OntoAct: Hypothesis-Based Neuro-Symbolic Agents through Ontological Abstraction

Donghyeok Lee; Bong Gyun Kang; Junyong Ahn; Sungroh Yoon

cs.AIOct 15, 2026

OntoAct: Hypothesis-Based Neuro-Symbolic Agents through Ontological Abstraction

Authors: Donghyeok Lee, Bong Gyun Kang, Junyong Ahn, Sungroh Yoon

Organizations: Interdisciplinary Program in Artificial Intelligence, Seoul National University, Seoul, South Korea · Department of Electrical and Computer Engineering, Seoul National University, Seoul, South Korea

Abstract

Learning from experience requires recovering reusable knowledge and determining when it applies to a new situation. We introduce OntoAct, a hypothesis-based neuro-symbolic framework built on ontological abstraction. OntoAct turns individual experiences into axioms over entity types, relations, and conditions, making their underlying patterns reusable across situations. A large language model (LLM) uses this world knowledge to propose hypotheses and actions, while a symbolic verifier assesses their supporting conditions against current observations. Verification identifies missing evidence, and new observations and intervention outcomes update hypothesis assessments and guide subsequent actions. We also introduce Tacit-FLE, a benchmark built on the Factorio Learning Environment (FLE) for diagnosing and repairing hidden faults in existing factories using inherited operational records. Its synthetic maintenance corpus distributes evidence across records, and an executed repair may leave the fault unresolved, requiring agents to reassess their diagnoses through interaction. Across Tacit-FLE, Mars, and ALFWorld, OntoAct achieves leading task performance with efficient interaction and inference.

Figures & tables

Figure 1: The LLM proposes, the ontology verifies. (a) OntoAct abstracts recorded experience into world knowledge over typed entities, relations, and conditions. (b) Symbolic checks assess proposed hypotheses and actions against current evidence. (c) Feedback guides observation, action revision, and hypothesis assessment.

Figure 2: The OntoAct framework. (a) The LLM proposes axioms from a reference corpus; symbolic admission checks their conformance to the schema and applicable evidence constraints. (b) The LLM proposes claims, hypotheses, and actions using the resulting world knowledge; ontology-based verification supplies feedback for revision, while observations and action outcomes update the reasoning context.

		Repair success (%) $\uparrow$
Scale	Method	Chain	Conjunction	Elimination	All	SI $\downarrow$	Turns $\downarrow$
Small	ReAct	0.0	71.4	75.0	39.3	4.6	22.2
	ReAct *	15.4	85.7	37.5	39.3	7.2	22.9
	Reflexion	7.7	71.4	75.0	42.9	5.3	23.2
	IfR	38.5	100.0	75.0	64.3	6.4	17.4
	NeSyC	0.0	0.0	0.0	0.0	0.0	30.0

Table 1: Repair performance on Tacit-FLE. Success (%) by type and overall (All); SI: mean spurious interventions; Turns: mean turns. Failed episodes use the full budget (Appendix A.4.1 ). Bold: best; underlined: second-best overall success, interventions, and turns. Intervention and turn rankings require at least one repair.

	OntoAct	WALL-E	IfR
Distinct candidates/episode	3.36	5.14	9.00
Fault in candidate set	32/36	17/36	24/36
Fault in first 3	25/36	10/36	4/36

Table 2: Candidate count and fault coverage. Medium Tacit-FLE (36 tasks); targets include unexecuted proposals. Coverage counts tasks with the true fault among all candidates or the first three.

World	ReAct	ReAct *	Reflexion	IfR	NeSyC	WALL-E	OntoAct
Reward $\uparrow$
Default	$10.10\pm 4.80$	$9.30\pm 1.64$	$7.90\pm 2.95$	$8.10\pm 2.35$	$0.70\pm 0.55$	$6.30\pm 6.06$	$\boldsymbol{12.10\pm 1.22}$
Terrain	$8.70\pm 2.30$	$12.16\pm 4.03$	$8.68\pm 5.66$	$9.30\pm 3.96$	$-0.90\pm 0.00$	$7.50\pm 3.78$	$\boldsymbol{12.30\pm 1.92}$
Survival	$7.10\pm 4.00$	$9.50\pm 3.44$	$6.30\pm 5.07$	$7.30\pm 4.82$	$0.98\pm 0.99$	$6.30\pm 4.97$	$\boldsymbol{10.46\pm 6.39}$
Task dep.	$4.10\pm 1.00$	$2.90\pm 1.30$	$4.10\pm 2.92$	$7.30\pm 1.10$	$0.90\pm 0.45$	$3.90\pm 0.84$	$\boldsymbol{11.30\pm 1.10}$
Terr. surv.	$6.10\pm 2.35$	$\boldsymbol{8.10\pm 2.35}$	$5.10\pm 2.92$	$6.10\pm 4.30$	$2.70\pm 1.82$	$3.90\pm 3.11$	$6.30\pm 3.03$

Table 3: Mars Reward and Score by world. Luna; five evaluation episodes per world. Reward combines achievement rewards and health changes (mean

\pm

sample SD); Score geometrically aggregates achievement success rates. Average excludes Default; its Reward SD is computed across matched-seed world averages. Bold/underlined: best/second-best. Metric definitions are in Appendix E.3 .

