Long-horizon agent learning

SCAD: Structured Credit Assignment and Distillation for Long-Horizon Agents

Shangyang Wu¹
Shuai Zhao^1,†
Ziyue Zhu¹
Jinyang Wu²
Anh Tuan Luu³
Haoran Luo^3,†

¹Beijing University of Posts and Telecommunications²Tsinghua University³Nanyang Technological University

^† Corresponding authors

Abstract

Training long-horizon agents to solve complex tasks requires effective supervision over extended interaction sequences. However, sparse terminal rewards obscure intermediate contributions, while on-policy distillation can lose informative teacher guidance as student-generated histories grow. To address this problem, we introduce SCAD, which organizes interactions into planning and bounded subtask execution, distills execution in local contexts, and refines planning credit through cross-rollout subtask prefix trees, with planning receiving full terminal credit and execution receiving positive terminal credit and teacher guidance. Across all evaluated benchmarks, SCAD improves macro-average accuracy over the strongest training baseline by 4.48 percentage points for text tasks and 4.19 points for multimodal tasks. SCAD effectively combines outcome-based credit assignment with teacher-guided distillation to improve planning and execution in long-horizon agents.

Method

Subtask-local distillation

Sustains useful teacher guidance over long horizons by distilling execution within bounded subtask-local contexts.

Cross-rollout planning credit

Refines credit for individual planning decisions by comparing terminal returns across rollouts with matched subtask–report histories.

Structured signal allocation

Preserves useful execution guidance in failed trajectories by retaining signed planning credit while blocking negative terminal credit to execution.

SCAD framework: Planner–Worker rollouts, subtask-local teacher supervision, cross-rollout prefix-tree credit, and structured signal allocation. — SCAD training framework. Click the figure to enlarge.

Main results

Three student settings across 19 text and multimodal benchmarks.

