ReHoPER: Receding-Horizon Planning for Enhanced Reasoning

Saeed Ahmadnia; Cornelia Caragea

cs.CLOct 1, 2026

ReHoPER: Receding-Horizon Planning for Enhanced Reasoning

Authors: Saeed Ahmadnia, Cornelia Caragea

Organizations: Department of Computer Science University of Illinois Chicago

Abstract

We propose ReHoPER, an inference-only, zero-shot method that improves large language models' reasoning by generating and answering intermediate questions along multiple paths before the final answer. It iteratively plans a horizon of candidate intermediate questions, selects one to answer, and replans from the updated history. ReHoPER is task-agnostic, using the same generic instructions across datasets and models without labeled data or task-specific prompt design. Across multiple datasets, including iLLC, a new controlled benchmark for compositional reasoning, ReHoPER outperforms strong baselines, with the largest gains in the most compositional settings. Our implementation and the iLLC generator are publicly available to support future work.

Figures & tables

Panel of Figure 1 — Figure 1 : ReHoPER overview. Layered frames indicate independent paths. Each path plans questions, answers one, and replans from updated history. A majority vote over the initial and per-step predictions across paths yields the final prediction.

Model	SD	SC	RH	SC	RH	SC	RH	SD	SC	RH	SC	RH	SC	RH	SD	SC	RH	SC	RH	SC	RH
	MuSR MM							MuSR OP							MuSR TA
		IO		CoT		PS			IO		CoT		PS			IO		CoT		PS
Gemma 3	57.6	54.4	61.2	61.2	62.0	61.6	65.6	50.0	50.4	55.1	51.6	55.5	54.7	55.9	50.0	58.8	47.6	56.0	51.2	58.0	56.0
gpt-oss	74.8	70.8	76.0	70.4	76.4	73.2	76.8	53.9	55.5	53.9	53.1	53.5	57.0	53.5	70.8	69.2	68.0	56.4	63.2	65.2	68.0
Llama 3.1 8B	54.0	56.8	62.4	56.4	60.8	57.2	63.6	44.5	49.2	53.1	52.0	48.4	52.0	48.4	34.8	35.2	43.6	43.2	47.6	45.2	47.6
Llama 3.1 70B	64.8	65.6	71.2	66.0	70.4	66.8	71.2	42.6	46.1	55.9	46.1	54.3	48.4	55.9	57.6	62.0	57.6	62.0	59.6	60.4	59.2

Table 1 : Accuracy (%) of ReHoPER (RH) on MuSR, MoreHopQA, and iLLC benchmarks under IO, CoT, and PS prefixes. SD refers to the Self-Discover baseline. SC denotes simple question answering with Self-Consistency using the corresponding output prefix, while RH applies our planning approach with that prefix. Green , red , and gray indicate performance changes of ReHoPER relative to SC. The bottom-right block reports averages over the eight displayed subsets for each model and prompt.

	MuSR MM									MoreHopQA
	IO			CoT			PS			IO			CoT			PS
Model	-RP	-LA	RH	-RP	-LA	RH	-RP	-LA	RH	-RP	-LA	RH	-RP	-LA	RH	-RP	-LA	RH
Gemma 3	57.6	58.8	61.2	59.2	58.4	62.0	62.4	62.4	65.6	64.0	66.0	64.7	66.7	66.7	68.7	70.0	69.3	74.0
gpt-oss	72.4	70.0	76.0	72.0	68.0	76.4	74.0	71.2	76.8	80.7	78.7	78.7	79.3	79.3	80.0	80.7	78.7	81.3
Llama 3.1 8B	61.6	58.4	62.4	61.6	61.2	60.8	64.8	60.4	63.6	52.0	57.3	57.3	56.7	58.7	60.7	60.7	66.0	63.3
Llama 3.1 70B	70.4	70.0	71.2	72.0	72.4	70.4	71.6	71.2	71.2	68.7	74.0	71.3	70.7	72.7	71.3	72.7	74.0	74.0

Table 2 : Planning ablations without replanning (-RP), without look-ahead (-LA), and the full ReHoPER method (RH) across three prompting styles, IO, CoT, and PS, on MuSR MM, MoreHopQA, and iLLC L2-4. Values are accuracies (%), and green and light green mark the best and second best variant, respectively, for each (model, dataset, prefix) triplet. The bottom-right block reports averages over all datasets for each model and prompt.