Figure 5: Mars Score and evaluation tokens. Luna, seven modified worlds. Equal world weights; input/output tokens exclude knowledge construction.

Method	Luna	Luna $R$	Terra
ReAct	7.95	13.06	13.91
ReAct *	10.64	15.23	16.48
Reflexion	7.33	12.91	12.09
IfR	9.57	20.60	18.86
WALL-E	6.47	9.93	7.79
OntoAct (ours)	15.92	22.99	21.57

Table 4: Mars Score across model conditions. Mean Score (%) over seven modified worlds. Bold/underlined: best/second-best per condition.

	OntoAct	IfR
Initial candidate violations	70/245 (28.6%)	20/107 (18.7%)
Redirected before execution	70/70 (100%)	0/20 (0%)
Issued command violations	0/136 (0%)	20/107 (18.7%)

Table 5: Violations of shared learned conditions. Resource conditions correctly stored by both methods. Luna, seven modified worlds, 35 episodes per method. Fractions use each method’s eligible candidates or commands at the specified stage; Appendix D.4 details matching and denominators.

Method	SR (%)	Steps	Calls	Tokens (k)
NeSyC †	26.1	38.6	9.3	281.32
WALL-E	96.3	14.8	28.8	47.70
OntoAct (ours)	98.5	13.5	13.7	32.16

Table 6: ALFWorld success and evaluation cost. Luna. SR: success (%); mean environment Steps, LLM Calls, and Tokens (k) per episode. † Initial object locations and inventory provided.

Appendix figures & tables35 assets

Supplementary material from the paper’s appendix.

Appendix

Figure A1: Tacit-FLE. (a) An agent inherits an existing plant, a finite reference corpus of maintenance logs and operating documents, and observations of a new fault. (b) Chain links dependencies across incidents; Conjunction combines repair precedents, plans, and observations; Elimination rules out known explanations contradicted by current evidence.

Source	Available information
Initial task	Turn budget, normal production rate (final products/minute), unit kinds, and coordinates.
Inherited records	Local incident observations, checks, repairs, and outcomes; buffer policy for Chain; planned outputs for Conjunction.
Current observations	Machine states, storage totals and changes, shipping; inspections and itemized readings on request.
Repair feedback	Episode termination after an accepted repair; continued interaction otherwise.
Evaluator only	Injected fault, accepted repair targets, full original plant specification, and coverage annotations.

Appendix

Table A1: Information boundary. Public records can disclose selected planned facts; the complete original specification remains evaluator metadata.

Command	Game seconds	Effect or returned information
inspect $<$ unit $>$	5	External state; power and material-transfer device readings require this probe.
open $<$ unit $>$	10	Itemized storage or input contents; configured output in Conjunction.
wait $<$ seconds $>$	specified	Advance the plant and receive an updated status board.
replace $<$ unit $>$	15	Restore the addressed unit to its original specification.
refill $<$ storage $>$ $<$ item $>$ $<$ count $>$	10	Insert the selected item into storage; success also depends on the repair target and item.

Appendix

Table A2: Agent actions in the evaluated protocol. Each action uses one turn. The plant advances by the indicated game time; LLM deliberation itself is not charged as game time. The status board is included in observations at no additional turn cost.

Figure A3: Small-scale plant examples (a)–(b). Each plant has one target product. One input stream splits and reconverges in (a), while (b) combines iron and copper through a circuit stage.

Figure A4: Small-scale plant examples (c)–(d). Each plant has one target product. A deeper assembly sequence in (c) contrasts with several material branches converging on one science pack in (d).

Figure A5: Medium-scale plant examples (a)–(b). Each plant has two target products. An independent short production line accompanies the deeper chain in (a); the two targets reuse upstream products in (b).

Figure A6: Medium-scale plant examples (c)–(d). Each plant has two target products. Parallel material branches in (c) contrast with the longer transport and engine chains in (d).

Figure A7: Large-scale plant examples (a)–(b). Each plant has four target products. Shorter parallel paths in (a) contrast with the reuse of a compact upstream circuit and belt network in (b).

