One Readout, Many Repairs: Diffusion-Guided Hierarchical Search for Tool-Agent Repair

Xiang Xia; Cheng Yan; Wuyang Zhang; Fan Xu; Zhijun Fan; Shuyuan Zhang; Yanyong Zhang

cs.AISep 28, 2026

One Readout, Many Repairs: Diffusion-Guided Hierarchical Search for Tool-Agent Repair

Authors: Xiang Xia, Cheng Yan, Wuyang Zhang, Fan Xu, Zhijun Fan, Shuyuan Zhang, Yanyong Zhang

Organizations: University of Science and Technology of China, Hefei, China

Abstract

Tool agents use large language models to act through external tools, yet successfully executed calls can still leave user requests unfulfilled. Tool-agent repair seeks alternative call sequences that execute successfully and fulfill the original requests. However, repair requires exploring both operation choices and their concrete realizations, making complete-sequence regeneration costly. Moreover, regeneration repeats operation selection even when failure arises from how those operations are realized. The resulting challenge is to reduce this repetition while preserving exploration of alternative operations and realizations. Therefore, we formulate repair as hierarchical search over operation supports, which we introduce as sets of permitted operation types that define reusable search regions for concrete tool-call sequences. We propose ReCommit, a training-free, diffusion-guided framework for improving tool-agent failure recovery while reducing repair computation. ReCommit amortizes operation-level proposal computation across repair trials by reusing operation-type scores from a single parallel readout of a masked diffusion language model. These scores guide search across supports, while realization search explores alternative entity bindings, arguments, and action composition within each support. Experiments on real failures across four enterprise services in the Agent-Diff benchmark show 75.9% and 63.2% relative recovery gains with 61.3% and 51.3% reductions in mean full-budget repair time at repair budgets $B=3$ and $B=13$ , respectively, over the strongest evaluated 8B comparison method. ReCommit achieves a favorable recovery--cost trade-off, including in comparisons with the evaluated 32B models.

Figures & tables

Figure 1: Two distinct repair needs from real failures in Agent-Diff ( Pyskło et al., 2026 ) . (a) In Linear, the failed attempt omits issueUpdate , which is needed to update the target issue. (b) In Slack, both realizations use chat.postMessage , but only the structured user mention satisfies the requested representation. Calls are abbreviated.

Figure 2: Overview of ReCommit. AR denotes autoregressive. Left: a single parallel dLLM readout supplies slot-level operation predictions, pooled into inclusion scores for support enumeration and ranking. Center: ranked supports define realization regions for different call sequences. Right: canonical realization $C(S_{i};x)$ grounds a schema-based call structure, while expressive realization $E(S_{i};x)$ searches alternative entity bindings, arguments, and action composition within the same support. Both modes use the original public context without cross-mode generation feedback. Bottom: repair trials follow the fixed order $C(S_{1};x),E(S_{1};x),C(S_{2};x),E(S_{2};x),\ldots$ . Within each trial, public-error pruning removes rejected calls and replays the remaining nonempty sequence once from the initial state.

Scale	Method	Budget $B=3$						Budget $B=13$
Scale	Method	Box	Calendar	Linear	Slack	Rate $\uparrow$	Time $\downarrow$	Box	Calendar	Linear	Slack	Rate $\uparrow$	Time $\downarrow$
32B	Direct	12	12	17	57	15.5	108.0	22	14	17	65	18.7	487.1
	Schema	30	23	25	65	22.7	143.1	42	31	27	67	26.5	649.8
	Feedback-Reflective	41	20	32	70	25.8	350.4	56	29	34	85	32.3	1528.3
8B	Direct	4	2	3	26	5.5	40.8	8	2	3	34	7.4	177.7
	Schema	10	7	18	31	10.5	44.8	13	9	18	36	12.0	197.0

Table 1: Fixed-budget recovery at

B=3

and

B=13

. Service columns report recovered episodes from Box (129), Calendar (180), Linear (156), and Slack (166). Recovery rate is episode-weighted (%), and time denotes mean full-budget repair time in seconds. Best and second-best values within each model-scale group at each budget are bolded and underlined . ReCommit uses LLaDA-8B for proposal and Qwen3-8B for realization. Comparison methods use Qwen3-8B and Qwen3-32B.

