ContractRL: Shielded Group-Relative Policy Optimization for Auditable Tool-Call Repair

Miaobo Hu; Shuhao Hu; Xiaobo Guo; Xin Wang; Bokun Wang; Yina Sa; Daren Zha

cs.AISep 29, 2026

ContractRL: Shielded Group-Relative Policy Optimization for Auditable Tool-Call Repair

Authors: Miaobo Hu, Shuhao Hu, Xiaobo Guo, Xin Wang, Bokun Wang, Yina Sa, Daren Zha

Organizations: School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing, China · Institute of Information Engineering, Chinese Academy of Sciences, Beijing, China

Abstract

Structured tool calls often fail after only a small number of fields violate a schema or an execution contract. Regenerating the complete object enlarges the action surface and makes repeated repair difficult to audit. We introduce ContractRL, a contract-constrained sequential repair protocol that models verifier-guided JSON repair as a bounded decision process. At each step the policy observes the candidate, typed verifier feedback, JSON Pointer, immutable repair history, and remaining budget; a contract-derived action mask filters malformed or prohibited RFC-6902 operations before a deterministic validator performs the transition. We specify a contract-constrained group-relative objective for patch, retry, and abstention decisions while keeping canonical targets and semantic labels outside the online state until trace freeze. Under identical verifier information, ContractRL attains 0.9362 semantic success with 34.4 generated tokens, compared with 0.9076 and 44.9 tokens for Patch-SFT and 0.9148 and 137.2 tokens for full regeneration over 192 cases per seed and five seeds. Policy optimization improves semantic success from 0.9186 for supervised ContractRL to 0.9375. A separate three-seed paired evaluation against Patch-SFT yields a semantic difference of +0.0396 (95% CI $[+0.0137,+0.0662], p=0.0039$ ). Feedback, action-mask, budget, and schema-shift analyses connect these gains to localized correction, while adversarial and multi-turn evaluations characterize the remaining failure modes.

Figures & tables

Figure 1: ContractRL inference and audit protocol. Blue panels expose the corrupted call and deterministic verifier; the purple loop proposes and revalidates bounded repairs. The ledger freezes before the orange audit accesses canonical targets. Solid arrows carry public state or control flow; the dashed boundary separates online repair from offline scoring.

Evidence unit	Cases / seeds	Schema valid $\uparrow$	Object match $\uparrow$	Attempts
BFCL mechanism replay	64	1.0000	1.0000	1.00
2B rank-8, validation block	64/4	0.9727	0.9141	1.00
2B rank-8, context-only	64/4	0.0820	0.0742	1.00
0.8B rank-4, validation block	64/8	0.9473	0.7441	1.00
0.8B rank-4, context-only	64/8	0.0000	0.0000	1.00

Table 1: Primary structural evidence. Adapter rates are per-adapter means on the fixed 64-case held-out file; the BFCL replay is a validator/ledger diagnostic whose supplied precomputed targets yield the displayed final-object match.

Arm	Output representation	Semantic success $\uparrow$	Output tokens $\downarrow$	p95 latency (s) $\downarrow$	Collateral edits $\downarrow$
Full regeneration	complete object	0.9148	137.2	2.54	0.1876
Constrained decoding	Constrained object	0.8879	146.1	2.80	0.1743
Patch-SFT	Bounded patch	0.9076	44.9	1.49	0.0488
ContractRL	Masked patch / abstain	0.9362	34.4	1.30	0.0169

Table 2: Same-information output-representation controls over 192 cases per seed and five seeds. Every arm receives the candidate, schema, typed error code, JSON Pointer, and public repair history; only the output representation and decoding rule change. Best values are bolded within each metric direction.

Figure 2: Semantic success and generation cost with identical public verifier information. All arms use 192 cases per seed and five seeds. Bars show the reported aggregates in Table 2 ; the output-token axis starts at zero.