Figure A8: Large-scale plant examples (c)–(d). Each plant has four target products. Direct shipping accompanies further assembly in (c), while (d) contains fifteen processing units across several intermediate stages.

Evidence source	Exact excerpt
Log log_6056	[OBSERVED] mat-193 stopped, input buffer empty. [CHECKED] site visit — mat-205 found destroyed.
Log log_7299	[OBSERVED] mat-228 stopped, input buffer empty. [CHECKED] site visit — mat-193 found destroyed.
Ops note buffer-policy	[NORM] Line-internal feed buffers are pulled just-in-time and normally sit empty; an empty internal buffer during a stall is expected, not evidence. [NORM] Raw-material intake buffers are kept stocked by the yard at all times. An intake buffer found empty with nothing arriving means its supply feed is severed — replace that storage.
Command responses	mat-205_in_W opened: empty. flat at zero.
Interpretation	The two logs connect mat-205 to mat-193 and mat-193 to mat-228. The operating note distinguishes an intake supply cut from normal internal-buffer emptiness.

Appendix

Table A3: Chain: an illustrative example. Record excerpts and command responses from a medium-scale evaluation case.

Evidence source	Exact excerpt
Log log_5841	[CHECKED] opened st-04 — configured output does not match the plan. [ACTION] replaced st-04 with a fresh unit (as-built configuration restored).
Plan norm_3982	- unit st-05: planned output mat-231
Command responses	st-05 opened: input buffer empty. configured output: mat-232.
Interpretation	A repair precedent on st-04 applies to a different unit: st-05 is planned to output mat-231 but currently outputs mat-232.

Appendix

Table A4: Conjunction: an illustrative example. Record excerpts and command responses from a medium-scale evaluation case.

Evidence source	Exact excerpt
Log log_7360	[CHECKED] site visit — mat-027 found destroyed.
Log log_8680	[CHECKED] opened storage mat-070_in_W — found empty, no inbound material.
Log log_3569	[OBSERVED] multiple machines in the zone stopped at once. [CHECKED] zone-wide simultaneous stoppage — power distribution loss confirmed.
Command responses	mat-027 opened: input buffer mat-089 x5, mat-074 x3. mat-027: old unit removed, replacement installed to plan. [SHIFT] not fixed yet; the shift goes on. st-00 opened: input buffer empty. mat-070: running.
Interpretation	Replacing mat-027 did not end the episode, which rejects the destroyed-unit explanation of log_7360. A running machine (mat-070) is incompatible with the zone-wide power loss of log_3569. The stopped st-00 matches none of the recorded incidents and becomes the unit to investigate.

Appendix

Table A5: Elimination: an illustrative example. Record excerpts and command responses from a medium-scale evaluation case.

Scale	Products	Type	Worlds	Documents/world	Episodes	$B$
Small	1	Chain	5	7–16	13	30
		Conjunction	5	5	7	30
		Elimination	5	4–5	8	30
Medium	2	Chain	5	32–40	14	60
		Conjunction	5	30	14	60
		Elimination	5	30	8	60

Appendix

Table A6: Evaluated Tacit-FLE profiles. Document counts include maintenance logs and operating documents in each corpus. Episodes correspond to the specified fault positions;

B

is the turn budget. Scale conditions vary jointly in target products, corpus sizes, and interaction budgets.

Check	Evidence	What it establishes
Plant behavior	Healthy production, followed by stalled shipping or a persistent change in operating readings after fault injection.	The injected fault has an observable effect on an operating plant.
Corpus construction	Required document fields and vocabulary, declared evidence coverage, and exclusion of test-fault repair records.	Compliance with the corpus construction conditions.
Reproducibility	Exact regeneration of all 45 evaluation corpora.	Consistent corpus inputs for the 104 evaluation episodes.

Appendix

Table A7: Benchmark construction checks. The checks cover simulated plant behavior, corpus conditions, and reproducible evaluation inputs.

Origin	Content	Role
Task inputs	Reference corpus, task description, observations, action interface	Use and any observation controls follow the experimental protocol.
Given representation	Schema $T$ , observation map $\rho$ , template semantics	Method-specific representation and interpretation choices; fixed before learning.
Induced knowledge	$A_{W}$ , $R_{W}$ , source records $\pi\subseteq D$	Constructed from the reference corpus and fixed during evaluation.
Episode evidence	$A_{t}$	Grounded in public observations; reset for each episode.
LLM proposals	Action $a$ ; structural and diagnostic claims where supported	Checked against the current knowledge and evidence.

Appendix

Table A8: Information accounting for OntoAct. The origins and roles of information used in construction and interaction.