$B$	Schedule	Supports	Recovered episodes $\uparrow$
$B$	Schedule	Supports	With-P	Without-P
$3$	Canonical-only	3	127	117
	Expressive-only	3	71	70
	Canonical resampling	2	122	108
	ReCommit	2	139	128
$13$	Canonical-only	13	141	137

Table 2: Realization-search recovery. Supports denotes the number of distinct supports visited. Without-P evaluates realizations without pruning. Trial budgets are equal, but computation differs. Best counts at each budget are bolded .

Appendix figures & tables15 assets

Supplementary material from the paper’s appendix.

Appendix

Service	Public tasks	Failure episodes
Box	40	129
Calendar	53	180
Linear	50	156
Slack	51	166
Total	194	631

Appendix

Table 3: Composition of the Agent-Diff repair evaluation pool. The task column counts distinct benchmark tasks; failure episodes are deduplicated unsuccessful tool-use attempts retained for repair. All methods are evaluated on the same pool.

Backbone	Repository identifier
Dream	Dream-org/Dream-v0-Instruct-7B
LLaDA 1.5	GSAI-ML/LLaDA-1.5
DiffuCoder	apple/DiffuCoder-7B-Instruct
SDAR	JetLM/SDAR-8B-Chat-b8
Nemotron	nvidia/Nemotron-Labs-Diffusion-8B

Appendix

Table 4: Diffusion backbones evaluated for support proposal beyond the default LLaDA-8B-Instruct. Repository identifiers specify the public Hugging Face checkpoints; all configurations use Qwen3-8B for realization under the same search schedule.

Service	Types	Operation vocabulary
Box	20	DELETE /collections/{id} , DELETE /comments/{id} , DELETE /files/{id} , DELETE /folders/{id} , DELETE /hubs/{id} , DELETE /tasks/{id} , POST /collections , POST /comments , POST /files/content , POST /files/{id}/content , POST /folders , POST /hubs , POST /hubs/{id}/manage_items , POST /tasks , PUT /collections/{id} , PUT /comments/{id} , PUT /files/{id} , PUT /folders/{id} , PUT /hubs/{id} , PUT /tasks/{id} .
Calendar	25	acl.delete , acl.insert , acl.patch , acl.update , acl.watch , calendarList.delete , calendarList.insert , calendarList.patch , calendarList.update , calendarList.watch , calendars.clear , calendars.delete , calendars.insert , calendars.patch , calendars.update , channels.stop , events.delete , events.import , events.insert , events.move , events.patch , events.quickAdd , events.update , events.watch , settings.watch .
Linear	13	commentCreate , commentDelete , commentUpdate , issueCreate , issueLabelCreate , issueLabelDelete , issueLabelUpdate , issueRelationCreate , issueUpdate , teamCreate , teamMembershipCreate , workflowStateArchive , workflowStateCreate .
Slack	14	chat.delete , chat.postMessage , chat.update , conversations.archive , conversations.create , conversations.invite , conversations.join , conversations.kick , conversations.leave , conversations.rename , conversations.setTopic , conversations.unarchive , reactions.add , reactions.remove .

Appendix

Table 5: State-changing operation types available to support search in each service. Types gives the vocabulary size; each operation support is a subset of the listed types. Read-only calls remain available for public lookups during realization.

Canonical $K$	Expressive $D$	With-P	Without-P
13	0	141	137
12	1	154	151
11	2	156	153
10	3	155	151
9	4	157	152
8	5	156	151

Appendix

Table 6: Recovery under canonical/expressive trial allocations at

B=13

. Each row uses the first

K

canonical and first

D

expressive realizations, with

K+D=13

and

D\leq K

. With-P and Without-P report recovered episodes with and without public-error pruning. Trial budgets are equal, but computation differs. The default

(K,D)=(7,6)

is shaded; column maxima are bolded.

Method	Executable trials	ERR (%)	Recovered episodes
32B comparison methods
Direct (32B)	771	40.73	98
Schema (32B)	871	46.01	143
Feedback-Reflective (32B)	1,067	56.37	163
8B methods
Direct	611	32.28	35

Appendix

Table 7: Execution validity and task recovery at

B=3

. The executable-trial count includes nonempty final realizations with no rejected calls. Executable realization rate (ERR) divides this count by the 1,893 repair trials. Recovered episodes additionally require task completion. Realization controls use the same support proposer as ReCommit at this budget.