Figure 2 : Answer aggregation for ReHoPER+PS on iLLC L2-4 across models. Top: accuracy over reasoning steps; bottom: accuracy over stochastic paths. CV is cumulative voting, CSV voting over each path’s latest prediction, and CSM their mean correctness. CPV/CPM are voting/mean correctness within a path; APV/APM are their averages across paths.

Table 5

Llama 3.1 70B	Runtime	1:16	3:37	2:42	1:26	4:23	4:12	2:11	6:23	7:56	2:37	9:04	11:15
		iLLC L1-4			iLLC L1-6			iLLC L2-4			iLLC L2-6
Model		SC 61	SC 181	RH	SC 61	SC 181	RH	SC 61	SC 181	RH	SC 61	SC 181	RH
Llama 3.1 70B	Accuracy	92.1	92.4	92.7	76.0	76.9	80.1	32.5	35.1	79.3	17.2	16.7	68.5
Mistral Small	Runtime	0:38	1:38	0:54	0:43	2:03	0:55	0:48	2:08	1:15	0:56	2:41	1:13
Mistral Small	Accuracy	90.7	90.8	91.3	76.1	76.9	73.5	3.5	2.4	11.6	0.0	0.1	2.0
Phi-4	Runtime	0:45	1:46	1:06	0:52	2:31	1:57	0:50	2:01	1:18	1:00	2:55	2:21

Table 4 : Runtime (h:mm) and accuracy (%) comparison between Self-Consistency (SC 61 ) and ReHoPER (RH) across iLLC L1 and L2 subsets. SC 181 uses 181 samples as a higher-compute SC comparison.

Appendix figures & tables25 assets

Supplementary material from the paper’s appendix.

Appendix

Model	SD	SC	RH	SC	RH	SC	RH	SD	SC	RH	SC	RH	SC	RH
	MuSR							iLLC
		IO		CoT		PS			IO		CoT		PS
Gemma 3	52.5	54.5	54.6	56.3	56.2	58.1	59.2	24.7	7.8	16.6	30.5	30.5	37.1	32.5
gpt-oss	66.5	65.2	66.0	60.0	64.4	65.1	66.1	62.2	85.9	83.9	78.3	83.4	52.7	83.9
Llama 3.1 8B	44.4	47.1	53.0	50.5	52.3	51.5	53.2	12.2	28.3	26.9	26.5	31.7	35.0	37.1
Llama 3.1 70B	55.0	57.9	61.6	58.0	61.4	58.5	62.1	39.9	40.3	69.9	44.5	69.9	56.6	77.3

Appendix

Table 5: Macro-aggregated accuracy (%) on three benchmark families. MuSR averages MM, OP, and TA; iLLC averages L1-4, L1-6, L1-8, L2-4, L2-6, and L2-8; and MoreHopQA is reported separately. The final block is an equal-weight average of these three family-level results. SD denotes Self-Discover. Green , red , and gray indicate improvements, decreases, and ties of ReHoPER (RH) relative to Self-Consistency (SC), respectively.

Figure 4 : Step-wise aggregation analysis for ReHoPER under the IO answer format. Each panel shows accuracy as a function of the reasoning step for one model–dataset pair. CV denotes cumulative voting over all predictions up to the current step; CSV votes over one carried-forward prediction per path at the current step; CSM reports the mean carried-forward per-path correctness. Shaded bands show the standard deviation for the per-path CSM diagnostic.

Figure 5 : Step-wise aggregation analysis for ReHoPER under the CoT answer format. Each panel shows accuracy as a function of the reasoning step for one model–dataset pair. CV denotes cumulative voting over all predictions up to the current step; CSV votes over one carried-forward prediction per path at the current step; CSM reports the mean carried-forward per-path correctness. Shaded bands show the standard deviation for the per-path CSM diagnostic.