Case	Proposal	Admission and reason
Complete	Established supply relation; both required conditions retained; inflow still unread.	Admitted: Missing readings can be investigated after admission.
Missing relation	The proposal names a supplier relation that has no supporting evidence.	Not admitted: Establish the relation before applying this pattern.
Missing condition	The proposal cites the pattern but checks only whether storage is empty.	Not admitted: Retain the required no-inflow condition.
No conditions	The explanation supplies no observable conditions.	Not admitted: An explanation needs a nonempty set of testable conditions.

Appendix

Table A9: Admission before observation-based assessment.

Case	Stock	Inflow	Values (stock, inflow) Eq. 7	Status Eq. 8	Reason and next step
A	0	Unread	$(\top,\mathord{?})$	Open	Empty stock supports one condition; inspect incoming supply to resolve the other.
B	0	None	$(\top,\top)$	Confirmed	Both expected conditions are supported. Check the proposed repair next.
C	0	Arriving	$(\top,\bot)$	Refuted	Observed inflow contradicts the no-inflow condition. Revise the explanation.
D	5	Unread	$(\bot,\mathord{?})$	Refuted	Nonzero stock already contradicts the claim; the unread inflow cannot rescue it.
E	Unread	Unread	$(\mathord{?},\mathord{?})$	Open	Neither condition has been resolved. Obtain an informative observation.

Appendix

Table A10: Different evidence for the same proposed explanation. Independent alternatives; each case evaluates both expected conditions.

Case	Proposal	Recorded status	Conditions (stock, inflow) Eq. 7	Backed	Reason
R1	refill ( $s_{2}$ )	Confirmed	$(\top,\top)$	Yes	Prescribed verb, matching target, and both conditions supported.
R2	replace ( $s_{2}$ )	Confirmed	$(\top,\top)$	No	The pattern prescribes a refill, not replacement.
R3	refill ( $s_{3}$ )	Confirmed	$(\top,\top)$	No	The target is neither the cause nor an entity tested by this hypothesis.
R4	refill ( $s_{2}$ )	Confirmed	$(\top,\mathord{?})$	No	Stored confirmation is insufficient when a current condition is unresolved.
O1	inspect ( $q_{2}$ )	Open	$(\top,\mathord{?})$	Not required	An observation does not require a confirmed repair hypothesis.

Appendix

Table A11: Hypothesis confirmation and action backing. Independent alternatives evaluated with

H^{\prime}=\{h\}

; the backing check uses this candidate only.

Event	Result	Meaning
The refill executes and the episode terminates successfully.	Task completed	The episode ends; no additional hypothesis status is needed.
The refill executes, but another episode turn follows.	Hypothesis refuted	Under the diagnostic outcome rule, the intervention contradicts this candidate-cause claim.
The environment refuses the refill command.	No intervention evidence	Refusal alone does not refute the hypothesis; inspect the feedback and current evidence.

Appendix

Table A12: Intervention outcomes after a backed refill.

Element	Tacit-FLE	Mars	ALFWorld
Corpus $D$	One plant’s maintenance logs and operating documents	Recorded transitions from one world	Recorded transitions from training games
Vocabulary $T$	Component types, material-flow relations, statuses, and public readings	Materials, items, creatures, action and effect templates	Action verbs and five observation predicates
Induced knowledge	Structural facts, diagnostic patterns, non-evidence axioms, practices, aggregate priors	Collect, make, place, walk, drink, creature, neighborhood, and blocked-action conditions	Necessary action conditions
Evidence batch	One document; practices and service samples over the joined corpus	The supplied transition buffer	By action, at most 24 transitions per batch
Corpus checks	Type-specific source, structure, support, and counterexample checks	Subject-linked citations; at most one contradictory transition	Action-linked citations; positive support; no successful counterexample
Observation map $\rho$	Roster, layout, public readings; stock changes inform inflow	Position, inventory, vitals, visible cells	Parsed scene, inventory, and receptacle state

Appendix

Table A13: Environment-specific instantiations of OntoAct. The common construction and interaction procedures use the following representations and checks.

	Small			Medium			Large
Method	Chain	Conj.	Elim.	Chain	Conj.	Elim.	Chain	Conj.	Elim.
ReAct	0/13	5/7	6/8	4/14	9/14	6/8	1/17	8/11	9/12
ReAct *	2/13	6/7	3/8	3/14	9/14	7/8	4/17	4/11	8/12
Reflexion	1/13	5/7	6/8	3/14	6/14	3/8	2/17	1/11	6/12
IfR	5/13	7/7	6/8	4/14	12/14	8/8	4/17	10/11	12/12
NeSyC	0/13	0/7	0/8	0/14	0/14	0/8	0/17	0/11	0/12

Appendix

Table A14: Repair success counts in the Luna condition. Successes over scheduled episodes behind Table 1 . The small, medium, and large suites contain 28, 36, and 40 episodes, respectively. Unsuccessful episodes remain in the denominators. Bold: best per column, ties included.