Proposal backbone	Recovered episodes $\uparrow$					Rate	Time
	Box	Calendar	Linear	Slack	Total	(%) $\uparrow$	(s) $\downarrow$
LLaDA-8B	23	9	32	75	139	22.03	33.60
Dream-7B ( Ye et al., 2025 )	19	6	34	72	131	20.76	30.67
LLaDA-1.5-8B ( Zhu et al., 2026b )	24	9	33	74	140	22.19	30.52
SDAR-8B ( Cheng et al., 2026 )	19	5	33	69	126	19.97	31.93
DiffuCoder-7B ( Gong et al., 2026 )	15	9	34	66	124	19.65	34.99

Appendix

Table 8: Diffusion-backbone comparison at

B=3

with a fixed realizer and realization schedule. Service columns and Total count recovered episodes; Rate is episode-weighted recovery (%), and Time is mean full-budget repair time (s). The default is shaded; column-best values are bolded.

$B$	Shared proposal	Downstream work	Total
1	0.113	8.606	8.718
3	0.113	33.486	33.599
7	0.113	83.279	83.392
13	0.113	158.556	158.669

Appendix

Table 9: ReCommit’s mean full-budget repair time (s), split into shared operation scoring and support ranking, and downstream realization, execution, and pruning replay. Each budget includes the full proposal cost once for the complete repair search.

Method	First-success time by service				Aggregate
	Box	Calendar	Linear	Slack	Recovered	Full	First
Direct (32B)	89.3	138.6	86.0	91.3	98	108.0	103.1
Schema (32B)	102.1	213.1	118.8	77.3	143	143.1	131.4
Feedback-Reflective (32B)	237.0	516.1	332.6	160.8	163	350.4	320.2
Direct	38.1	61.5	28.5	30.1	35	40.8	40.3
Schema	39.3	68.2	33.1	30.2	66	44.8	43.6

Appendix

Table 10: Offline first-success repair time at

B=3

. Service columns report mean time (s). Aggregate columns give recovered episodes and mean full-budget (Full) and offline first-success (First) times (s). All means include unrecovered episodes at their full-budget cost. Comparison methods use Qwen3-8B unless marked 32B; both ReCommit variants use Qwen3-8B for realization.

Method	First-success time by service				Aggregate
	Box	Calendar	Linear	Slack	Recovered	Full	First
Direct (32B)	373.9	625.1	380.6	353.7	118	487.1	441.9
Schema (32B)	398.0	928.9	519.9	300.5	167	649.8	553.9
Feedback-Reflective (32B)	729.2	2154.3	1446.5	578.2	204	1528.3	1273.3
Direct	168.9	264.2	124.9	124.8	47	177.7	173.6
Schema	181.0	293.3	141.2	121.0	76	197.0	187.4

Appendix

Table 11: Offline first-success repair time at

B=13

Method	$B=1$	$B=3$	$B=5$	$B=7$	$B=9$	$B=11$	$B=13$
Direct (32B)	7 32.9	12 92.7	15 150.2	18 218.9	20 282.7	21 346.5	22 408.0
Schema (32B)	20 40.5	30 113.5	36 192.0	38 274.7	39 355.9	41 439.2	42 520.5
Feedback-Reflective (32B)	23 92.5	41 277.7	50 431.3	55 602.9	55 763.9	56 924.0	56 1084.8
Direct	4 13.1	4 38.3	5 64.2	6 90.6	6 119.9	6 146.8	8 171.6
Schema	9 14.1	10 40.3	10 69.6	10 98.6	10 128.1	11 157.4	13 188.7
Reflective	6 10.4	8 33.7	9 57.1	10 80.5	11 103.6	11 126.2	11 148.3

Appendix

Table 12: Recovery–cost results across budgets on Box (129 episodes). Cells give recovered episodes (top) and mean full-budget repair time in seconds (bottom). Comparison methods use Qwen3-8B unless marked 32B; both shaded ReCommit variants use Qwen3-8B for realization.