Figure 6 : Step-wise aggregation analysis for ReHoPER under the PS answer format. Each panel shows accuracy as a function of the reasoning step for one model–dataset pair. CV denotes cumulative voting over all predictions up to the current step; CSV votes over one carried-forward prediction per path at the current step; CSM reports the mean carried-forward per-path correctness. Shaded bands show the standard deviation for the per-path CSM diagnostic.

Figure 7 : Path-wise aggregation analysis for ReHoPER under the IO answer format. CV cumulatively votes over predictions from paths

1

through

j

, including the shared initial prediction only once. CPV and CPM summarize the current path alone, while APV and APM show the average path-local vote and mean correctness across all paths.

Figure 8 : Path-wise aggregation analysis for ReHoPER under the CoT answer format. CV cumulatively votes over predictions from paths

1

through

j

, including the shared initial prediction only once. CPV and CPM summarize the current path alone, while APV and APM show the average path-local vote and mean correctness across all paths.

Figure 9 : Path-wise aggregation analysis for ReHoPER under the PS answer format. CV cumulatively votes over predictions from paths

1

through

j

, including the shared initial prediction only once. CPV and CPM summarize the current path alone, while APV and APM show the average path-local vote and mean correctness across all paths.

Figure 10 : Accuracy gains from allocating iterations to steps vs. paths under the IO answer format.

Figure 11 : Accuracy gains from allocating iterations to steps vs. paths under the CoT answer format.

Figure 12 : Accuracy gains from allocating iterations to steps vs. paths under the PS answer format.

Model	#IQ	#IQ/Path std	#uniqIQ	#uniqIQ/Path std	%uniqIQ	%uniqIQ/Path
	MuSR MM
Gemma 3	60.00	$10.00_{\mathrm{0.00}}$	59.29	$10.00_{\mathrm{0.00}}$	98.82%	100.00%
gpt-oss	39.38	$6.56_{\mathrm{3.72}}$	39.29	$6.56_{\mathrm{3.72}}$	99.77%	100.00%
Llama 3.1 8B	59.95	$9.99_{\mathrm{0.02}}$	59.54	$9.99_{\mathrm{0.02}}$	99.32%	100.00%
Llama 3.1 70B	60.00	$10.00_{\mathrm{0.00}}$	56.72	$10.00_{\mathrm{0.00}}$	94.53%	100.00%
Mistral Small	59.96	$9.99_{\mathrm{0.01}}$	55.86	$9.99_{\mathrm{0.01}}$	93.16%	100.00%

Appendix

Table 6 : Statistics of generated intermediate questions (IQs) in ReHoPER on MuSR and MoreHopQA . We report the total number of generated IQs, the average number of IQs per path with standard deviation, the number of unique IQs, the average number of unique IQs per path with standard deviation, and the corresponding uniqueness percentages.

Model	#IQ	#IQ/Path std	#uniqIQ	#uniqIQ/Path std	%uniqIQ	%uniqIQ/Path
	iLLC L1-4
Gemma 3	34.98	$5.83_{\mathrm{0.90}}$	17.23	$5.83_{\mathrm{0.90}}$	49.25%	100.00%
gpt-oss	53.96	$8.99_{\mathrm{1.31}}$	35.02	$8.99_{\mathrm{1.31}}$	64.89%	100.00%
Llama 3.1 8B	23.58	$3.93_{\mathrm{2.47}}$	17.78	$3.93_{\mathrm{2.47}}$	75.40%	100.00%
Llama 3.1 70B	39.54	$6.59_{\mathrm{0.58}}$	10.14	$6.59_{\mathrm{0.58}}$	25.65%	100.00%
Mistral Small	38.87	$6.48_{\mathrm{1.66}}$	17.58	$6.48_{\mathrm{1.66}}$	45.23%	100.00%

Appendix

Table 7 : Statistics of generated intermediate questions (IQs) in ReHoPER on iLLC . We report the total number of generated IQs, the average number of IQs per path with standard deviation, the number of unique IQs, the average number of unique IQs per path with standard deviation, and the corresponding uniqueness percentages.