Luna
	ReAct	ReAct *	Reflexion	IfR	OntoAct
SR $\uparrow$	52.8	52.8	33.3	66.7	83.3
SI $\downarrow$	11.3	13.4	15.4	11.9	2.4
T $\downarrow$	41.8	38.7	46.4	31.7	19.4
Luna $R$
SR $\uparrow$	44.4	38.9	52.8	86.1	97.2

Appendix

Table A15: Model conditions on medium Tacit-FLE (36 tasks). SR: repair success (%); SI: mean spurious interventions; T: mean turns. Bold/underline: best/second-best within each condition.

Reward $\uparrow$
World	ReAct	ReAct *	Reflexion	IfR	NeSyC	WALL-E	OntoAct
Default	$12.30\pm 4.82$	$\boldsymbol{15.46\pm 5.35}$	$11.30\pm 2.28$	$10.70\pm 3.21$	$5.77\pm 0.58$	$9.90\pm 2.77$	$10.90\pm 3.56$
Terrain	$13.10\pm 2.55$	$10.10\pm 5.87$	$14.50\pm 1.14$	$\boldsymbol{15.98\pm 2.06}$	$-0.57\pm 0.58$	$10.30\pm 3.11$	$14.06\pm 5.67$
Survival	$9.70\pm 5.18$	$13.70\pm 5.73$	$9.10\pm 6.78$	$12.50\pm 3.51$	$4.07\pm 3.55$	$10.50\pm 6.54$	$\boldsymbol{13.88\pm 5.60}$
Task dep.	$5.70\pm 2.30$	$2.10\pm 0.71$	$2.90\pm 1.30$	$\boldsymbol{7.70\pm 3.13}$	$1.77\pm 1.53$	$3.70\pm 2.51$	$5.70\pm 1.14$
Terr. surv.	$7.50\pm 2.61$	$5.30\pm 3.27$	$6.90\pm 2.59$	$8.30\pm 3.11$	$5.10\pm 1.00$	$7.10\pm 2.65$	$\boldsymbol{8.50\pm 3.44}$

Appendix

Table A16: Mars performance with Luna

R

. Luna

R

; five evaluation episodes per world. Reward is mean

\pm

sample SD; Score pools achievement success rates across episodes. Average excludes Default; its Reward SD is computed across matched-seed world averages. Bold/underlined: best/second-best. Metric definitions are in Appendix E.3 .

Reward $\uparrow$
World	ReAct	ReAct *	Reflexion	IfR	NeSyC	WALL-E	OntoAct
Default	$10.10\pm 2.74$	$10.70\pm 2.97$	$10.30\pm 1.10$	$\boldsymbol{13.70\pm 2.88}$	$3.50\pm 1.95$	$11.30\pm 2.28$	$12.50\pm 2.19$
Terrain	$11.30\pm 4.21$	$10.48\pm 5.82$	$11.10\pm 4.85$	$\boldsymbol{15.38\pm 4.55}$	$2.28\pm 2.16$	$10.12\pm 4.09$	$14.04\pm 3.43$
Survival	$8.10\pm 5.39$	$10.28\pm 6.87$	$7.10\pm 3.39$	$12.78\pm 8.06$	$0.70\pm 0.55$	$7.10\pm 4.06$	$\boldsymbol{15.50\pm 1.52}$
Task dep.	$6.70\pm 3.51$	$6.50\pm 4.34$	$6.10\pm 3.94$	$7.10\pm 2.24$	$0.10\pm 0.00$	$4.70\pm 1.82$	$\boldsymbol{8.10\pm 3.94}$
Terr. surv.	$7.50\pm 2.70$	$8.10\pm 4.85$	$\boldsymbol{9.90\pm 1.64}$	$5.30\pm 1.30$	$1.90\pm 1.10$	$4.10\pm 2.35$	$6.50\pm 3.13$

Appendix

Table A17: Mars performance with Terra. Terra; five evaluation episodes per world. Reward is mean

\pm

World	Method	Seed 0	Seed 1	Seed 2	Seed 3	Seed 4	Mean $\pm$ SD
Default	ReAct	7.10	17.10	7.10	13.10	6.10	$10.10\pm 4.80$
	ReAct *	10.10	11.10	7.10	8.10	10.10	$9.30\pm 1.64$
	Reflexion	8.10	3.10	8.10	11.10	9.10	$7.90\pm 2.95$
	IfR	6.10	7.10	11.10	6.10	10.10	$8.10\pm 2.35$
	NeSyC	1.10	0.10	1.10	1.10	0.10	$0.70\pm 0.55$
	WALL-E	5.10	16.10	7.10	0.10	3.10	$6.30\pm 6.06$

Appendix

Table A18: Mars reward per evaluation seed (1/2). Luna; all seven methods and seeds 0–4 from Table 3 . Each seed column is one recorded episode reward; the final column reports the mean and sample SD across the five seeds.