Arm	Split	Seeds	Schema valid $\uparrow$	Semantic success $\uparrow$	Mean retries $\downarrow$	Output tokens $\downarrow$	Mean latency (s) $\downarrow$	Collateral edits $\downarrow$
ContractRL	held-out	3	0.9768	0.9314	0.0547	34.6	0.83	0.0189
full regeneration	held-out	3	0.9392	0.9027	0.0000	138.9	1.76	0.2214
constrained decoding	held-out	3	0.9901	0.8663	0.0000	147.8	2.11	0.1986
Patch-SFT	held-out	3	0.9446	0.8918	0.1176	46.7	0.97	0.0618
ContractRL	nested-array	3	0.9187	0.8516	0.1239	42.9	1.04	0.0297

Table 3: Matched semantic and efficiency evaluation over three seeds. Held-out rows use 64 cases per seed; the nested/array row is a separate case-family cohort. Rates are case means, while retries, generated tokens, mean latency, and collateral edits use the frozen resource ledger; p95 latency is reported separately in the efficiency table.

Appendix figures & tables42 assets

Supplementary material from the paper’s appendix.

Appendix

Stage	Recorded operation	Target visible?
Observe	Read the corrupted call and typed verifier code plus the reported JSON Pointer.	no
Patch	Emit a bounded JSON Patch operation list and apply it to the current object.	no
Validate	Run schema and cross-field checks; return typed errors on failure.	no
Retry	Spend the next of at most two patch attempts, preserving prior patches and validator responses.	no
Freeze	Hash the ordered patch ledger, attempt count, additional-retry count, and collateral fields.	no
Score	Join the canonical patch and object and compute their distinct equalities.	post-hoc

Appendix

Table 4: Execution contract for one repair episode. The target patch and target object are held out from the online loop; validator feedback is the only retry signal.

Evaluation	Population	Outcome
Structural replay	64 cases	64/64 final validity with one precomputed operation per case; 32 add and 32 replace operations.
Adapter evaluation	64 cases; 4 or 8 seeds	2B: 0.9727 schema validity and 0.9141 object match. 0.8B: 0.9473 and 0.7441.
Context-only control	Same adapter cases	2B: 0.0820 validity and 0.0742 object match. 0.8B: both 0.0000.
Transaction audit	9 cases; 8 repairable	7/9 final validity and 6/8 accepted repairs, with rollback, wrong-path, and immutable-field checks.
Case-matched policies	3 seeds	ContractRL: 0.9314 semantic success, 34.6 output tokens, and 1.30 s p95 latency.
Public-information control	192 cases per seed; 5 seeds	ContractRL: 0.9362 semantic success versus 0.9076 for Patch-SFT and 0.9148 for full regeneration.

Appendix

Table 5: Evaluation units and principal outcomes. Each population retains its own case and seed denominator.

Field	Stage	Audit rule
case identifier	split	unique within a manifest and disjoint across train/evaluation
public evidence	online	candidate, schema, pointer, and typed feedback only
patch ledger	repair	ordered operations, validator status, and retry count are append-only
semantic trace	execution	dispatched accepted object, tool outcome, and side-effect record
resource ledger	aggregation	I/O tokens, calls, retries, latency, and seed
result join	scoring	canonical object and task label are joined after trace freeze

Appendix

Table 6: Field-level interpretation of the evidence ledger.

Figure 3: Adapter structural results on 64 held-out cases. Bars are seed means: four seeds for 2B and eight for 0.8B. The validation-block and context-only conditions use the same model-specific case sets.

Figure 4: Structural diagnostics with explicit denominators. The deterministic replay uses 64 cases; the independent audit uses nine cases, of which eight require repair. The right panel records rollback and rejection checks in the independent audit.

Figure 5: Eight-case complete-output diagnostic. Labels show the numerator and denominator for each endpoint; token-cap hits are failures and are plotted separately from successful-output rates.

Arm	Input	Output budget	Online feedback	Recorded cost
full regeneration	corrupted call + schema	complete object	none	I/O tokens, latency
constrained decoding	corrupted call + grammar	complete object	none	I/O tokens, latency
Patch-SFT	corrupted call + typed error	patch, max two ops	validator report	I/O, retries, latency
ContractRL	same public block	patch, max two ops	typed ledger	all above plus collateral edits

Appendix

Table 7: Comparator accounting contract.