Qwen3-4B

Teacher: Qwen3-4B-Instruct-2507 · 11 benchmarks

Qwen3-4B: per-benchmark and overall macro accuracy in percent, mean and sample standard deviation over three seeds.
Method	BrowseComp-Plus	WebWalkerQA	HotpotQA	2WikiMultiHopQA	MuSiQue	NQ	TriviaQA	BamboogleOOD	PopQAOOD	GAIAOOD	xbenchOOD	Overall avg.
Teacher	23.61 ± 0.5	36.67 ± 3.1	60.51 ± 0.9	52.31 ± 1.5	11.28 ± 0.9	68.72 ± 1.8	85.13 ± 0.9	56.92 ± 0.0	56.92 ± 0.0	25.42 ± 0.0	23.67 ± 0.6	45.56 ± 0.6
Vanilla	0.00 ± 0.0	2.92 ± 0.7	31.28 ± 0.9	21.54 ± 0.0	11.79 ± 0.9	28.21 ± 0.9	46.67 ± 1.8	24.10 ± 0.9	21.03 ± 0.9	2.82 ± 1.0	3.67 ± 0.6	17.64 ± 0.5
GRPO	19.44 ± 4.6	28.75 ± 1.3	40.00 ± 1.5	46.15 ± 1.5	10.77 ± 1.5	63.59 ± 1.8	75.90 ± 1.8	32.82 ± 5.4	42.05 ± 6.2	15.25 ± 0.0	25.00 ± 3.6	36.34 ± 1.4
FoldGRPO	17.22 ± 0.5	34.58 ± 5.1	52.31 ± 5.5	52.31 ± 1.5	13.33 ± 0.9	63.08 ± 0.0	71.28 ± 1.8	41.54 ± 1.5	55.38 ± 1.5	13.56 ± 6.1	21.00 ± 1.0	39.60 ± 0.9
Tree-GRPO	18.33 ± 2.2	35.83 ± 0.7	50.77 ± 0.0	51.28 ± 1.8	18.46 ± 1.5	61.03 ± 0.9	78.46 ± 8.1	38.46 ± 4.1	52.31 ± 1.5	18.64 ± 1.7	25.33 ± 4.5	40.81 ± 0.3
HiPER	21.67 ± 2.5	32.92 ± 3.1	54.36 ± 4.7	46.15 ± 4.1	16.92 ± 1.5	66.15 ± 1.5	73.85 ± 1.5	43.59 ± 0.9	53.85 ± 1.5	18.64 ± 1.7	28.33 ± 1.2	41.49 ± 0.5
SFT	19.17 ± 3.8	33.33 ± 7.5	47.69 ± 1.5	48.21 ± 5.4	11.79 ± 1.8	63.08 ± 1.5	71.28 ± 0.9	40.00 ± 1.5	52.31 ± 1.5	17.51 ± 9.3	22.33 ± 1.2	38.79 ± 2.2
OPD	19.72 ± 3.4	28.75 ± 1.3	51.79 ± 0.9	45.64 ± 1.8	13.85 ± 1.5	65.64 ± 5.4	71.79 ± 5.4	37.44 ± 0.9	47.18 ± 6.2	11.86 ± 1.7	25.00 ± 0.0	38.06 ± 1.0
SDAR	15.83 ± 3.8	35.00 ± 1.3	44.10 ± 0.9	54.87 ± 1.8	8.72 ± 7.6	58.46 ± 1.5	82.56 ± 6.2	35.90 ± 1.8	56.41 ± 5.4	15.82 ± 1.0	26.67 ± 1.2	39.49 ± 2.2
ATOD	21.39 ± 3.4	33.75 ± 3.3	53.85 ± 1.5	50.77 ± 0.0	13.85 ± 1.5	64.10 ± 0.9	75.90 ± 1.8	45.13 ± 8.5	56.92 ± 1.5	17.51 ± 8.4	24.67 ± 0.6	41.62 ± 1.0
TCOD	20.83 ± 0.0	30.83 ± 0.7	47.69 ± 1.5	49.23 ± 1.5	11.79 ± 0.9	59.49 ± 0.9	80.00 ± 0.0	41.54 ± 8.1	50.77 ± 1.5	14.12 ± 1.0	22.00 ± 0.0	38.94 ± 0.7
MemOPD	21.67 ± 3.0	36.25 ± 1.3	50.26 ± 0.9	46.67 ± 1.8	14.87 ± 1.8	62.56 ± 0.9	76.92 ± 5.5	44.10 ± 0.9	56.92 ± 1.5	22.60 ± 1.0	20.67 ± 3.5	41.23 ± 0.9
SCAD (Ours)	23.06 ± 2.9	40.42 ± 2.6	55.38 ± 1.5	56.41 ± 2.4	18.97 ± 1.8	71.28 ± 3.9	84.10 ± 1.8	45.64 ± 0.9	58.97 ± 6.2	23.16 ± 3.5	29.67 ± 0.6	46.10 ± 1.2