Model	#IQ	#IQ/Path std	#uniqIQ	#uniqIQ/Path std	%uniqIQ	%uniqIQ/Path
	MuSR MM
Gemma 3	48.73	8.12 1.07	48.53	8.12 1.07	99.58%	100.00%
gpt-oss	59.70	9.95 0.10	59.62	9.95 0.10	99.87%	99.99%
Llama 3.1 8B	46.69	7.78 2.42	46.64	7.78 2.42	99.90%	99.97%
Llama 3.1 70B	50.04	8.34 1.20	49.74	8.34 1.20	99.41%	100.00%
Mistral Small	58.78	9.80 0.38	58.12	9.79 0.39	98.87%	99.94%

Appendix

Table 8 : Statistics of generated intermediate questions (IQs) in ReHoPER without replanning on MuSR MM, MoreHopQA, and iLLC L2-4. We report the total number of generated IQs, the average number of IQs per path with standard deviation, the number of unique IQs, the average number of unique IQs per path with standard deviation, and the corresponding uniqueness percentages.

Model	#IQ	#IQ/Path std	#uniqIQ	#uniqIQ/Path std	%uniqIQ	%uniqIQ/Path
	MuSR MM
Gemma 3	60.00	10.00 0.00	58.76	9.95 0.09	97.93%	99.53%
gpt-oss	59.97	10.00 0.01	59.29	9.88 0.21	98.86%	98.88%
Llama 3.1 8B	60.00	10.00 0.00	57.86	9.69 0.47	96.43%	96.87%
Llama 3.1 70B	60.00	10.00 0.00	59.11	9.95 0.10	98.51%	99.49%
Mistral Small	60.00	10.00 0.00	57.53	9.62 0.52	95.88%	96.15%

Appendix

Table 9 : Statistics of generated intermediate questions (IQs) in ReHoPER without look-ahead on MuSR MM, MoreHopQA, and iLLC L2-4. We report the total number of generated IQs, the average number of IQs per path with standard deviation, the number of unique IQs, the average number of unique IQs per path with standard deviation, and the corresponding uniqueness percentages.

Step	T std	%T/N std	R std	%R/T std	%R/N std	T std	%T/N std	R std	%R/T std	%R/N std	T std	%T/N std	R std	%R/T std	%R/N std
	Gemma 3
	MuSR MM (N: 250)					MoreHopQA (N: 150)					iLLC L2-4 (N: 250)
1 $\rightarrow$ 2	250.00 0.00	100.0 0.0	249.83 0.37	99.9 0.1	99.9 0.1	138.17 2.27	92.1 1.5	122.50 3.69	88.7 3.0	81.7 2.5	250.00 0.00	100.0 0.0	209.33 21.02	83.7 8.4	83.7 8.4
2 $\rightarrow$ 3	250.00 0.00	100.0 0.0	250.00 0.00	100.0 0.0	100.0 0.0	124.33 2.49	82.9 1.7	105.50 2.14	84.9 2.6	70.3 1.4	250.00 0.00	100.0 0.0	131.00 34.24	52.4 13.7	52.4 13.7
3 $\rightarrow$ 4	250.00 0.00	100.0 0.0	249.83 0.37	99.9 0.1	99.9 0.1	118.17 3.13	78.8 2.1	101.33 3.68	85.8 2.0	67.6 2.5	248.17 1.21	99.3 0.5	73.83 16.87	29.7 6.8	29.5 6.7
4 $\rightarrow$ 5	250.00 0.00	100.0 0.0	250.00 0.00	100.0 0.0	100.0 0.0	106.00 5.32	70.7 3.5	90.50 3.91	85.4 2.3	60.3 2.6	241.17 7.60	96.5 3.0	74.83 18.96	31.1 8.0	29.9 7.6

Appendix

Table 10 : Trajectory statistics under the second-next analysis.

T

is the number of valid step transitions,

R

is the number of replanned transitions, and

N

is the total number of instances.