Check	Training	Held-out	Result
case ID overlap	426 IDs	64 IDs	0 exact overlaps
candidate fingerprint	426 hashes	64 hashes	0 exact overlaps
schema hash	147 hashes	26 hashes	0 exact overlaps
pointer/error template	392 templates	61 templates	0 exact overlaps
tool-name disjointness	118 names	23 names	0 overlapping names

Appendix

Table 8: Split and case-integrity checks.

Scenario	Arm	Schema valid $\uparrow$	Semantic success $\uparrow$	Mean retries $\downarrow$	Latency (s) $\downarrow$
single-field	all	0.9674	0.9128	0.0509	1.81
multi-field	all	0.8497	0.7892	0.1698	2.13
nested array	all	0.8985	0.8307	0.1376	2.05
wrong pointer	patch arms	0.8172	0.7449	0.2264	1.64
malformed feedback	all	0.7846	0.7168	0.2471	1.96
prompt injection	all	0.9453	0.8989	0.0714	1.90

Appendix

Table 9: Semantic and adversarial results. Latency is measured in seconds for each scenario.

Arm	Mean retries $\downarrow$	p95 retries $\downarrow$	Output tokens $\downarrow$	p95 latency (s) $\downarrow$	Semantic success $\uparrow$
ContractRL	0.0547	1	34.6	1.30	0.9314
full regeneration	0.0000	0	138.9	2.58	0.9027
constrained decoding	0.0000	0	147.8	2.85	0.8663
Patch-SFT	0.1176	1	46.7	1.55	0.8918

Appendix

Table 10: Retry and efficiency results under the matched three-seed ledger. Latency is reported at p95; mean latency appears in Table 3 .

Arm	Case family	First valid $\uparrow$	Final valid $\uparrow$	Mean retries $\downarrow$	Pointer fix $\uparrow$	Collateral edits $\downarrow$	Semantic success $\uparrow$
ContractRL	missing-required	0.9815	0.9952	0.0185	0.9937	0.0054	0.9598
ContractRL	wrong-type	0.9618	0.9861	0.0382	0.9804	0.0137	0.9406
Patch-SFT	mixed	0.8897	0.9446	0.1103	0.9198	0.0618	0.8918
full regeneration	mixed	0.9392	0.9392	0.0000	n/a	0.2214	0.9027

Appendix

Table 11: Typed-feedback transition outcomes. Rows retain their recorded case family; mixed rows are matched within that cohort, and rates are case means from the frozen ledger.

Arm	Case family	Seeds	Prompt tokens $\downarrow$	Output tokens $\downarrow$	Validator calls $\downarrow$	p95 latency (s) $\downarrow$	Semantic success $\uparrow$
ContractRL	single-field	3	1111	31.8	1.028	1.13	0.9502
ContractRL	nested/array	3	1135	42.9	1.124	1.52	0.8516
constrained decoding	mixed	3	1079	147.8	1.000	2.85	0.8663
full regeneration	mixed	3	1072	138.9	n/a	2.58	0.9027

Appendix

Table 12: Resource-normalised comparison. Case family, model seed, token budget, and validator implementation are bound before scoring; rows are reported by cohort and are not pooled across families.

Arm	Case family	Cases	First valid $\uparrow$	Semantic success $\uparrow$	Mean retries $\downarrow$	Output tokens $\downarrow$	p95 latency (s) $\downarrow$	Collateral edits $\downarrow$
full regeneration	single-field	128	0.9487	0.9164	0.0000	136.2	2.44	0.2093
constrained decoding	single-field	128	0.9951	0.8826	n/a	143.7	2.67	0.1908
Patch-SFT	nested/array	192	0.8463	0.8097	0.1534	51.5	1.69	0.1027
ContractRL	single-field	128	0.9716	0.9502	0.0280	31.8	1.13	0.0096
ContractRL	ambiguous/multi-field	192	0.7639	0.7935	0.2361	48.6	1.84	0.0432

Appendix

Table 13: Structural and semantic boundary ablation. Rows retain their recorded case family; model seed units, output budgets, and the execution sandbox are fixed within each cohort. The measured cohorts contain 128 single-field, 192 nested/array, and 192 ambiguous/multi-field cases.

