System 1 decision models (represented by Jev) provide fast, low-cost action selection, but existing agent integrations do not reliably translate these local advantages into task completion. At certain points during execution, an agent may repeatedly select the same actions without making progress toward its goal. LLM-based policies offer strong reasoning capabilities that support planning within an agent, but invoking them at every step incurs substantial overhead. The effective coordination should therefore preserve Jev's fast local action selection while introducing LLM reasoning when execution stalls. We observe that recurring state-action pairs can signal stalled execution. We therefore propose LoopRoute, a simple but effective action router that uses this signal to allocate control within an agent. The decision model acts by default. When a loop is detected, the LLM provides brief, feedback-driven assistance and then returns control to the decision model. Later loops can request further assistance. The rule requires neither additional training nor model-specific confidence calibration. Across three benchmarks and seven LLMs, LoopRoute improves task success over standalone Jev with only limited use of the LLM policy. The mean gains are 21.0%, 60.1%, and 30.1% on ALFWorld, MapReader, and TextWorld cooking, respectively. Compared with full LLM control, LoopRoute also substantially reduces the mean cumulative API time and API-call cost per task. The same rule also improves the success rates of six frozen decision models.
Figures & tables
Figure 1: Local selection versus long-horizon execution with standalone Jev (robot). (a) Static tool selection on BFCL Multiple. (b) Illustrative loops stall a multi-step task. Loop onsets are averaged over all tasks. Repeated-step shares are pooled over failed tasks only. TextWorld denotes cooking (Appendix D ).
Figure 2: Overview of LoopRoute . The System 1 decision model proposes an action from a shared execution context. The router checks how often the proposed state–action pair has been executed: proposals below the repetition threshold execute directly, while a repeated proposal triggers at most two feedback-driven LLM actions before control returns to the decision model. Both policies update the same execution context after each environment transition.
Table 1: An ALFWorld trajectory in which Jev cycles through microwave actions after storing an uncooled tomato. LoopRoute intercepts its third proposal to examine microwave 1 under the same observed state. The LLM policy takes two actions and returns control. Jev completes the task at step 34. Jev alone reaches the 100-step limit. The icons mark stalled selection, LLM assistance, and resumed progress.
ALFWorld
MapReader
TextWorld
Model
Policy
SR (%)
Time (s)
Cost ($)
SR (%)
Time (s)
Cost ($)
SR (%)
Time (s)
Cost ($)
Jev
Alone
77.6
13.4
0.0074
39.0
20.9
0.0059
54.0
16.5
0.0096
Jev
+ Rank
88.8
13.4
0.0062
62.0
16.4
0.0050
66.0
13.0
0.0081
Jev
+ Requery
82.8
15.3
0.0077
62.0
22.4
0.0072
66.0
17.0
0.0116
GPT-6-Astra
Full
100.0
64.7
0.4600
100.0
161.6
0.4582
100.0
193.8
0.7517
Routed
100.0
18.5
0.0528
100.0
85.1
0.1771
83.0
79.8
0.2700
Table 2: Task success and resource use. SR denotes the success rate (%). Alone uses only Jev. Rank and Requery apply loop-breaking controls without LLM assistance. Full uses the LLM policy throughout. Routed uses Jev + LoopRoute . Time and cost are per-task means and include both policies for routed runs. The bottom rows average the seven LLM comparisons equally. Mean gain gives the signed success-rate difference (Routed minus Full) and the relative time/cost reductions ( ↓ ). The success-rate difference is expressed on the percentage scale, not as relative growth. Resource reductions are computed per pair before averaging. Cost accounting is in Appendix A .
Figure 3: Reduction in LLM requests under LoopRoute across seven LLMs. Each bar compares routed and complete policies using the same LLM and matched task sets, on a common 0–100% scale. Task success, per-task time, and USD cost are in Table 2 . The success counts and mean episode lengths are in Appendix A .
Policy
SR (%)
LLM
Env.
ALFWorld
LoopRoute
99.3
0.98
28.16
Random (lower)
96.3
0.81
27.46
Random (higher)
97.8
1.15
23.33
MapReader
LoopRoute
99.0
4.50
21.97
Table 3: Comparison with random intervention. LLM and Env. are executed LLM and total environment actions per task. MapReader and TextWorld random use is within 10% of routed LLM actions. Both ALFWorld random rounds fall outside this tolerance. All rounds appear in Appendix E .
ALFWorld
MapReader
TextWorld
τ
SR (%)
Req.
SR (%)
Req.
SR (%)
Req.
1
100.0
32
80.0
718
85.0
256
2
100.0
14
80.0
538
85.0
281
3
95.0
22
60.0
640
80.0
313
Table 4: Threshold sensitivity with Jev and DeepSeek-V4.1-Flash on 20 additional tasks per benchmark. SR is task success. Req. counts all LLM HTTP requests. The takeover budget remains B=2 .
ALFWorld
MapReader
TextWorld
B
SR (%)
Req.
SR (%)
Req.
SR (%)
Req.