Step	T std	%T/N std	R std	%R/T std	%R/N std	T std	%T/N std	R std	%R/T std	%R/N std	T std	%T/N std	R std	%R/T std	%R/N std
	Llama 3.1 8B
	MuSR MM (N: 250)					MoreHopQA (N: 150)					iLLC L2-4 (N: 250)
1 $\rightarrow$ 2	250.00 0.00	100.0 0.0	250.00 0.00	100.0 0.0	100.0 0.0	109.67 5.12	73.1 3.4	101.83 4.52	92.9 1.4	67.9 3.0	221.83 8.63	88.7 3.5	219.17 9.10	98.8 1.0	87.7 3.6
2 $\rightarrow$ 3	250.00 0.00	100.0 0.0	249.00 1.00	99.6 0.4	99.6 0.4	105.67 4.15	70.4 2.8	99.00 6.14	93.6 2.8	66.0 4.1	210.33 9.62	84.1 3.8	199.00 9.27	94.6 2.3	79.6 3.7
3 $\rightarrow$ 4	249.67 0.47	99.9 0.2	248.83 0.90	99.7 0.3	99.5 0.4	94.67 2.49	63.1 1.7	88.67 2.21	93.7 2.3	59.1 1.5	201.83 11.91	80.7 4.8	186.50 9.39	92.5 1.0	74.6 3.8
4 $\rightarrow$ 5	250.00 0.00	100.0 0.0	249.00 0.82	99.6 0.3	99.6 0.3	79.67 2.29	53.1 1.5	74.17 2.91	93.1 1.7	49.4 1.9	176.50 13.46	70.6 5.4	154.50 13.51	87.5 1.9	61.8 5.4

Appendix

Table 11 : Trajectory statistics under the second-next analysis.

T

is the number of valid step transitions,

R

is the number of replanned transitions, and

N

is the total number of instances.

	iLLC L2-4 (Development Set)												MoreHopQA
	IO				CoT				PS				IO				CoT				PS
Model	T=.2	T=.3	T=.5	T=.7	T=.2	T=.3	T=.5	T=.7	T=.2	T=.3	T=.5	T=.7	T=.2	T=.3	T=.5	T=.7	T=.2	T=.3	T=.5	T=.7	T=.2	T=.3	T=.5	T=.7
Gemma 3	6.8	6.8	5.6	7.6	5.6	7.6	5.6	8.4	4.0	4.8	6.0	6.8	64.7	67.3	64.0	66.0	66.0	66.0	66.7	67.3	70.7	71.3	74.7	73.3
gpt-oss	90.8	92.4	93.2	94.8	89.2	92.0	93.6	94.8	89.6	92.8	91.2	93.6	78.7	78.7	81.3	80.0	78.0	79.3	80.0	80.0	80.0	80.7	82.0	79.3
Llama 3.1 8B	10.8	8.8	16.8	20.0	16.4	12.4	18.4	11.2	40.4	30.0	38.4	42.4	57.3	58.7	56.7	55.3	56.0	57.3	58.0	62.0	58.7	60.7	63.3	64.7
Mistral Small	8.0	7.2	9.2	7.2	5.6	8.8	9.2	6.0	8.8	8.4	8.4	6.4	66.0	70.7	71.3	70.0	67.3	68.0	67.3	72.0	70.7	70.7	70.7	68.0

Appendix

Table 12 : Accuracy (%) of ReHoPER under different sampling temperatures (T) on iLLC L2-4 (development set) and MoreHopQA, using IO, CoT, and PS answer-format prefixes. Green and light green mark the best and second best result, respectively, for each (model, dataset, prompting-style) triplet.