Field	Online arm	Scoring arm	Validation
candidate JSON	yes	input digest	schema hash
typed verifier code	yes	input digest	enum check
JSON Pointer	yes	pointer metric	RFC-6901 parse
schema version	yes	split key	version match
patch ledger	append-only	retry/cost	digest check
canonical patch	no	exact patch match	post-hoc join

Appendix

Table 14: ContractRL evidence contract.

Operation	Case family	Validator result	Retry cost	Validator test
add	missing-required	schema/cross-field	0.018	localized repair
replace	wrong-type	schema/cross-field	0.038	type correction
remove	multi-field	schema/cross-field	0.154	deletion boundary
move/copy	nested/array	schema/cross-field	0.124	pointer semantics
wrong path	adversarial	pointer failure	1.000	fail-closed handling
ambiguous report	noisy verifier	typed retry	1.000	feedback robustness

Appendix

Table 15: Patch vocabulary and transition outcomes.

Check	Unit	Read
case ID overlap	record	exact split separation
candidate fingerprint overlap	record	input separation
target fingerprint overlap	record	label separation
schema hash overlap	schema	schema coverage
target-path overlap	pointer	error separation
operation-sequence overlap	patch	patch separation

Appendix

Table 16: Split-integrity checks for the recorded repair protocol.

Arm	Case family	Schema valid $\uparrow$	Semantic success $\uparrow$	Side-effect error $\downarrow$	Mean retries $\downarrow$
ContractRL	single-field	0.9907	0.9502	0.0084	0.0280
ContractRL	nested/array	0.9187	0.8516	0.0274	0.1239
full regeneration	mixed	0.9392	0.9027	0.0451	n/a
constrained decoding	mixed	0.9901	0.8663	0.0396	n/a
Patch-SFT	mixed	0.9446	0.8918	0.0308	0.1176

Appendix

Table 17: Semantic execution results.

Case family	Validator condition	First valid $\uparrow$	Final valid $\uparrow$	Semantic success $\uparrow$
wrong path	reject or correct	0.6948	0.8421	0.7796
immutable pointer	reject	0.0000	0.0000	0.0000
ambiguous verifier	typed retry	0.6324	0.8587	0.8065
nested/array	local patch	0.8463	0.9187	0.8516
multi-field	bounded sequence	0.7639	0.8672	0.7935
malformed output	reject	0.0000	0.0000	0.0000

Appendix

Table 18: Adversarial failure taxonomy and measured metrics.

Arm	Prompt tokens $\downarrow$	Output tokens $\downarrow$	Calls	Mean retries $\downarrow$	p95 latency (s) $\downarrow$	Semantic success $\uparrow$
ContractRL	1114.2	34.6	1.055	0.0547	1.30	0.9314
full regeneration	1072.4	138.9	1.000	0	2.58	0.9027
constrained decoding	1079.1	147.8	1.000	0	2.85	0.8663
Patch-SFT	1112.7	46.7	1.118	0.1176	1.55	0.8918

Appendix

Table 19: Resource accounting for matched comparators. Latency is reported at p95.

Method	Update interface	Constraint mechanism	Evaluation focus
VeriHarness	Candidate retry with structured feedback	External acceptance and call budget	TextWorld terminal success
PatchBoard	JSON Patch over shared state	Schema, write contracts, and transactional kernel	Multi-agent ALFWorld episodes
Schema RL	Complete structured output	Schema reward in training	Schema-conformant generation
SLOT	Output post-processing	Structural conversion	Structured output recovery
XGrammar	Token-level generation	Context-free grammar	Coverage and generation overhead
ContractRL	Bounded patch, retry, or abstention	Action mask and deterministic validator	Semantic repair, collateral edits, and cost

Appendix

Table 20: Closest method interfaces and evaluation targets, based on the cited primary sources.

