Despite advances in LLM-based table reasoning, most studies focus on single-table settings. Multi-table QA requires multi-hop reasoning over complex table relationships and evidence dispersed across long contexts, often without verifiable intermediate traces. To address these challenges, we introduce EviDA, an evidence-centric dual-agent framework comprising an Evidence Construction Agent and a Reasoning and Answering Agent. The Evidence Construction Agent generates and executes table-manipulation programs over complete tables to produce question-specific evidence and verifiable intermediate traces, simplifying subsequent reasoning. The Reasoning and Answering Agent then performs multi-hop reasoning over this evidence to derive final answers. We further introduce EviDARL, which combines agent-specific fine-grained rewards with parameter-efficient alternating reinforcement learning to improve collaboration between the agents. By optimizing each agent in turn while holding the other fixed, EviDARL enables bidirectional adaptation between evidence construction and answer reasoning. Experiments on MMQA and TQA-Bench demonstrate that EviDARL substantially outperforms competitive open-model and multi-table reasoning baselines under constrained model and training budgets.
Figures & tables
Figure 1: An illustrative comparison between EviDARL and single-agent approaches. Single-agent methods may ground reasoning in incorrect information, leading to erroneous reasoning traces and answers, while EviDARL constructs relevant evidence across tables and derives correct answers.
Figure 2: Overview of EviDA and EviDARL. (A) EviDA connects executable evidence construction with evidence-grounded reasoning through a verifiable intermediate interface and a fallback mechanism. (B) EviDARL employs mode-specific rollouts and alternating GRPO updates to enable stable co-evolution between the two agents.
MMQA
TQA-Bench
Base Model
2-table
3-table
Avg
8K
64K
VME
EM
EM
EM
Acc
Acc
Claude-Sonnet-5
59.28
63.73
60.69
93.40
92.00
✗
EviDA(Claude-Sonnet-5)
68.78 ↑
68.63 ↑
68.73 ↑
93.40 →
92.80 ↑
✓
Deepseek-V4-Flash
57.92
63.73
59.75
91.80
90.20
✗
EviDA(Deepseek-V4-Flash)
60.18 ↑
66.67 ↑
62.23 ↑
92.60 ↑
91.40 ↑
✓
Table 1: Effectiveness of EviDA across different base models on MMQA and TQA-Bench. VME indicates whether the method constructs verifiable intermediate evidence.
MMQA
TQA-Bench
Base Model
Method
2-table
3-table
Avg
8K
64K
EM
PM
EM
PM
EM
PM
Acc
Acc
Proprietary LLMs
Claude-Sonnet-5
Prompt + Thinking
59.28
71.04
63.73
73.53
60.69
71.83
93.40
92.00
Claude-Opus-4.8
Prompt + Thinking
57.47
64.25
59.80
69.61
58.20
65.94
92.80
91.40
Deepseek-V4-Flash
Prompt + Thinking
57.92
77.38
63.73
75.49
59.75
76.78
91.80
90.20
Table 2: Main results on MMQA and TQA-Bench. Comparison with proprietary LLMs, table-specific LLMs, and multi-table reasoning methods. The best and second-best results in each column are shown in bold and underlined, respectively.
Comparison
MMQA
2-table
3-table
EM
ECS
ES
EM
ECS
ES
EviDARL
70.59
95.02
72.38
61.76
88.24
68.89
EviSARL-AC
57.92
75.57
62.87
48.04
73.53
60.00
EviSARL-FG
61.99
90.95
57.12
57.84
81.37
57.83
Table-R1
59.73
NA
NA
56.86
NA
NA
Table 3: Comparison of dual-agent and single-agent optimization on MMQA.
Settings
MMQA
TQA-Bench
2-table
3-table
Avg
8K
64K
EM
ECS
ES
EM
ECS
ES
EM
ECS
ES
Acc
ECS
ES
Acc
ECS
ES
Llama3.1-8B-Instruct
13.57
NA
NA
10.78
NA
NA
12.69
NA
NA
48.60
NA
NA
36.40
NA
NA
EviDA
18.10
72.85
4.97
12.74
60.78
3.23
16.41
69.04
4.48
46.80
67.40
21.66
37.60
63.20
23.73
SFT-EviDA
54.75
79.64
54.55
48.04
71.57
58.90
52.63
77.09
55.82
74.20
84.60
71.16
72.80
83.40
72.42
SFT-EviDARL
67.42
94.57
70.81
53.92
90.20
57.61
63.16
93.19
66.78
87.20
98.40
86.79
86.80
97.80
88.55
Table 4: Ablation of training components on MMQA and TQA-Bench. EM, ECS, and ES denote exact match, evidence-construction success rate, and evidence sufficie ncy. Best and second best results are bolded and underlined.
Figure 3: Error distributions of EviDARL on (a) MMQA and (b) TQA-Bench, decomposed into reasoning- and evidence-related failures.
Figure 4: Performance of EviDARL on MMQA and TQA-Bench across question types. Bars report the proportions of correct and incorrect predictions within each category.
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 5: System and user prompt templates for EC Agent and RA Agent in EviDA.
Training Setting
MMQA
2-table
3-table
EM
ECS
ES
EM
ECS
ES
Sequential
44.80
97.74
45.83
31.68
92.09
34.78
By-Epoch (ours)
70.59
95.02
72.38
61.76
88.24
68.89
By-Half-Epoch
69.23
95.48
71.09
61.76
87.25
70.79
By-Step
65.16
94.12
69.23
59.80
88.24
67.78
Appendix
Table 5: Comparison of sequential and alternating dual-agent training schedules on MMQA. The By-Epoch and By-Half-Epoch variants switch the optimized agent every epoch and half-epoch, respectively. The By-Step variant first completes one epoch-level alternation cycle and then switches the active agent after every optimization step.
Method
MMQA
Thinking
E2E
LLM
Code
TTFT(ms)
Gen. Tokens
Table-R1
12.67
12.67
0.000
31.68
1821
✓
EviDARL
5.67
5.66
0.014
5670.6
523
✗
EC Agent
4.93
4.91
0.014
24.6
457
✗
RA Agent
0.75
0.75
0.000
53.7
66
✗
Appendix
Table 6: Inference latency of Table-R1 and EviDARL and the two EviDARL agent stages. E2E, LLM, and Code denote the runtime of the entire workflow, LLM inference, and code execution, respectively, measured in seconds.
Figure 6: Ablation results on MMQA and TQABench with Qwen3-8B and Llama-3.1-8B-Instruct. We compare EviDARL without task-specific post-training, EC Agent+ with only the EC Agent trained, RA Agent+ with only the RA Agent trained, and EviDARL with both agents optimized through alternating RL.
Institute of Systems Engineering, Academy of Military Sciences, Beijing, 100101, China · School of Information Resource Management, Renmin University of China, Beijing, 100872, China