average number of entities traversed in the internal reasoning; “% gold in thoughts” is the percentage of examples in which the gold answer appears anywhere in the internal reasoning.

System	100	500	1k	2k	5k	10k
SAS	0.191	0.240	0.252	0.250	0.260	0.263
95% CI	(0.158, 0.201)	(0.238, 0.268)	(0.251, 0.282)	(0.248, 0.275)	(0.247, 0.295)	(0.252, 0.299)
Avg. thinking tokens	100	474	800	1103	1260	1307
Seq	0.198	0.223	0.225	0.252	0.252	0.258
95% CI	(0.174, 0.206)	(0.214, 0.224)	(0.205, 0.232)	(0.227, 0.255)	(0.229, 0.253)	(0.247, 0.281)
Avg. thinking tokens	99.5	494	889	1321	1693	1808

Appendix

Table 3: FRAMES full results for Qwen3-30B-A3B.

System	100	500	1k	2k	5k	10k
SAS	0.200	0.250	0.260	0.271	0.271	0.271
95% CI	(0.196, 0.214)	(0.248, 0.255)	(0.260, 0.271)	(0.269, 0.276)	(0.268, 0.276)	(0.268, 0.276)
Avg. thinking tokens	100	495	916	1345	1453	1462
Seq	0.220	0.226	0.229	0.229	0.231	0.231
95% CI	(0.212, 0.238)	(0.229, 0.235)	(0.223, 0.248)	(0.223, 0.240)	(0.226, 0.243)	(0.225, 0.244)
Avg. thinking tokens	99.6	495	907	1383	1766	1811

Appendix

Table 4: MuSiQue 4-hop full results for Qwen3-30B-A3B.

System	100	500	1k	2k	5k	10k
SAS	0.365	0.427	0.448	0.454	0.455	0.456
95% CI	(0.359, 0.378)	(0.407, 0.432)	(0.421, 0.459)	(0.431, 0.463)	(0.434, 0.464)	(0.434, 0.465)
Avg. thinking tokens	100	466	550	574	635	960
Seq	0.387	0.396	0.391	0.393	0.394	0.397
95% CI	(0.365, 0.395)	(0.373, 0.397)	(0.360, 0.400)	(0.369, 0.405)	(0.372, 0.399)	(0.374, 0.400)
Avg. thinking tokens	99.6	493	749	959	1080	1337

Appendix

Table 5: FRAMES full results for DeepSeek-R1-Distill-Llama-70B.

System	100	500	1k	2k	5k	10k
SAS	0.294	0.383	0.407	0.418	0.419	0.417
95% CI	(0.282, 0.305)	(0.372, 0.397)	(0.400, 0.413)	(0.412, 0.429)	(0.413, 0.429)	(0.402, 0.429)
Avg. thinking tokens	100	412	519	574	635	724
Seq	0.328	0.332	0.320	0.327	0.323	0.327
95% CI	(0.324, 0.347)	(0.317, 0.361)	(0.313, 0.336)	(0.312, 0.357)	(0.318, 0.344)	(0.320, 0.351)
Avg. thinking tokens	99.6	479	788	1017	1105	1124

Appendix

Table 6: MuSiQue 4-hop full results for DeepSeek-R1-Distill-Llama-70B.

System	100	500	1k	2k	5k	10k
SAS acc.	0.263	0.340	0.331	0.334	0.344	0.338
95% CI	(0.249, 0.266)	(0.319, 0.358)	(0.324, 0.346)	(0.322, 0.353)	(0.323, 0.370)	(0.321, 0.361)
Self-count tok.	90	399	637	884	1407	1687
Visible words	124	212	248	254	250	251
Assumed visible tok.	177	303	354	363	357	359
Sequential acc.	0.272	0.287	0.287	0.283	0.289	0.285

Appendix

Table 7: MuSiQue 4-hop full results for Gemini-2.5-Flash.

System	100	500	1k	2k	5k	10k
SAS acc.	0.308	0.391	0.413	0.412	0.419	0.428
95% CI	(0.288, 0.320)	(0.369, 0.405)	(0.397, 0.420)	(0.408, 0.418)	(0.410, 0.425)	(0.413, 0.435)
Self-count tok.	84	442	816	884	1526	1522
Visible words	124	226	273	254	276	275
Assumed visible tok.	177	323	390	363	394	393
Sequential acc.	0.393	0.391	0.400	0.414	0.392	0.392

Appendix

Table 8: MuSiQue 4-hop full results for Gemini-2.5-Pro.

System	100	500	1k	2k	5k	10k
SAS acc.	0.333	0.487	0.551	0.532	0.545	0.547
95% CI	(0.315, 0.341)	(0.473, 0.496)	(0.537, 0.570)	(0.510, 0.537)	(0.541, 0.564)	(0.524, 0.551)
Self-count tok.	89	374	535	714	1057	1267
Visible words	138	221	254	278	272	275
Assumed visible tok.	197	316	363	397	389	393
Sequential acc.	0.462	0.494	0.507	0.526	0.524	0.516

Appendix

Table 9: FRAMES full results for Gemini-2.5-Flash.

System	100	500	1k	2k	5k	10k
SAS acc.	0.368	0.600	0.680	0.700	0.700	0.692
95% CI	(0.358, 0.376)	(0.578, 0.608)	(0.640, 0.704)	(0.673, 0.705)	(0.691, 0.705)	(0.661, 0.718)
Self-count tok.	84	494	807	1119	1158	1210
Visible words	138	261	322	337	337	330
Assumed visible tok.	197	373	460	481	481	471
Sequential acc.	0.654	0.660	0.670	0.690	0.680	0.691

Appendix

Table 10: FRAMES full results for Gemini-2.5-Pro.