Arm	Output form	Decoding	Seeds	Audit field
ContractRL	JSON Patch	unconstrained	3	ledger and validator
full regeneration	complete object	unconstrained	3	token baseline
constrained decoding	complete object	grammar constrained	3	syntax control
Patch-SFT	JSON Patch	unconstrained	3	learned patch
semantic execution	accepted object	sandbox	3	task endpoint

Appendix

Table 21: ContractRL baseline configuration.

Case family	Pointer type	Operations	Collateral edits $\downarrow$	Semantic success $\uparrow$	Read
single-field	exact	1	0	0.9502	mechanism screen
single-field	wrong path	1	0.0000	0.7796	pointer boundary
nested/array	exact	1.24	0.0297	0.8516	structure generalization
multi-field	mixed	1.71	0.0432	0.7935	operation budget

Appendix

Table 22: Pointer and collateral-edit metrics.

Arm	First valid $\uparrow$	Final valid $\uparrow$	Mean retries $\downarrow$	p95 retries $\downarrow$	Observed behaviour
ContractRL	0.9453	0.9768	0.0547	1	typed local correction
Patch-SFT	0.8897	0.9446	0.1176	1	learned correction
full regeneration	0.9392	0.9392	0	0	regen.
constrained decoding	0.9901	0.9901	0	0	syntax control

Appendix

Table 23: Retry-policy results.

Field	Denominator	Source	Read
schema validity	all cases	validator	structural gate
object match	held-out cases	hidden object	canonical-object metric
semantic	executed cases	sandbox result	task endpoint
tool error	executed cases	sandbox trace	failure boundary
collateral edit	accepted patches	pointer diff	locality metric
latency/cost	all attempts	resource ledger	efficiency endpoint

Appendix

Table 24: Semantic scoring fields.

Arm	Model rank	Schema valid $\uparrow$	Object match $\uparrow$	Semantic success $\uparrow$	Statistical uncertainty
ContractRL	0.8B/r4	0.9508	0.7523	0.8124	0.0028
ContractRL	2B/r8	0.9765	0.9188	0.9331	0.0012
ContractRL	7B/r16	0.9831	0.9329	0.9446	0.0009
Patch-SFT	2B/r8	0.9451	0.8679	0.8924	0.0018
full regeneration	2B	0.9403	0.8896	0.9041	0.0014
constrained decoding	2B	0.9905	0.8426	0.8671	0.0016

Appendix

Table 25: Model-size comparison across five recorded seeds and 192 cases per configuration. ContractRL rows use the same fixed cases and public evidence; reference arms retain their recorded information conditions. Statistical uncertainty accompanies the success estimates. Best values are bolded within each metric direction.

Arm	Model rank	Mean retries $\downarrow$	Output tokens $\downarrow$	p95 latency (s) $\downarrow$
ContractRL	0.8B/r4	0.0831	33.1	1.22
ContractRL	2B/r8	0.0538	34.8	1.32
ContractRL	7B/r16	0.0415	35.6	1.68
Patch-SFT	2B/r8	0.1162	46.9	1.56
full regeneration	2B	n/a	139.4	2.60
constrained decoding	2B	n/a	148.1	2.86

Appendix

Table 26: Resource metrics for the model-size comparison: 192 cases per seed and five seeds. Best values are bold within the ContractRL capacity sweep and within the 2B comparator set.

Panel of Figure 6 — Figure 6: Five-seed ContractRL scaling over 192 cases per seed. Point labels report generated tokens; points show aggregate semantic success.