World	Method	Seed 0	Seed 1	Seed 2	Seed 3	Seed 4	Mean $\pm$ SD
Terr. surv.	ReAct	4.10	8.10	3.10	7.10	8.10	$6.10\pm 2.35$
	ReAct *	7.10	6.10	12.10	7.10	8.10	$8.10\pm 2.35$
	Reflexion	3.10	7.10	6.10	1.10	8.10	$5.10\pm 2.92$
	IfR	3.10	2.10	6.10	6.10	13.10	$6.10\pm 4.30$
	NeSyC	3.10	5.10	0.10	2.10	3.10	$2.70\pm 1.82$
	WALL-E	7.10	0.10	3.10	2.10	7.10	$3.90\pm 3.11$

Appendix

Table A19: Mars reward per evaluation seed (2/2). Luna; all seven methods and seeds 0–4 from Table 3 . Each seed column is one recorded episode reward; the final column reports the mean and sample SD across the five seeds.

	Standard observations					Initial-state disclosure
Method	Success	Steps	Calls	In (k)	Out (k)	Success	Steps	Calls	In (k)	Out (k)
NeSyC	0/134	49.0	1.0	5.82	0.06	35/134	38.6	9.3	269.33	11.99
WALL-E	129/134	14.8	28.8	47.08	0.61	133/134	7.8	15.6	27.83	0.38
OntoAct (ours)	132/134	13.5	13.7	32.00	0.15	134/134	6.7	8.3	23.28	0.15

Appendix

Table A20: ALFWorld with and without initial-state disclosure (Luna). Each method uses the same learned knowledge and 134 unseen tasks in both conditions, with one attempt and a 49-step budget per task. Success gives completed tasks; Steps, Calls, and input (In) and output (Out) tokens in thousands are means over all 134 episodes, including failures. The disclosure condition supplies initial object locations and inventory to every method.

Method	Pick (24)	Clean (31)	Heat (23)	Cool (21)	Examine (18)	Pick two (17)	All
NeSyC †	75.0	0.0	34.8	0.0	0.0	52.9	26.1
WALL-E	100.0	96.8	95.7	100.0	88.9	94.1	96.3
OntoAct (ours)	100.0	96.8	100.0	100.0	94.4	100.0	98.5

Appendix

Table A21: ALFWorld success (%) by task family under the Luna condition. The six families follow Shridhar et al. (2021) ; parentheses give the number of unseen tasks. All is success over all 134 tasks, weighted by family size. Best values are bold; second-best All values are underlined. Ties share the same formatting. † Initial object locations and inventory are additionally provided to NeSyC.

Small
	Construction (per plant)		Evaluation (per episode)
Method	Calls	Tokens (k)	Calls	Tokens (k)
ReAct	0	0	22.2	154.6
ReAct *	0	0	22.9	166.6
Reflexion	6.8	2.3	23.2	170.4
IfR	6.8	3.7	13.8	135.3

Appendix

Table A22: Tacit-FLE LLM usage (Luna). Construction calls and tokens are per plant (15 plants per scale); evaluation calls and tokens are per episode (28, 36, and 40 episodes). Tokens sum input and output in thousands (k). ReAct and ReAct * construct no knowledge; NeSyC, which executes no repair action (Appendix D.1 ), is omitted.

Luna
Method	Steps	Calls	In (k)	Out (k)	Input/call (k)	Total tokens (k)
ReAct	198.6	60.8	109.13	1.42	1.79	110.55
ReAct *	252.5	83.0	354.22	2.12	4.27	356.35
Reflexion	249.5	75.8	174.74	2.05	2.31	176.79
IfR	218.2	55.2	298.10	12.45	5.40	310.55
NeSyC	177.9	183.1	2237.98	13.65	12.22	2251.63

Appendix

Table A23: Mars evaluation resource use. Means over 40 episodes per method and model (eight worlds, five episodes each). Steps are environment steps; Calls, In, Out, and Total tokens are per episode. Input/call is mean input tokens per call; Total tokens sums input and output before rounding. Token quantities are in thousands (k). Knowledge-construction costs are reported separately in Appendix D.6 .

