Temporal Knowledge Graph Question Answering (TKGQA) requires language model agents to retrieve evidence, select relevant facts, and perform multi-step temporal reasoning. However, the trajectory-level outcome signal in GRPO does not explicitly distinguish the quality of intermediate decisions: downstream failures can penalize useful evidence acquisition, while identical rewards in all-failed groups eliminate the relative outcome signal despite differences in intermediate progress. We propose Credit Redistribution with Evidence-Driven Opportunities (CREDO), which uses intermediate evidence to recover local supervision, including in all-failed groups, and determine where outcome supervision acts. CREDO assigns positive stage-level credit through comparisons among rollouts of the same question and uses retained answer overlap as an opportunity proxy to shift part of the outcome-policy weight toward reasoning and answering. The framework preserves terminal rewards and trajectory-level advantages, while redistribution conserves total outcome-policy weight per trajectory. Following the public evaluation protocol of prior work, CREDO raises Overall accuracy from 78.0% to 82.4% and Multiple accuracy from 55.0% to 65.6% on MultiTQ relative to the strongest published baseline in our comparison. On TimelineKGQA, it raises cross-KG accuracy from 30.9% to 36.4% relative to our reproduced baseline via direct transfer to ICEWS Actor under matched training and evaluation settings.
Figures & tables
Figure 1: Motivation of CREDO. In a mixed-outcome group, useful intermediate decisions can receive negative outcome credit when the final answer is wrong. CREDO uses intermediate evidence for within-question comparisons and evidence-conditioned outcome redistribution.
Figure 2: Framework of CREDO. CREDO evaluates retrieved and retained evidence to recover local credit through within-question comparisons and redistribute outcome-policy weight toward reasoning and answering. The illustration uses ρ=0.4 ; the experimental default is ρ=0.04 .
1
for each (x,G)∈B do
2
Sample K trajectories with πold ; compute original rewards and GRPO advantages Ai .
3
Extract evidence and stage masks; determine token availability tis .
4
Compute qiQ , cisel , qiE , and qiR using Eqs. ( 1 )–( 2 ).
5
Determine comparison validity vis ; set ωiE=qiQ and ωiR=cisel .
6
for each stage s∈{Q,E,R} do
7
Compute peer comparisons Fis and local advantages ais (Eqs. ( 4 )–( 5 )).
Table 3
Figure 3: All-wrong groups and those with active local credit in CREDO Stage 1 (8B). Both denominators count all sampled groups.
Model
Overall
Question Type
Answer Type
Multiple
Single
Entity
Time
TKG Embedding-based methods
EmbedKGQA
0.206
0.134
0.235
0.290
0.001
CronKGQA
0.279
0.134
0.337
0.328
0.156
MultiQA
0.293
0.159
0.347
0.349
0.157
Prompt-based LLMs
Table 1: Performance comparison on MultiTQ. CREDO follows the normalized substring-matching protocol in Temp-R1’s public evaluation code. Baseline scores are adopted from Table 3 of Temp-R1 ( Gong et al., 2026 ) . The best and second-best scores under each metric are highlighted in colors, respectively; ties receive the same highlight.
Model
CronQuestion KG (In Domain)
ICEWS Actor (Out of Domain)
Overall
Simple
Medium
Complex
Overall
Simple
Medium
Complex
RAG Baseline
0.235
0.704
0.092
0.009
0.265
0.660
0.128
0.011
LLaMA2-7B
0.169
0.049
0.143
0.282
0.111
0.035
0.066
0.322
GPT-4o
0.206
0.069
0.130
0.376
0.113
0.051
0.035
0.353
RTQA
0.298
0.608
0.218
0.135
–
–
–
–
Temp-R1 ∗
0.508
0.844
0.241
0.446
0.309
0.447
0.123
0.324
Table 2: Performance comparison on TimelineKGQA across in-domain (CronQuestion KG) and out-of-domain (ICEWS Actor) settings. ∗ Our reproduction, using the same base training configuration as CREDO and Temp-R1’s public scoring protocol (Appendix A.2 ). Other baseline results are adopted from Temp-R1 ( Gong et al., 2026 ) . The best and second-best scores under each metric are highlighted in colors, respectively; ties receive the same highlight.