Feedback condition	Public signal	Schema valid $\uparrow$	Object match $\uparrow$	Semantic success $\uparrow$	Mean retries $\downarrow$	Collateral edits $\downarrow$	Fail-closed rate $\uparrow$
full typed block	code + pointer + schema context	0.9766	0.9189	0.9317	0.0539	0.0191	0.9728
error code only	code	0.9018	0.8027	0.8236	0.1442	0.0476	0.9147
pointer only	JSON Pointer	0.8694	0.7462	0.7815	0.1817	0.0589	0.8793
natural-language text	free-form message	0.8847	0.7724	0.8001	0.1668	0.0527	0.8872
context only	candidate + schema	0.0831	0.0749	0.0678	0.4383	0.1762	0.6146

Appendix

Table 27: Verifier-feedback ablation over 192 cases per seed and five seeds. The rows remove one typed field or replace the block with a weaker representation. Best values are bolded within each metric direction.

Training arm	Model/rank	Semantic success $\uparrow$	Mean retries $\downarrow$	Output tokens $\downarrow$	Collateral edits $\downarrow$	p95 latency (s) $\downarrow$
Patch-SFT	2B/r8	0.8927	0.1159	46.6	0.0607	1.55
ContractRL-SFT	2B/r8	0.9186	0.0762	39.3	0.0314	1.42
ContractRL policy optimization	2B/r8	0.9375	0.0496	34.3	0.0168	1.30

Appendix

Table 28: Supervised initialization versus policy optimization over 192 cases per seed and five seeds. The comparison fixes the public state, case split, model family, validator, and episode budgets, then changes the training objective. Best values are bolded within each metric direction.

Policy variant	Semantic success $\uparrow$	Mean retries $\downarrow$	Abstain rate	Collateral edits $\downarrow$	Unsafe dispatches $\downarrow$	Output tokens $\downarrow$	p95 latency (s) $\downarrow$
full ContractRL objective	0.9375	0.0496	0.0718	0.0168	0	34.3	1.30
w/o contract action mask	0.8971	0.1098	0.0531	0.0649	26	42.7	1.59
w/o repair history	0.9162	0.0847	0.0695	0.0319	4	38.6	1.43
w/o learned retry/abstention	0.9064	0.0000	0.0000	0.0396	7	31.5	1.18
w/o collateral objective	0.9328	0.0524	0.0642	0.0476	0	35.8	1.34
structure-only objective	0.8723	0.0661	0.0249	0.0874	12	40.4	1.50

Appendix

Table 29: Component-wise policy ablation over 192 cases per seed and five seeds. Each row removes one decision or objective term while preserving the same public state, case IDs, validator, and scoring join. Best values are bolded within each metric direction.

Max. ops.	Max. attempts	First valid $\uparrow$	Final valid $\uparrow$	Semantic success $\uparrow$	Mean retries $\downarrow$	Output tokens $\downarrow$	Collateral edits $\downarrow$
1	1	0.9126	0.9126	0.8794	0.0000	28.9	0.0098
1	2	0.9126	0.9497	0.9078	0.0758	32.2	0.0127
2	2	0.9453	0.9768	0.9314	0.0547	34.6	0.0189
3	2	0.9498	0.9812	0.9348	0.0521	39.1	0.0338
4	3	0.9514	0.9851	0.9362	0.0648	47.3	0.0571

Appendix

Table 30: Patch-budget and operation-count ablation. Each row uses the same corruption cases; the reported budget is read from the joint semantic and collateral frontier.

Figure 7: Patch-budget trade-off from Table 30 . Settings 1–5 are respectively (operations, attempts) = (1,1), (1,2), (2,2), (3,2), and (4,3). The teal bar marks the two-operation, two-attempt contract. Increasing the budget improves semantic success while increasing output tokens and collateral edits.

Structure family	Cases	Schema valid $\uparrow$	Object match $\uparrow$	Semantic success $\uparrow$	Mean retries $\downarrow$	Collateral edits $\downarrow$	p95 latency (s) $\downarrow$
flat single-field	128	0.9908	0.9396	0.9506	0.0276	0.0094	1.15
nested object	96	0.9457	0.8584	0.8798	0.0968	0.0246	1.39
array index	96	0.9169	0.8152	0.8469	0.1327	0.0334	1.53
union / optional	96	0.9054	0.7996	0.8302	0.1451	0.0358	1.60
long pointer	96	0.8892	0.7811	0.8134	0.1635	0.0397	1.71
multi-field	96	0.8696	0.7487	0.7951	0.2289	0.0445	1.83

Appendix

Table 31: Schema-structure evaluation. The table separates locality and semantic degradation by corruption shape.

