Let's Think in Two Steps: Mitigating Agreement Bias in MLLMs with Self-Grounded Verification
Organizations: Georgia Institute of Technology
Abstract
Verifiers--functions assigning rewards to agent behavior--have been key to AI progress in math, code, and games. However, extending gains to domains without clear-cut success criteria (e.g., computer use) remains a challenge: while humans can recognize desired outcomes, translating this intuition into scalable rules is nontrivial. Multimodal LLMs (MLLMs) offer a promising solution, given their world knowledge, human-preference alignment, and reasoning skills. We evaluate MLLM verifiers across web navigation, computer use, and robotics, spanning 13 models, 28 judge designs, diverse agent trajectories, and distinct downstream applications. We identify a critical limitation: a strong tendency for MLLMs to over-validate agent behavior, a phenomenon we term agreement bias. This bias is pervasive across models, is resilient to test-time scaling, and harms pipelines relying on MLLM judgments (e.g., self-improvement). We discuss principles for design and evaluation of verifiers, and introduce SGV, a method that improves verification by harnessing MLLMs' own capabilities via modulation of (un)conditional generation. First, an MLLM generates broad priors about desired behavior, independent of the data under evaluation. Then, conditioned on generated priors, it reasons over and judges a candidate trajectory. Our methods yield more human-aligned verifiers, with gains of up to 25pp in failure detection, 14pp in accuracy, and benefits extending to downstream applications. In self-improvement and online supervision, our verifiers boost task completion of a GUI specialist in OSWorld, a diffusion policy in robomimic, and a ReAct agent in VisualWebArena--surpassing the previous SOTA by 20pp. Finally, we release an updated version of VisualWebArena, featuring improved evaluators, high-fidelity environment parallelism with >10x speedups, and VWA-Lite, a 1/3-scale subset with comparable evaluation fidelity.
Figures & tables
| Model | (a) No SGV | (b) SGV | (b) - (a) | ||||||||||
| Acc | TPR | TNR | Bias | dSkew | Acc | TPR | TNR | Bias | dSkew | Acc | Bias | dSkew | |
| Gemini 2.0 | 61 | 96 | 42 | 36 | 34 | 69 | 94 | 55 | 27 | 23 | +9 | -9 | -11 |
| Gemini 2.5 Lite | 55 | 96 | 34 | 41 | 39 | 65 | 90 | 51 | 29 | 25 | +9 | -11 | -13 |
| Gemini 2.5 Flash | 68 | 94 | 55 | 27 | 24 | 80 | 88 | 76 | 12 | 8 | +12 | -15 | -16 |
| Gemini-2.5-Flash (T) | 74 | 92 | 64 | 21 | 18 | 82 | 89 | 78 | 10 | 6 | +8 | -11 | -13 |
| Qwen3-32b | 69 | 92 | 57 | 25 | 21 | 76 | 88 | 71 | 14 | 11 | +7 | -11 | -10 |
Appendix figures & tables34 assets
Supplementary material from the paper’s appendix.
Appendix
| Model | (a) No SGV | (b) SGV | (b) - (a) | ||||||||||
| Acc. | TPR | TNR | Bias | dSkew | Acc. | TPR | TNR | Bias | dSkew | Acc. | Bias | dSkew | |
| Gemini 2.0 | 56 | 93 | 28 | 38 | 35 | 65 | 90 | 42 | 28 | 23 | +9 | -10 | -12 |