Luna
Method	Calls	In (k)	Out (k)	Input/call (k)	Total tokens (k)
ReAct	0	0	0	—	0
ReAct *	0	0	0	—	0
Reflexion	20.0	12.76	2.61	0.64	15.38
IfR	20.0	18.28	3.88	0.91	22.16
NeSyC	63.5	3768.39	33.08	59.34	3801.47

Appendix

Table A24: Mars knowledge-construction resource use. Calls, In, Out, and Total tokens are per construction run. NeSyC constructs its rules once per world (eight runs); its rule refinements during play are counted as evaluation (Table A23 ). ReAct and ReAct * require no LLM-based knowledge construction; their input length is undefined. Input/call is mean input tokens per call; Total tokens sums input and output before rounding. Token quantities are in thousands (k).

Method	Calls	In (k)	Out (k)	Total tokens (k)
NeSyC	79	1989.08	8.62	1997.70
WALL-E	254	349.37	14.84	364.21
OntoAct (ours)	15	116.43	0.94	117.36

Appendix

Table A25: ALFWorld knowledge-construction LLM usage (Luna). Totals for one construction run per method on the shared 238-transition corpus. Each resulting knowledge base is reused across 134 evaluation tasks. Token quantities use thousands of tokens (k); Total tokens sums input and output before rounding. Costs exclude corpus collection and evaluation.

Scale	Method	Targets	C@1	C@3	C@5	C@all	Precision
Small	OntoAct (ours)	2.68	50.0	75.0	82.1	82.1	23/75
	WALL-E	4.79	3.6	35.7	50.0	50.0	14/134
	IfR	5.79	10.7	28.6	50.0	64.3	18/162
Medium	OntoAct (ours)	3.36	33.3	69.4	75.0	88.9	32/121
	WALL-E	5.14	0.0	27.8	36.1	47.2	17/185
	IfR	9.00	0.0	11.1	38.9	66.7	24/324

Appendix

Table A26: Replacement-candidate diagnostics. Targets is the mean number of distinct proposed replacement targets per episode. C@

k

and C@all are true-fault coverage (%) within the first

k

targets and the complete sequence. Precision is correct-target/total episode–target pairs. Small, medium, and large contain 28, 36, and 40 episodes.

Checkpoint / measure	Small	Medium	Large
All tracked	20/140	28/272	34/192
Checkpoint +1	19/68	23/112	31/98
Checkpoint +2	19/48	21/61	28/68
Checkpoint +3	19/32	20/36	28/42
Final	18/21	20/21	25/32
True-target retention	18/20	20/28	25/34

Appendix

Table A27: Selection within the hypothesis ledger. Target-aligned/all tracked at baseline, then target-aligned/retained at later checkpoints (retained: open or confirmed). The last two rows give final target-aligned retention and off-target refutation; terminal-only hypotheses (9/15/14 by scale) are excluded.

Method	Small	Medium	Large
OntoAct (ours)	29.3% (22/75)	25.4% (29/114)	36.4% (36/99)
WALL-E	2.3% (3/133)	5.2% (12/232)	3.8% (14/364)
IfR	9.9% (18/182)	5.6% (24/426)	3.7% (26/694)
ReAct	8.9% (11/124)	4.7% (15/321)	3.5% (17/482)
ReAct *	5.6% (11/197)	4.3% (17/392)	2.2% (15/688)
Reflexion	8.6% (12/140)	3.1% (12/387)	1.3% (8/600)

Appendix

Table A28: Precision of executed replacement commands. Each cell gives the percentage and correct-target/total commands. Repeated targets are counted each time. These final executions are distinct from proposed candidates and native gate passes.

Ontology-Based Reasoning

See all 15 papers

May 4, 2026

ANNEAL: Adapting LLM Agents via Governed Symbolic Patch Learning

LLM-based agents can recover from individual execution errors, yet they repeatedly fail on the same fault when the underlying process knowledge--operator schemas, preconditions, and constraints--remains unrepaired. Existing self-evolving approaches address this gap by updating prompts, memory, or model weights, but none directly repair the symbolic structures that encode how tasks are executed, and few provide the governance guarantees required for safe deployment. We introduce ANNEAL, a neuro-symbolic agent that converts recurring failures into governed symbolic edits of a process knowledge graph without modifying foundation model weights. Its core mechanism, Failure-Driven Knowledge Acquisition (FDKA), localizes the responsible operator, synthesizes a typed patch through constrained LLM generation, and validates the proposal via multi-dimensional scoring, symbolic guardrails, and canary testing before commit. Every accepted edit carries full provenance and deterministic rollback capability. Across four domains and 27 multi-seed runs, ANNEAL is the only evaluated system that commits persistent structural repairs--strong baselines such as ReAct and Reflexion achieve high episodic recovery yet retain 72--100% holdout failure rates on recurring faults, whereas ANNEAL reduces these to 0% in the tested recurring-failure settings. Ablation confirms that removing FDKA eliminates all structural repairs and drops success rate by up to 26.7 percentage points. These results suggest that governed symbolic repair offers a complementary paradigm to weight-level and prompt-level adaptation for persistent fault elimination.