Feedback condition	Cases	First valid $\uparrow$	Final valid $\uparrow$	Semantic success $\uparrow$	Fail-closed rate $\uparrow$	Collateral edits $\downarrow$
clean typed report	96	0.9451	0.9770	0.9320	0.9715	0.0187
stale schema version	96	0.5813	0.7368	0.6685	0.9896	0.0079
wrong pointer	96	0.6969	0.8438	0.7792	0.9687	0.0206
ambiguous pointer	96	0.6347	0.8595	0.8058	0.9514	0.0278
malformed feedback	96	0.6116	0.7889	0.7182	0.9844	0.0104
injected tool text	96	0.9074	0.9516	0.9007	0.9911	0.0176

Appendix

Table 32: Verifier-noise and adversarial-feedback evaluation. Leakage is the fraction of cases in which hidden target information is exposed or copied into the public trace.

H	Trajectories	Completion	Semantic success $\uparrow$	Tool errors $\downarrow$	Mean retries $\downarrow$	Cumulative tokens $\downarrow$	Cumulative latency (s) $\downarrow$	Collateral edits $\downarrow$
1	192	0.9761	0.9321	0.0187	0.0548	34.8	0.84	0.0188
3	192	0.9256	0.8293	0.0518	0.1619	104.7	2.55	0.0426
5	192	0.8794	0.7548	0.0796	0.2735	175.2	4.31	0.0591
8	192	0.8073	0.6457	0.1214	0.4438	281.8	7.05	0.0807

Appendix

Table 33: Long-horizon and multi-turn tool trajectory evaluation. A trajectory is successful only if all accepted steps satisfy the semantic invariant.

Held-out family	Cases	Schema valid $\uparrow$	Object match $\uparrow$	Semantic success $\uparrow$	Pointer fix $\uparrow$	Collateral edits $\downarrow$
file	96	0.9621	0.8876	0.9024	0.9513	0.0216
calendar	96	0.9564	0.8748	0.8917	0.9436	0.0241
database	96	0.9312	0.8359	0.8546	0.9127	0.0368
HTTP	96	0.9487	0.8613	0.8795	0.9294	0.0285
code-edit	96	0.9258	0.8216	0.8429	0.9018	0.0417

Appendix

Table 34: Cross-tool-family transfer evaluation. Train and held-out family columns are recorded explicitly so that transfer is not confused with interpolation. Each row trains on the other four families among file, calendar, database, HTTP, and code-edit.

Arm / threshold	Coverage	Semantic risk $\downarrow$	Fail-closed rate $\uparrow$	ECE	AUROC	Brier
ContractRL / 0.50	0.9482	0.0719	0.7821	0.0458	0.9187	0.0794
ContractRL / 0.70	0.8156	0.0456	0.8816	0.0371	0.9281	0.0651
ContractRL / 0.80	0.6693	0.0298	0.9349	0.0316	0.9364	0.0557
ContractRL / 0.90	0.4047	0.0174	0.9782	0.0249	0.9449	0.0436
full regeneration / matched	1.0000	0.0973	0.0000	0.0941	0.8368	0.1297

Appendix

Table 35: Calibration and selective-abstention evaluation. Thresholds are fixed in advance; semantic risk is measured only on the covered subset.

Panel of Figure 8 — Figure 8: Coverage–risk frontier for selective repair. Each point corresponds to a confidence threshold fixed before held-out scoring.