| Gemini 2.5 Lite | 56 | 91 | 26 | 36 | 32 | 62 | 84 | 42 | 28 | 21 | +5 | -8 | -10 |
| Gemini 2.5 | 65 | 90 | 47 | 26 | 21 | 76 | 84 | 71 | 11 | 5 | +11 | -15 | -16 |
| Gemini-2.5 (T) | 70 | 90 | 55 | 23 | 18 | 77 | 86 | 70 | 12 | 6 | +7 | -11 | -12 |
| GPT-4.1 Mini | 59 | 94 | 31 | 35 | 31 | 72 | 84 | 62 | 11 | 4 | +13 | -24 | -27 |
| Model | (a) No SGV | (b) SGV | (b) - (a) | ||||||||||
| Acc. | TPR | TNR | Bias | dSkew | Acc. | TPR | TNR | Bias | dSkew | Acc. | Bias | dSkew | |
| Gemini 2.0 | 66 | 99 | 56 | 33 | 33 | 74 | 97 | 67 | 25 | 24 | +8 | -9 | -9 |
| Gemini 2.5 Lite | 55 | 100 | 42 | 45 | 45 | 68 | 96 | 60 | 31 | 29 | +13 | -15 | -16 |
| Gemini 2.5 | 71 | 97 | 63 | 28 | 27 | 83 | 92 | 81 | 13 | 11 | +13 | -15 | -17 |
| Gemini 2.5 (T) | 78 | 95 | 73 | 20 | 18 | 87 | 91 | 86 | 8 | 5 | +9 | -12 | -14 |
| GPT-4.1 Mini | 60 | 99 | 49 | 40 | 39 | 75 | 99 | 69 | 24 | 24 | +16 | -16 | -16 |
| Base Agent | + SGV, outcome verifier | + SGV, verify every 5 steps | |
| All | 22 | 25 | 27 |
| Chrome | 29 | 30 | 35 |
| GIMP | 24 | 28 | 31 |
| LO Calc | 15 | 15 | 16 |
| LO Impress | 25 | 35 | 29 |
| LO Writer | 30 | 44 | 41 |
| Original | Improved Oracle | |
| Precision | 85.2 | 100 |
| TPR | 58.2 | 92 |
| TNR | 95.9 | 100 |
| Acc | 85.1 | 98 |
| Method | Precision |
| Rule-based (VWA Oracle, original) | 85 |
| Rule-based (VWA Oracle, ours) | 100 |
| No-SGV Baseline (Gemini 2.5, no Thinking) | 73 |
| WebJudge (GPT-o4, SOTA) | 75 |
| SGV (Gemini 2.5, no Thinking) | 80 |
| No-SGV Baseline (GPT-o4) | 80 |
| Metric | Before Calibration | Platt | SGV |
| ACC | 63 | 65 | 72 |
| TPR | 92 | 92 | 82 |
| TNR | 37 | 40 | 64 |
| Bias | 29 | 28 | 10 |
| dSkew | 25 | 23 | 4 |
| AUC | 65 | 65 | 74 |
| Score bucket | Before Calib. | Platt | Isotonic | SGV |
| 10% | 11% | 16% | 6% | |
| 2% | 10% | 0% | 8% | |
| 5% | 3% | 4% | 13% | |
| 4% | 1% | 4% | 5% | |
| 2% | 1% | 1% | 1% | |
| 2% | 74% | 74% | 3% |
| Metric | NoSGV temp=0 | NoSGV | Platt | Platt p15 | Platt 0.5 | SGV temp=0 | SGV | SGV |
| ACC | 65 | 65 | 72 | 69 | 71 | 76 | 75 | 78 |
| TPR | 92 | 93 | 77 | 85 | 86 | 84 | 85 | 79 |
| TNR | 41 | 40 | 68 | 56 | 57 | 71 | 67 | 78 |
| Bias | 27 | 28 | 6 | 20 | 16 | 11 | 11 | 1 |
| dSkew | 23 | 24 | 2 | 15 | 9 | 5 | 5 | 0 |
| AUC | 67 | 66 | 73 | 75 | 75 | 76 | 76 | 83 |
| Verification Mode | OSWorld | VisualWebArena |
| Baseline Agent | 21.7 | 45.0 |
| No SGV | ||
| + Outcome-based | 23.4 | 46.1 |
| + Every 5 Steps | 24.3 | 46.9 |
| SGV | ||
| + Outcome-based | 24.9 | 54.0 |
| Metric | All | 1–5 | 6–19 | 20 |
| % Success | 47 | 57 | 36 | 17 |
| N Samples | 2730 | 1587 | 1218 | 198 |
| % Total. | 100 | 58 | 45 | 7 |
| No-SGV | ||||
| ACC | 67 | 68 | 61 | 65 |
| TPR | 90 | 91 | 88 | 64 |
| Metric | All | 1–10 | 11–20 | 21–30 | 30 |
| % Success | 22 | 54 | 20 | 16 | 2 |
| N Samples | 1044 | 300 | 198 | 111 | 435 |
| % Total. | 100 | 29 | 19 | 11 | 42 |
| No-SGV | |||||
| ACC | 76 | 71 | 59 | 67 | 89 |