	iLLC L2-4 (Development Set)												MoreHopQA
	IO				CoT				PS				IO				CoT				PS
Model	5-12	6-10	10-6	12-5	5-12	6-10	10-6	12-5	5-12	6-10	10-6	12-5	5-12	6-10	10-6	12-5	5-12	6-10	10-6	12-5	5-12	6-10	10-6	12-5
Gemma 3	6.0	6.8	5.6	4.8	8.0	8.0	5.6	5.6	6.4	7.6	6.0	5.6	66.0	64.7	64.0	62.7	66.7	64.7	66.7	64.0	74.0	74.7	74.7	74.0
gpt-oss	94.0	94.4	93.2	94.8	93.6	93.6	93.6	93.6	93.2	94.4	91.2	92.8	78.0	78.0	81.3	80.0	78.0	81.3	80.0	79.3	79.3	82.0	82.0	78.7
Llama 3.1 8B	23.6	21.2	16.8	10.8	19.6	20.8	18.4	16.0	38.4	40.0	38.4	33.6	54.7	56.0	56.7	54.7	60.7	59.3	58.0	56.0	65.3	63.3	63.3	61.3
Mistral Small	10.0	10.0	9.2	5.6	12.8	12.8	9.2	8.0	11.2	9.6	8.4	6.4	70.7	71.3	71.3	72.0	67.3	67.3	67.3	67.3	71.3	70.7	70.7	71.3

Appendix

Table 13 : Accuracy (%) of ReHoPER under different {

n

}-{

m_{\max}

} configurations on iLLC L2-4 (development set) and MoreHopQA, using IO, CoT, and PS answer-format prefixes. Green and light green mark the best and second best result, respectively, for each (model, dataset, prompting-style) triplet.

Model	SC	RH	SC	RH	SC	RH	SC	RH	SC	RH	SC	RH
	MoreHopQA						iLLC L2-4
	IO		CoT		PS		IO		CoT		PS
Gemma 3	61.8 0.4	65.8 1.0	64.7 0.7	68.7 0.0	66.7 0.7	74.7 1.2	7.3 0.2	6.9 1.3	7.6 0.0	6.9 0.9	3.6 0.8	8.1 1.7
Llama 3.1 8B	60.7 1.2	57.1 1.7	60.0 0.7	60.0 1.8	62.4 1.5	63.1 0.4	11.9 0.8	17.5 4.1	12.5 3.2	18.4 2.0	30.4 1.1	41.6 1.4
Mistral Small	70.2 0.4	71.3 0.0	69.1 1.0	68.7 1.2	72.2 0.4	71.1 1.4	1.9 0.5	10.5 0.6	2.4 0.4	9.6 1.7	3.1 0.8	10.8 1.2
Phi-4	77.3 0.7	76.0 1.8	76.4 0.4	80.4 1.0	77.6 0.4	78.4 1.7	13.2 0.4	54.3 0.8	10.0 0.4	62.0 1.7	9.7 0.8	62.7 1.3

Appendix

Table 14 : Robustness of ReHoPER (RH) across repeated runs on MoreHopQA and iLLC L2-4 under IO, CoT, and PS answer-format prefixes. We report accuracy (%) as mean std over three runs. SC denotes simple question answering with Self-Consistency using the corresponding output prefix, while RH applies our planning approach but evaluates with the same prefix. Green , red , and gray indicate performance changes relative to SC.

	iLLC L1-8							iLLC L2-8
		IO		CoT		PS			IO		CoT		PS
Model	SD	SC	RH	SC	RH	SC	RH	SD	SC	RH	SC	RH	SC	RH
Gemma 3	26.4	7.6	12.4	36.0	28.8	52.8	32.8	0.0	0.4	0.0	0.0	0.0	0.0	0.0
gpt-oss	58.8	84.0	76.4	78.4	74.8	77.6	76.0	45.6	76.4	80.0	61.6	80.8	6.0	81.2
Llama 3.1 8B	7.6	31.2	27.6	24.0	39.2	39.2	34.4	0.0	0.8	2.0	0.0	4.0	0.4	2.4
Llama 3.1 70B	56.0	59.2	62.8	48.8	63.2	68.8	70.4	0.8	1.2	44.0	3.6	46.4	12.8	54.0

Appendix

Table 15 : Accuracy (%) of ReHoPER (RH) across iLLC L1-8 and iLLC L2-8 benchmarks under IO, CoT, and PS answer-format prefixes. SD refers to the Self-Discover baseline. Furthermore, SC denotes simple question answering with self-consistency using the corresponding output prefix, while RH applies our planning approach but evaluates with the same prefix. Green , red , and gray indicate performance changes relative to SC.