Metric	ContractRL	Patch-SFT	$\Delta$	95% CI	$p$ -value	Effect size
schema validity	0.9768	0.9446	$+0.0322$	$[+0.0121,+0.0525]$	0.0027	0.31
semantic success	0.9314	0.8918	$+0.0396$	$[+0.0137,+0.0662]$	0.0039	0.35
mean retries	0.0547	0.1176	$-0.0629$	$[-0.0932,-0.0315]$	0.0018	-0.42
generated tokens	34.6	46.7	$-12.1$	$[-15.8,-8.5]$	$<0.0001$	-0.91
collateral edits	0.0189	0.0618	$-0.0429$	$[-0.0596,-0.0271]$	$<0.0001$	-0.58
p95 latency	1.30	1.55	$-0.25$	$[-0.36,-0.14]$	0.0003	-0.63

Appendix

Table 36: Paired uncertainty and effect-size report. Each interval is computed from the same case IDs for ContractRL and its matched baseline.

Arm	Audited cases	Agreement	False accept $\downarrow$	False reject $\downarrow$	Abstain agreement $\uparrow$	Median review time
ContractRL	120	0.9417	0.0167	0.0416	0.9250	18.6
Patch-SFT	120	0.8917	0.0417	0.0666	0.8583	22.9
full regeneration	120	0.8583	0.0750	0.0667	0.7833	31.7
constrained decoding	120	0.8750	0.0583	0.0667	0.8167	29.6

Appendix

Table 37: Blinded human–validator agreement audit. The audit separates agreement, false acceptance, false rejection, and review time for ContractRL and matched controls.

Case family	ID	Predicted patch	Validator result	Semantic outcome	Human label
single-field	SF-03	add /timezone	accept	success	safe
nested / array	NA-11	replace /items/2/id	accept	failure	unsafe
wrong pointer	WP-04	replace /user/id	reject: wrong-path	not executed	abstain
malformed feedback	MF-06	replace /date	reject: malformed	not executed	abstain
multi-field	MU-09	add /city; replace /date	accept	failure	unsafe
immutable / unsafe	IU-02	replace /account/id	reject: immutable	not executed	unsafe

Appendix

Table 38: Case-level failure audit. The sample is stratified across repair and failure families and retains representative successes and failures.

Measurement family	Evaluation unit	Result
Structural repair	64-case replay and adapter seeds	Table 1
Public-information control	192 cases per seed; five seeds	Table 2
Policy training	Fixed public state; five seeds	Table 28
Components	Contract decisions and costs	Table 29
Paired comparison	Shared cases and seeds	Table 36
Independent audit	120 blinded cases per arm	Table 37

Appendix

Table 39: Measurement index for the reproduction records.

LLM Self-Correction

Large language model self-correction studies how language models detect and repair errors in their own answers, reasoning, and actions.

See all 159 papers

One Readout, Many Repairs: Diffusion-Guided Hierarchical Search for Tool-Agent Repair
Sep 28, 2026Xiang Xia, Cheng Yan, Wuyang Zhang +4LLM Agent Reliability
University of Science and Technology of China, Hefei, China
Contract2Tool: Learning Preconditions and Effects for Reliable Tool-Augmented LLM Agents
Jun 5, 2026Rahul Suresh Babu, Laxmipriya Ganesh IyerAI Agent Reliability Tool-Use Evaluation
Independent Researcher United States of America
RubricRefine: Improving Tool-Use Agent Reliability with Training-Free Pre-Execution Refinement
May 10, 2026Will LeVine, Brendan Evers, Sam Saltwick +1LLM Self-Refinement Rubric-Based Evaluation
Anduril Industries

Sep 28, 2026J/K move · Enter open · S save

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

Xiang Xia, Cheng Yan, Wuyang Zhang, Fan Xu+3

University of Science and Technology of China, Hefei, China

LLM Agent Reliability

Open paper AskPDF arXiv

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.

Abstract

LLM Self-Correction

Explore similar work

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

Contract2Tool: Learning Preconditions and Effects for Reliable Tool-Augmented LLM Agents

RubricRefine: Improving Tool-Use Agent Reliability with Training-Free Pre-Execution Refinement