		Original		Light paraphrase		Deep paraphrase
Model	Budget	SAS	Seq	SAS	Seq	SAS	Seq
Qwen3-30B-A3B	1k	0.260 (0.260, 0.271)	0.229 (0.223, 0.248)	0.249 (0.236, 0.260)	0.224 (0.175, 0.234)	0.268 (0.264, 0.286)	0.225 (0.214, 0.234)
Qwen3-30B-A3B	2k	0.271 (0.269, 0.276)	0.229 (0.223, 0.240)	0.264 (0.254, 0.274)	0.204 (0.200, 0.220)	0.272 (0.263, 0.286)	0.236 (0.231, 0.248)
Gemini-2.5-Flash	1k	0.331 (0.324, 0.346)	0.287 (0.270, 0.308)	0.326 (0.320, 0.331)	0.278 (0.266, 0.298)	0.358 (0.346, 0.373)	0.308 (0.283, 0.325)
Gemini-2.5-Flash	2k	0.334 (0.322, 0.353)	0.283 (0.271, 0.296)	0.329 (0.323, 0.340)	0.289 (0.271, 0.304)	0.348 (0.332, 0.356)	0.304 (0.277, 0.309)

Appendix

Table 11: Paraphrase results with 95% bootstrap confidence intervals.

Model	SAS	Sequential
Gemini-2-Flash-Lite	0.258 (0.245, 0.282)	0.255 (0.226, 0.266)
Gemini-2-Flash	0.347 (0.333, 0.366)	0.308 (0.296, 0.321)
Gemini-2.5-Flash	0.296 (0.276, 0.319)	0.294 (0.279, 0.316)
Gemini-2.5-Pro	0.399 (0.381, 0.400)	0.396 (0.372, 0.403)
Gemini-3-Pro-Preview	0.491 (0.472, 0.500)	0.489 (0.461, 0.513)

Appendix

Table 12: MuSiQue 4-hop accuracy across Gemini model versions with unlimited thinking, with 95% bootstrap confidence intervals.

Method	Level	SAS	Sequential
Deletion	$\alpha=0.3$	0.245 (0.234, 0.248)	0.221 (0.218, 0.229)
Deletion	$\alpha=0.5$	0.216 (0.210, 0.226)	0.219 (0.217, 0.235)
Deletion	$\alpha=0.7$	0.229 (0.221, 0.242)	0.215 (0.213, 0.225)
Masking	$\alpha=0.3$	0.234 (0.223, 0.255)	0.221 (0.211, 0.229)
Masking	$\alpha=0.5$	0.223 (0.205, 0.238)	0.223 (0.206, 0.231)
Masking	$\alpha=0.7$	0.220 (0.215, 0.224)	0.231 (0.224, 0.248)

Appendix

Table 13: Context-degradation results on MuSiQue 4-hop with Qwen3-30B-A3B and a fixed 1000-token thinking budget.

Explore similar work

Benefits and Limitations of Communication in Multi-Agent Reasoning
Oct 14, 2025Michael Rizvi-Martel, Satwik Bhattamishra, Neil Rathi +2Multi-Agent Reasoning Communication Complexity
Mila & Universit´e de Montr´eal · University of Oxford · Stanford University +1
When Do Multi-Agent Systems Help? An Information Bottleneck Perspective
Jul 17, 2026Wendi Yu, Lianhao Zhou, Xiangjue Dong +6Multi-Agent LLM Systems Information Bottleneck
Department of Computer Science & Engineering, Texas A&M University · Department of Electrical & Computer Engineering, Texas A&M University · Computing and Data Sciences Directorate, Brookhaven National Laboratory
Token-Efficient Multi-Agent Collaboration via System One-Guided Computational Division of Labor
Oct 6, 2026Zihan Zhou, Xinzhe Hu, Hanxu Yang +2Multi-Agent LLM Systems Multi-Agent System Optimization
University of Electronic Science and Technology of China · Southwestern University of Finance and Economics

Oct 14, 2025J/K move · Enter open · S save

Benefits and Limitations of Communication in Multi-Agent Reasoning

Michael Rizvi-Martel, Satwik Bhattamishra, Neil Rathi, Guillaume Rabusseau+1

Mila & Universit´e de Montr´eal · University of Oxford · Stanford University+1

Multi-Agent Reasoning Communication Complexity Multi-Agent Communication Multi-Agent LLMs

Open paper AskPDF arXiv

Chain-of-thought prompting has popularized step-by-step reasoning in large language models, yet model performance still degrades as problem complexity and context length grow. By decomposing difficult tasks with long contexts into shorter, manageable ones, recent multi-agent paradigms offer a promising near-term solution to this problem. However, the fundamental capacities of such systems are poorly understood. In this work, we propose a theoretical framework to analyze the expressivity of multi-agent systems. We apply our framework to three algorithmic families: state tracking, recall, and

k

-hop reasoning. We derive bounds on (i) the number of agents required to solve the task exactly, (ii) the quantity and structure of inter-agent communication, and (iii) the achievable speedups as problem size and context scale. Our results identify regimes where communication is provably beneficial, delineate tradeoffs between agent count and bandwidth, and expose intrinsic limitations when either resource is constrained. We complement our theoretical analysis with a set of experiments on pretrained LLMs using controlled synthetic benchmarks. Empirical outcomes confirm the tradeoffs between key quantities predicted by our theory. Collectively, our analysis offers principled guidance for designing scalable multi-agent reasoning systems.

Abstract

Explore similar work

Benefits and Limitations of Communication in Multi-Agent Reasoning

When Do Multi-Agent Systems Help? An Information Bottleneck Perspective

Token-Efficient Multi-Agent Collaboration via System One-Guided Computational Division of Labor