Safayat Bin Hakim, Keyan Guo, Wenkai Tan +3

University of Maryland, Baltimore County · University at Buffalo · University of Colorado Boulder +1

LLM Agent Self-Improvement AI Agent Governance Neuro-Symbolic AI Knowledge Graphs

Ask PDF

Apr 1, 2026

Ontology-Constrained Neural Reasoning in Enterprise Agentic Systems: A Neurosymbolic Architecture for Domain-Grounded AI Agents

Enterprise adoption of Large Language Models (LLMs) is constrained by hallucination, domain drift, and the inability to enforce regulatory compliance at the reasoning level. We present a neurosymbolic architecture implemented within the Foundation AgenticOS (FAOS) platform that addresses these limitations through ontology-constrained neural reasoning. We introduce a three-layer ontological framework--Role, Domain, and Interaction ontologies--grounding LLM-based enterprise agents. We formalize asymmetric neurosymbolic coupling: current enterprise systems constrain agent inputs (context assembly, tool discovery, governance thresholds) but not outputs, and we propose mechanisms extending this coupling to output-side validation (response checking, reasoning verification, compliance enforcement). A controlled experiment (1,800 runs across five industries and three LLMs: Claude Sonnet 4, Qwen 2.5 72B, Gemma 4 26B) finds ontology-coupled agents significantly outperform ungrounded agents on Metric Accuracy (p < .001) and Role Consistency (p < .001) across all three models with large effect sizes (Kendall's W = .46-.64). Improvements are greatest where LLM parametric knowledge is weakest--particularly in Vietnam-localized domains, where ontology lift is 2x that of English domains. Contributions: (1) a formal three-layer enterprise ontology model; (2) a taxonomy of neurosymbolic coupling patterns; (3) ontology-constrained tool discovery via SQL-pushdown scoring; (4) a proposed framework for output-side ontological validation; (5) empirical evidence for the inverse parametric knowledge effect--ontological grounding value is inversely proportional to LLM training-data coverage of the domain; (6) cross-model replication establishing model-independence; (7) a production system serving 22 industry verticals with 650+ agents.

Thanh Luong Tuan, Abhijit Sanyal

Golden Gate University, San Francisco · Foundation AgenticOS (FAOS) · Associate Director, Data, Digital & IT, Novartis Healthcare Pvt. Ltd., Hyderabad, India

Ontology-Based Reasoning Enterprise AI Agents LLM Grounding LLM Agents Neuro-Symbolic Reasoning LLM Agent Reliability

Ask PDF

Apr 29, 2026

AGEL-Comp: A Neuro-Symbolic Framework for Compositional Generalization in Interactive Agents

Large Language Model (LLM)-based agents exhibit systemic failures in compositional generalization, limiting their robustness in interactive environments. This work introduces AGEL-Comp, a neuro-symbolic AI agent architecture designed to address this challenge by grounding actions of the agent. AGEL-Comp integrates three core innovations: (1) a dynamic Causal Program Graph (CPG) as a world model, representing procedural and causal knowledge as a directed hypergraph; (2) an Inductive Logic Programming (ILP) engine that synthesizes new Horn clauses from experiential feedback, grounding symbolic knowledge through interaction; and (3) a hybrid reasoning core where an LLM proposes a set of candidate sub-goals that are verified for logical consistency by a Neural Theorem Prover (NTP). Together, these components operationalize a deduction--abduction learning cycle: enabling the agent to deduce plans and abductively expand its symbolic world model, while a neural adaptation phase keeps its reasoning engine aligned with new knowledge. We propose an evaluation protocol within the Retro Quest simulation environment to probe for compositional generalization scenarios to evaluate our AGEL agent. Our findings clearly indicate the better performance of our AGEL model over pure LLM-based models. Our framework presents a principled path toward agents that build an explicit, interpretable, and compositionally structured understanding of their world.

Mahnoor Shahid, Hannes Rothe

Universität Duisburg-Essen, Essen, Germany

Neuro-Symbolic AI Inductive Logic Programming LLM Agent Planning Compositional Generalization

Ask PDF

Abstract

Ontology-Based Reasoning

Explore similar work

ANNEAL: Adapting LLM Agents via Governed Symbolic Patch Learning

Ontology-Constrained Neural Reasoning in Enterprise Agentic Systems: A Neurosymbolic Architecture for Domain-Grounded AI Agents

AGEL-Comp: A Neuro-Symbolic Framework for Compositional Generalization in Interactive Agents