Method	$B=1$	$B=3$	$B=5$	$B=7$	$B=9$	$B=11$	$B=13$
Direct (32B)	9 53.6	12 141.0	12 236.7	13 345.5	14 447.0	14 549.6	14 650.3
Schema (32B)	20 86.5	23 226.6	26 384.9	29 553.8	31 712.4	31 880.4	31 1048.3
Feedback-Reflective (32B)	17 186.1	20 545.0	21 881.3	24 1271.1	25 1643.4	29 2022.7	29 2398.2
Direct	2 20.6	2 61.7	2 102.3	2 143.3	2 183.7	2 224.4	2 265.0
Schema	7 23.2	7 69.3	8 115.8	9 161.5	9 208.2	9 254.9	9 301.4
Reflective	6 22.0	7 66.8	8 112.5	10 156.0	10 201.1	10 246.3	10 290.9

Appendix

Table 13: Recovery–cost results across budgets on Calendar (180 episodes). Cells give recovered episodes (top) and mean full-budget repair time in seconds (bottom). Comparison methods use Qwen3-8B unless marked 32B; both shaded ReCommit variants use Qwen3-8B for realization.

Method	$B=1$	$B=3$	$B=5$	$B=7$	$B=9$	$B=11$	$B=13$
Direct (32B)	10 31.0	17 88.4	17 146.6	17 210.6	17 272.8	17 336.0	17 399.2
Schema (32B)	22 44.6	25 124.8	25 209.9	25 301.0	26 387.3	27 476.5	27 563.6
Feedback-Reflective (32B)	30 121.5	32 355.9	34 594.5	34 851.3	34 1105.3	34 1358.6	34 1610.2
Direct	2 9.7	3 28.7	3 48.5	3 67.3	3 86.7	3 106.4	3 125.8
Schema	17 11.6	18 34.4	18 57.3	18 80.6	18 103.4	18 126.9	18 149.2
Reflective	20 11.1	20 32.6	20 54.2	20 75.8	20 97.1	20 119.0	20 140.3

Appendix

Table 14: Recovery–cost results across budgets on Linear (156 episodes). Cells give recovered episodes (top) and mean full-budget repair time in seconds (bottom). Comparison methods use Qwen3-8B unless marked 32B; both shaded ReCommit variants use Qwen3-8B for realization.

Method	$B=1$	$B=3$	$B=5$	$B=7$	$B=9$	$B=11$	$B=13$
Direct (32B)	46 37.4	57 102.7	61 170.4	62 245.4	64 313.2	64 385.8	65 454.3
Schema (32B)	56 34.1	65 92.7	66 153.0	66 218.2	66 278.4	66 338.5	67 399.3
Feedback-Reflective (32B)	55 67.6	70 190.9	76 318.1	77 456.6	81 588.6	84 723.5	85 852.7
Direct	24 10.5	26 31.5	30 53.1	32 74.0	33 94.7	33 115.7	34 136.7
Schema	27 11.7	31 31.8	33 51.9	34 72.8	34 93.8	35 114.5	36 135.1
Reflective	21 8.8	27 27.1	28 44.9	28 62.3	28 78.6	29 95.6	29 112.8

Appendix

Table 15: Recovery–cost results across budgets on Slack (166 episodes). Cells give recovered episodes (top) and mean full-budget repair time in seconds (bottom). Comparison methods use Qwen3-8B unless marked 32B; both shaded ReCommit variants use Qwen3-8B for realization.

Service	Episodes	Recovered episodes $\uparrow$		Exclusive recoveries
		AR-proposer control	ReCommit	ReCommit only	AR-proposer control only
Box	129	16	23	9	2
Calendar	180	7	9	2	0
Linear	156	31	32	3	2
Slack	166	53	75	23	1
Total	631	107	139	37	5

Appendix

Table 16: ReCommit and the AR-proposer control at

B=3

. The realizer, public-error pruning, and canonical/expressive schedule are fixed. Each proposer retains its own support-ordering rule. The last two columns count episodes recovered by one configuration but missed by the other. Higher recovery counts within each row are bolded .

$M$	Recovered episodes by service				Total recovered		Time
	Box	Calendar	Linear	Slack	With-P	Without-P	(s)
4	11	5	29	67	112	110	25.99
8	23	9	32	75	139	128	33.60
12	22	9	29	71	131	92	55.20
16	18	12	26	65	121	71	58.37
20	16	17	27	65	125	62	63.44

Appendix

Table 17: Effect of the masked-slot count

M

on recovery and cost at

B=3

. Service columns give recovered episodes with public-error pruning; With-P and Without-P give totals with and without pruning. Time is mean full-budget repair time with pruning (s). The shared default

M=8

is shaded; the best aggregate recoveries are bolded.