Model (as used)	Org / Family	Params	Context
google/gemma-3-12b-it	Google / Gemma 3	12B	128K
openai/gpt-oss-20b	OpenAI / gpt-oss	21B (3.6B active)	128K
meta-llama/Llama-3.1-8B-Instruct	Meta / Llama 3.1	8B	128K
hugging-quants/Meta-Llama-3.1-70B-Instruct-AWQ-INT4	Meta / Llama 3.1	70B	128K
stelterlab/Mistral-Small-24B-Instruct-2501-AWQ	Mistral / Mistral Small 3	24B	32K
microsoft/phi-4	Microsoft / Phi	14B	16K

Appendix

Table 16 : Models used in our experiments. We list the exact checkpoint, model family, parameter scale, and context length reported by the corresponding model card.

Panel of Figure 13 — Figure 13 : Instruction $I_{\mathrm{rh}}$ used for ReHoPER.

Panel of Figure 14 — Figure 14 : Instruction $I_{\mathrm{fixed}}$ used for the ablation without replanning.

Panel of Figure 15 — Figure 15 : Instruction $I_{\mathrm{single}}$ used for the ablation without look-ahead.

Dataset	Task	Test Set Size	Answer Format
Dataset	Task	Per Task	Answer Format
MuSR	Murder Mysteries, Team Allocation	250	Choice number
MuSR	Object Placement	256	Choice number
MoreHopQA	-	150	Short text answer
iLLC	L1-4, L1-6, L1-8, L2-4, L2-6, L2-8	250	Concatenated letters

Appendix

Table 17: Summary of reasoning datasets used for evaluation.

Explore similar work

PoTRE: Test-Time Reasoning inspired by Cognitive Heterogeneity
Jul 22, 2026Anmol Kankariya, Sercan Ö. ArıkLLM Reasoning Strategies Large Language Model Planning
Applied AI, Google Cloud
Learning Composable Chains-of-Thought
May 28, 2025Fangcong Yin, Zeyu Leo Liu, Liu Leqi +2LLM Reasoning Strategies Chain-of-Thought Reasoning
The University of Texas at Austin · Princeton University
HyperGuide: Hyperbolic Guidance for Efficient Multi-Step Reasoning in Large Language Models
May 22, 2026Yuyu Liu, Haotian Xu, Yanan He +3LLM Reasoning Strategies Few-Step Generation
Department of Computer Science Stony Brook University · Department of Applied Mathematics and Statistics Stony Brook University · Department of Computer Science Yale University +2

Jul 22, 2026, cs.AIJ/K move · Enter open · S save

PoTRE: Test-Time Reasoning inspired by Cognitive Heterogeneity

Anmol Kankariya, Sercan Ö. Arık

Applied AI, Google Cloud

LLM Reasoning Strategies Large Language Model Planning Classical Planning Cognitive Science Ensemble

Open paper AskPDF arXiv

While Large Language Models (LLMs) excel at many tasks, they frequently struggle with complex reasoning that requires long-horizon planning and iterative error correction. Furthermore, standard single-stream prompting proves brittle when models encounter novel abstractions or rigorous domain constraints. We introduce PoTRE (Poly-Topological Reasoning Ensembles), a heterogeneous framework that decouples inference into four agents: (1) Adversarial Refinement Agent, (2) Hierarchical strategic Planning Agent, (3) Spectrum Search Agent, and (4) Direct Chain Agent. A final Task-Adaptive Aggregation Layer dynamically reconciles these perspectives -- via final candidate selection, semantic synthesis, or neuro-symbolic verification -- to produce a robust global solution. We evaluate PoTRE on three frontier benchmarks: ARC-AGI-2, Humanity's Last Exam (HLE), and PRBench Finance. PoTRE achieves state-of-the-art accuracy of 49.92% on HLE, surpassing the previous best official score. We demonstrate that this architectural heterogeneity achieves improved reasoning performance using similar or fewer inference tokens compared to heavily scaled homogeneous baselines.

Abstract

Explore similar work

PoTRE: Test-Time Reasoning inspired by Cognitive Heterogeneity

Learning Composable Chains-of-Thought

HyperGuide: Hyperbolic Guidance for Efficient Multi-Step Reasoning in Large Language Models