1
100.0
13
75.0
645
85.0
241
2
100.0
14
80.0
538
85.0
281
3
95.0
52
75.0
615
80.0
243
Table 5: Takeover-budget sensitivity on the same 20 tasks per benchmark as Table 4 , with Jev, DeepSeek-V4.1-Flash, and τ=2 . SR is task success. Req. counts all LLM HTTP requests. The B=2 trajectories are reused from that study.
ALFWorld
MapReader
Cooking
Decision model
Alone (%)
Routed (%)
Gain
Alone (%)
Routed (%)
Gain
Alone (%)
Routed (%)
Gain
Jev
77.6
94.8
+17.2%
39.0
97.0
+58.0%
54.0
83.0
+29.0%
JEV-27B-VL
85.1
97.0
+11.9%
67.0
99.0
+32.0%
74.0
84.0
+10.0%
StartLux-4B
53.7
93.3
+39.6%
32.0
91.0
+59.0%
28.0
43.0
+15.0%
OneJev1.6-4B
20.9
93.3
+72.4%
28.0
98.0
+70.0%
14.0
23.0
+9.0%
Kev-4B
17.9
94.8
+76.9%
17.0
99.0
+82.0%
17.0
50.0
+33.0%
Table 6: Transfer across six frozen decision models with unchanged τ=2 and B=2 , using DeepSeek-V4.1-Flash for LLM assistance. Alone and Routed give success rates in percent. Gain is their signed absolute difference (Routed minus Alone), not relative growth. The complete LLM policy achieves 94.0% on ALFWorld, 98.0% on MapReader, and 92.0% on cooking. Resource results are in Appendix A . Laya cooking results are unavailable.
Method
Success
Lost
Limit
Jev
54
3
43
+ Rank
66
13
21
+ LoopRoute
83
11
2
DeepSeek + StateAct
92
6
0
Table 7: Success and selected failure outcomes in 100 cooking games. Routed and StateAct runs use DeepSeek-V4.1-Flash. Lost denotes a game loss after an irreversible action. Limit denotes exhaustion of the action budget. Requery matches Rank.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Decision model / policy
LLM requests
Saved (%)
Time (s)
Saved (%)
ALFWorld (134 tasks)
Full LLM
2,941
–
157.3
–
Jev
230
92.2
23.6
85.0
JEV-27B-VL
176
94.0
47.6
69.8
StartLux-4B
419
–
70.8
–
OneJev1.6-4B
678
76.9
163.4
-3.8
Appendix
Table 8: Resource use for the transfer runs in Table 6 . Full LLM is the complete DeepSeek-V4.1-Flash policy. Saved is the relative reduction against a matching reference, with negative values indicating increases. Time is the per-task sum for both policies. Dashes mark unmatched comparisons. Jev costs are already reported in Table 2 . Other selectors lack consistently priced usage for both policies.
ALFWorld
MapReader
Cooking
LLM
Full
Routed
Steps F/R
Full
Routed
Steps F/R
Full
Routed
Steps F/R
GPT-6-Astra
134/134
134/134
9.5/27.1
100/100
100/100
12.7/22.1
100/100
83/100
15.0/22.3
Claude Opus 5
134/134
132/134
11.0/28.6
100/100
99/100
13.0/25.4
91/100
81/100
16.5/25.0
Gemini 3.8 Flash
134/134
134/134
8.2/28.2
100/100
100/100
12.3/21.9
100/100
83/100
14.4/22.6
Grok 4.5
134/134
134/134
11.7/27.5
98/98
98/98
12.3/21.7
100/100
87/100
14.4/22.6
GLM-5.3
134/134
133/134
10.2/29.5
100/100
100/100
12.8/23.6
95/100
89/100
15.2/25.4
Appendix
Table 9: Success counts and mean episode lengths for Table 2 . Full and Routed report successes over the evaluated task count. Steps F/R gives mean environment actions for complete and routed policies, including unsuccessful episodes. Request comparisons use the same task IDs as these counts.
Routing rule
Success
Requests
Repetition only ( LoopRoute )
127/134
230
+ confidence <0.15
127/134
399
Appendix
Table 10: Confidence-gate comparison on 134 matched ALFWorld tasks. Requests include HTTP retries.
Mean
Repeated-step share (%)
Benchmark
onsets
All tasks
Failed tasks
ALFWorld
1.522
28.26
47.63
MapReader
2.370
83.52
87.87
TextWorld
2.090
67.96
77.46
Appendix
Table 11: Standalone Jev repetition. ALFWorld, MapReader, and TextWorld contain 134, 100, and 100 tasks, of which 30, 61, and 46 fail. Failed-task shares are respectively 1,429/3,000, 5,360/6,100, and 3,375/4,357 steps.
Policy/round
p
Success
LLM
Env.
ALFWorld ( N=134 )
LoopRoute
–
133
131
3,773
Random initial
0.019396
125
78
3,885
Random extra 1
0.032575
129
109
3,680
Random extra 2
0.039150
131
154
3,126
MapReader ( N=100 )
Appendix
Table 12: All random-comparison rounds. Success counts retain the full N for each row. LLM and Env. are total executed actions, not per-task means. Initial evaluation contains 668 trajectories and the additional rounds contain 568.
LLM agent routing selects models, tools, skills, or collaboration workflows for each task or step to balance capability, cost, latency, and reliability.