| TPR | 96 | 97 | 97 | 89 | 67 |
| Verifier | G-2.5 | G-2.5 | G-2.5 | G-2.0 | G-2.0 |
| Agent | G-2.5 | G-2.0 | GPT4o | G-2.5 | GPT4o |
| Acc | (65, 76) | (70, 78) | (68, 77) | (56, 65) | (55, 65) |
| TPR | (90, 84) | (84, 77) | (80, 76) | (93, 90) | (85, 74) |
| TNR | (47, 71) | (63, 79) | (63, 77) | (28, 42) | (42, 61) |
| Bias | (26, 11) | (20, 7) | (20, 8) | (38, 28) | (35, 19) |
| dSkew | (21, 5) | (14, 2) | (13, 3) | (35, 23) | (30, 12) |
| # | Method | Accuracy (%) | TPR (%) | TNR (%) | Bias (%) | dSkew (%) |
| 1 | Base MLLM Verifier | 65 | 90 | 47 | 26 | 21 |
| 2 | + First Step | 66 | 90 | 49 | 25 | 20 |
| 3 | SGV, Unified | 74 | 85 | 67 | 14 | 8 |
| 4 | SGV | 76 | 84 | 71 | 11 | 5 |
| Setting | No SGV Baseline | SGV Baseline | (1) | (2) | (3) | (4) | (5) |
| First Step | – | Q | GP | G (k=3) | G+Q+GP (k=3) | ||
| Verify | G | G | G | G | G | G | G |
| Acc | 65 | 76 | 77 | 76 | 76 | 78 | 79 |
| TPR | 90 | 84 | 84 | 84 | 84 | 87 | 82 |
| TNR | 47 | 71 | 71 | 70 | 70 | 70 | 76 |
| Bias | 26 | 11 | 11 | 11 | 11 | 12 | 6 |
| Setting | NoSGV Base | (1) | SGV Base | (2) | (3) | (4) | NoSGV Base (T) | (5) | SGV Base (T) | (6) |
| First Step | – | – | G | G + Web | G-Lite | G + Noise | – | G (T) + Web | G (T) | G |
| Verify | G | G + Web | G | G | G | G | G (T) | G (T) | G (T) | G (T) |
| Acc | 65 | 64 | 76 | 74 | 74 | 73 | 69 | 66 | 77 | 76 |
| TPR | 90 | 92 | 84 | 84 | 82 | 81 | 90 | 73 | 86 | 86 |
| TNR | 47 | 44 | 71 | 65 | 69 | 68 | 55 | 61 | 70 | 68 |
| Bias | 26 | 29 | 11 | 17 | 11 | 14 | 23 | 10 | 12 | 13 |
| Method | All VWA | Shopping | Classifieds | |
| Base Agent | 36 | 38 | 27 | 41 |
| + Verifier, no SGV | 36 | 39 | 27 | 40 |
| + Verifier, SGV | 40 | 41 | 29 | 46 |
| Method | Model | (a) Full Benchmark | (b) VWA Lite | (b-a) | ||||||
| All | Shop. | Class. | All | Shop. | Class. | All | ||||
| Search Agent Koh et al. [2024b] | GPT-4o | 27 | 29 | 21 | 30 | 28 | 29 | 22 | 30 | +0.3 |
| ReAct | Gemini 2.0 | 36 | 38 | 27 | 41 | 36 | 35 | 26 | 43 | -0.4 |
| ReAct + SGV | Gemini 2.0 | 40 | 41 | 29 | 46 | 40 | 40 | 31 | 47 | +0.2 |
| ReAct | Gemini 2.5 | 47 | 49 | 36 | 52 | 46 | 47 | 35 | 52 | -0.8 |
| ReAct + SGV | Gemini 2.5 | 54 | 58 | 44 | 56 | 54 | 57 | 42 | 59 | -0.1 |
| Speedup | Total Time | Avg / Task | |
| Original | |||
| Full 910 tasks | – | 2d:20h:09m | 04:31m |
| Full 910 tasks + Reset | – | 3d:18h:31m | 05:58m |
| 5 Environments in Parallel | |||
| Full 910 tasks * | 7 | 12h:13m | 00:50m |
| VisualWebArena-Lite * | 21 | 04h:04m | 00:50m |
| Method | All VWA | Classifieds | Shopping | |
| Base Agent - Before Refinements | 41 | 46 | 35 | 42 |
| Base Agent - After Refinements | 47 | 49 | 36 | 51 |
| Agent + SGV - Before Refinements | 48 | 53 | 42 | 49 |
| Agent + SGV - After Refinements | 54 | 56 | 43 | 58 |
LLM Agent Verification
LLM agent verification checks whether agents’ actions, evidence, policies, and completion claims satisfy explicit requirements.
See all 161 papers