Explore similar work

Jun 4, 2026cs.AI

When Tools Fail: Benchmarking Dynamic Replanning and Anomaly Recovery in LLM Agents

Existing benchmarks evaluate Tool-Integrated Reasoning (TIR) in LLMs on idealized ''happy paths'', largely overlooking real-world tool failures. We introduce ToolMaze, a benchmark for dynamic path discovery and error recovery in TIR agents. To separate systematic replanning from blind trial-and-error, ToolMaze adopts a two-dimensional design: DAG-based topological complexity and a

2 \times 2

taxonomy of tool perturbations (explicit/implicit, transient/permanent). Evaluations show that perturbations degrade performance across nearly all models, with the sharpest drops under implicit semantic failures. Driven by systemic over-trust in corrupted outputs, Perturbation Recovery Rate (PRR) plummets by around 37% in these scenarios, while complex topologies trap agents in futile trial-and-error loops. Crucially, agentic fault-tolerance improves with model scale

3.66\times

slower than basic task execution, highlighting dynamic replanning as a distinct bottleneck unaddressed by model scaling or prompting. Data and code are available at https://github.com/Zhudongsheng75/ToolMaze.

Dongsheng Zhu, Xuchen Ma, Yucheng Shen +5

Shanghai AI Laboratory · East China Normal University · Soochow University +2

Large Language Model Agents Fault Tolerance Iterative Re-Planning

Ask PDF

Sep 14, 2026cs.LG

ParaRecover: A Process-Level Benchmark for Error Localization and Recovery in Parallel Tool-Use Agents

Existing agent benchmarks mainly evaluate final task success or tool-call correctness, providing limited insight into whether agents can reliably diagnose and recover from intermediate execution failures. This limitation becomes particularly critical in multi-turn parallel tool-use scenarios, where errors may propagate across dependent branches and trigger cascading failures. We introduce ParaRecover, a process-level benchmark for evaluating error localization and recovery in multi-turn parallel tool-use agents. Built upon a fine-grained taxonomy of 14 error types covering planning dependencies, tool selection, and argument matching, the benchmark comprises 10,626 instances spanning two difficulty levels. To enable finegrained, process-oriented evaluation, we further propose the SDE rubric, which measures structural integrity, diagnostic reasoning, and evolutionary strategy during agent execution.Experiments across more than ten mainstream LLMs reveal that even state-of-the-art models still struggle with multi-turn error propagation,implicit tool-use failures, and precise replanning. Moreover, we demonstrate that the SDE rubric provides effective supervision signals for improving agents' reflective recovery capabilities. Our data and code are available at https://github.com/gbw206/ParaRecover.

Bowen Guan, Zhentao Yin, Yanming Shen

School of Computer Science and Technology, Dalian University of Technology, China

Localization

Ask PDF

Jul 21, 2026cs.AI

PhoenixRepair: Rethinking Repair Strategy Exploration in Software Agents

While Large Language Models have greatly advanced automated issue resolution, existing agent-based methods exhibit a fundamental limitation in their insufficient exploration of repair strategies. This insufficiency manifests in two key aspects. First, the exploration of multiple potential edit locations is limited. Second, the exploration of repair attempts at each location is also insufficient. To address these challenges, we present PhoenixRepair, a multi-agent framework that systematically explores multiple candidate edit locations and performs iterative reflection and refinement on patch generation, thereby expanding the search space of repair strategies. Our framework begins with multi-location sampling, optionally augmented with graph-based localization information for difficult tasks, followed by iterative reflection and refinement to generate better patches, culminating in final-round generation guided by distilled insights from all historical attempts. Experiments on SWE-bench-Verified demonstrate that PhoenixRepair achieves the largest relative improvement of 7.8% over SWE-agent under DeepSeek-V3.1, and attains the highest resolved rate of 76.0% Pass@1 under MiniMax-M2.5. Meanwhile, it achieves higher fault localization accuracy than existing approaches. Our code is available at https://github.com/DeepSoftwareAnalytics/PhoenixRepair.

Tianyue Jiang, Yanlin Wang, Xin He +7

Sun Yat-sen University, China · State Key Laboratory of Blockchain and Data Security, Zhejiang University, China · School of Cyber Science and Engineering, Huazhong University of Science and Technology, China +1

Automated Program Repair Repair Bug Detection Fault Diagnosis Efficient Exploration Precise Editing

Ask PDF