Qwen3-1.7B

Teacher: Qwen3-4B-Instruct-2507 · 11 benchmarks

Qwen3-1.7B: per-benchmark and overall macro accuracy in percent, mean and sample standard deviation over three seeds.
Method	BrowseComp-Plus	WebWalkerQA	HotpotQA	2WikiMultiHopQA	MuSiQue	NQ	TriviaQA	BamboogleOOD	PopQAOOD	GAIAOOD	xbenchOOD	Overall avg.
Teacher	23.61 ± 0.5	36.67 ± 3.1	60.51 ± 0.9	52.31 ± 1.5	11.28 ± 0.9	68.72 ± 1.8	85.13 ± 0.9	56.92 ± 0.0	56.92 ± 0.0	25.42 ± 0.0	23.67 ± 0.6	45.56 ± 0.6
Vanilla	1.11 ± 0.5	2.92 ± 0.7	17.95 ± 0.9	15.38 ± 1.5	5.13 ± 0.9	26.15 ± 1.5	31.28 ± 0.9	12.82 ± 0.9	14.36 ± 0.9	3.95 ± 1.0	6.00 ± 1.0	12.46 ± 0.2
GRPO	9.44 ± 0.5	27.92 ± 1.9	41.03 ± 2.4	28.21 ± 3.2	7.69 ± 3.1	40.51 ± 2.4	73.85 ± 3.1	24.10 ± 0.9	39.49 ± 1.8	7.34 ± 2.0	25.67 ± 0.6	29.57 ± 0.6
FoldGRPO	14.17 ± 0.8	31.67 ± 2.6	45.13 ± 3.2	37.44 ± 0.9	10.26 ± 0.9	47.18 ± 2.4	77.44 ± 1.8	26.67 ± 4.4	44.62 ± 1.5	9.04 ± 2.6	29.33 ± 1.5	33.90 ± 1.0
Tree-GRPO	15.00 ± 0.8	29.17 ± 1.4	48.72 ± 3.2	39.49 ± 1.8	10.26 ± 3.2	43.59 ± 1.8	79.49 ± 2.4	29.74 ± 3.6	43.59 ± 0.9	11.30 ± 2.0	29.33 ± 0.6	34.52 ± 0.3
HiPER	16.39 ± 1.3	32.50 ± 1.2	42.56 ± 3.2	39.49 ± 1.8	10.77 ± 2.7	45.64 ± 3.9	80.00 ± 4.1	30.26 ± 0.9	42.56 ± 3.2	10.17 ± 3.4	30.00 ± 3.0	34.58 ± 1.7
SFT	11.11 ± 1.0	27.92 ± 4.4	41.03 ± 0.9	29.74 ± 2.4	7.69 ± 1.5	46.15 ± 1.5	76.41 ± 1.8	27.18 ± 2.4	39.49 ± 1.8	7.34 ± 2.6	24.33 ± 3.2	30.76 ± 1.1
OPD	15.00 ± 3.0	30.42 ± 4.7	42.56 ± 3.9	37.44 ± 1.8	7.69 ± 1.5	47.69 ± 1.5	81.54 ± 1.5	28.72 ± 3.2	44.10 ± 3.2	9.04 ± 1.0	29.33 ± 2.5	33.96 ± 1.9
SDAR	11.67 ± 1.7	31.67 ± 0.7	40.00 ± 4.1	37.44 ± 0.9	6.67 ± 2.4	43.08 ± 1.5	81.03 ± 3.2	26.67 ± 3.9	42.56 ± 1.8	13.56 ± 1.7	28.67 ± 2.1	33.00 ± 1.3
ATOD	17.78 ± 1.3	35.00 ± 4.3	44.62 ± 1.5	41.54 ± 1.5	9.23 ± 1.5	45.64 ± 2.4	80.00 ± 4.6	30.26 ± 1.8	46.15 ± 3.1	7.34 ± 4.3	31.00 ± 1.0	35.32 ± 0.5
TCOD	18.33 ± 1.4	33.33 ± 3.1	43.59 ± 0.9	36.92 ± 6.7	7.18 ± 2.4	48.21 ± 3.6	81.54 ± 1.5	28.21 ± 1.8	45.64 ± 3.9	11.86 ± 1.7	30.67 ± 2.3	35.04 ± 1.7
MemOPD	15.28 ± 0.5	35.42 ± 0.7	43.08 ± 3.1	38.97 ± 2.4	6.67 ± 3.6	48.72 ± 3.9	77.44 ± 1.8	28.72 ± 2.4	45.13 ± 1.8	10.73 ± 2.6	30.67 ± 2.9	34.62 ± 0.9
SCAD (Ours)	19.17 ± 1.7	42.50 ± 2.5	50.26 ± 2.4	47.18 ± 3.2	11.79 ± 2.4	50.26 ± 2.4	84.62 ± 3.1	31.79 ± 2.4	47.18 ± 2.4	14.69 ± 2.6	36.00 ± 2.0	39.58 ± 1.1