Figure 4: MultiTQ component comparisons: (a) Overall; (b) Multiple. GRPO uses the Temp-R1 reported results ( Gong et al., 2026 ) . CWO disables redistribution; Random changes its target tokens; No-gate adds ungated answer-region redistribution to CWO. See Appendix A.2 for selection details.
Figure 5: CREDO diagnostics. (a) Stage-wise accuracy on 500 MultiTQ inputs; the GRPO Stage 2 score is from our reproduction under the same training and evaluation settings as CREDO. (b) Additional substring-score loss from temporal rather than generic distraction. (c) Substring accuracy with temporal distractors, averaged over three decoding seeds on 67 questions. (d) Official Multiple types from the paired evaluation, with group sizes below the axis. Sources: Appendix A.2 .
λlocal
ρ
Multiple (%)
0.03
0.04
45.00
0.01
0.04
43.75
0.05
0.04
45.00
0.03
0.02
45.00
0.03
0.06
45.62
Table 3: Peak Multiple accuracy on MultiTQ (3B), using the same evaluation settings as the 8B experiments. The default configuration is shown in bold.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Setting
Stage 1
Stage 2
Scheduled updates
80
500
Questions per update × rollouts
60×5
2×8
PPO minibatch trajectories
30
2
Data / sequence parallelism
1 / 8
2 / 4
Learning rate
5×10−7
5×10−7
Gradient clipping
5
1
Appendix
Table 4: Common base training configuration for the five in-house 8B methods on MultiTQ. Generation limits and response-container limits measure different quantities.
Category
N
GRPO
CWO
CREDO
Equal
174
156 (89.66)
151 (86.78)
152 (87.36)
Before/After
94
81 (86.17)
89 (94.68)
87 (92.55)
First/Last
72
65 (90.28)
62 (86.11)
68 (94.44)
Equal Multi
33
22 (66.67)
28 (84.85)
28 (84.85)
Before Last
69
34 (49.28)
38 (55.07)
41 (59.42)
After First
58
27 (46.55)
33 (56.90)
36 (62.07)
Appendix
Table 5: Paired MultiTQ results: correct answers (accuracy, %). CREDO and CWO also provide the main results; GRPO is our matched in-house baseline. Aggregate rows overlap the individual categories.
Comparator / subset
N
Corrected
Regressed
Difference
95% CI
GRPO / Overall
500
50
23
+5.40
[2.20,8.80]
CWO / Overall
500
34
23
+2.20
[−0.80,5.20]
GRPO / Multiple
160
33
11
+13.75
[6.25,21.88]
CWO / Multiple
160
19
13
+3.75
[−3.13,10.63]
Appendix
Table 6: Paired supplementary comparisons. “Corrected” means only CREDO is correct; “Regressed” means only the comparator is correct. Differences and intervals are in percentage points.
Evidence condition
GRPO
CWO
CREDO
Supporting facts only
178 (88.56)
195 (97.01)
194 (96.52)
Generic, k=2
164 (81.59)
184 (91.54)
196 (97.51)
Generic, k=4
170 (84.58)
184 (91.54)
198 (98.51)
Generic, k=8
154 (76.62)
181 (90.05)
196 (97.51)
Temporal, k=2
157 (78.11)
180 (89.55)
196 (97.51)
Temporal, k=4
152 (75.62)
178 (88.56)
197 (98.01)
Appendix
Table 7: Answer containment under fixed evidence. Entries are correct generations out of 201 (substring accuracy, %), from 67 questions and three decoding seeds.
Evidence condition
GRPO
CWO
CREDO
Complete evidence
60 (33.33)
114 (63.33)
125 (69.44)
Generic, k=4
58 (32.22)
96 (53.33)
119 (66.11)
Temporal, k=4
62 (34.44)
97 (53.89)
100 (55.56)
Appendix
Table 8: Multiple-question answer containment with complete support retained. Entries are correct generations out of 180 (substring accuracy, %), from 60 questions and three decoding seeds. Bold indicates the highest accuracy in each condition.
Institute of Information Engineering, Chinese Academy of Sciences, Beijing, China · School of Cyber Security, University of Chinese Academy of Sciences, Beijing, China