Qwen3-VL-2B

Teacher: Qwen3-VL-4B · 8 benchmarks

Qwen3-VL-2B: per-benchmark and overall macro accuracy in percent, mean and sample standard deviation over three seeds.
Method	MMSearch-Plus	VDR-Bench	VisBrowse-Bench	MMSearch	SimpleVQA	ID avg.	FVQAOOD	InfoSeekOOD	LiveVQAOOD	Overall avg.
Teacher	16.19 ± 1.0	16.79 ± 0.9	5.64 ± 1.8	40.61 ± 1.0	45.71 ± 2.9	24.99 ± 1.0	56.19 ± 1.6	37.14 ± 2.9	28.89 ± 2.2	30.90 ± 1.0
Vanilla	0.95 ± 0.0	0.74 ± 0.7	1.54 ± 1.5	23.64 ± 1.8	20.00 ± 2.9	9.37 ± 1.2	25.71 ± 2.9	11.43 ± 0.0	8.89 ± 0.0	11.61 ± 0.4
GRPO	3.49 ± 2.4	3.46 ± 1.9	0.51 ± 0.9	40.00 ± 1.8	42.86 ± 0.0	18.06 ± 0.5	47.62 ± 1.6	27.62 ± 1.6	19.26 ± 2.6	23.10 ± 0.4
MMSearch-R1	9.52 ± 1.0	12.35 ± 1.9	5.13 ± 0.9	40.61 ± 4.6	46.67 ± 10.0	22.85 ± 3.5	56.19 ± 1.6	32.38 ± 1.6	22.22 ± 2.2	28.13 ± 2.6
VSearcher	10.16 ± 3.3	14.32 ± 6.0	6.15 ± 1.5	37.58 ± 2.1	47.62 ± 1.6	23.17 ± 1.5	51.43 ± 2.9	36.19 ± 1.6	23.70 ± 2.6	28.39 ± 1.7
SFT	4.76 ± 4.4	6.67 ± 0.7	2.56 ± 2.4	35.76 ± 6.4	39.05 ± 1.6	17.76 ± 0.9	48.57 ± 0.0	24.76 ± 1.6	17.04 ± 7.8	22.40 ± 0.7
OPD	7.30 ± 3.3	9.88 ± 0.4	2.56 ± 0.9	33.33 ± 4.6	38.10 ± 1.6	18.23 ± 1.7	54.29 ± 2.9	27.62 ± 1.6	19.26 ± 1.3	24.04 ± 0.8
VOLD	10.79 ± 3.8	10.12 ± 2.6	3.08 ± 1.5	36.36 ± 1.8	49.52 ± 1.6	21.98 ± 0.3	50.48 ± 3.3	33.33 ± 1.6	26.67 ± 2.2	27.54 ± 0.9
HyperEyes	11.43 ± 2.5	12.59 ± 0.7	5.64 ± 3.9	36.36 ± 1.8	49.52 ± 10.0	23.11 ± 3.0	53.33 ± 1.6	35.24 ± 1.6	27.41 ± 5.6	28.94 ± 2.6
SCAD (Ours)	15.24 ± 3.8	15.56 ± 0.7	8.21 ± 6.4	41.21 ± 6.4	51.43 ± 2.9	26.33 ± 1.7	59.05 ± 1.6	44.76 ± 7.2	29.63 ± 1.3	33.13 ± 1.5

Accuracy (%), mean ± sample standard deviation over three seeds. Overall averages weight each benchmark equally. Bold values indicate the best student mean in each column, excluding the teacher. OOD benchmarks are absent from training. Scroll horizontally to view all benchmarks.

Citation

BibTeXShow citation

Download .bib

@misc{wu2026scad,
  title  = {SCAD: Structured Credit Assignment and
            Distillation for Long-Horizon Agents},
  author = {Wu, Shangyang and Zhao, Shuai and Zhu, Ziyue
            and Wu, Jinyang and Luu, Anh Tuan
            and Luo, Haoran},
  year   = {2026},
  eprint = {2610.03372},
  archivePrefix = {arXiv},
  primaryClass = {cs.LG},
  url    = {https://arxiv.org/abs/2610.03372}
}