Multi-Channel Mitigation of Source-Trust Shortcuts in Fact-Checking RL Agents

Jianchang Su; Yiwei Yang; Wei Zhang

cs.AISep 29, 2026

Multi-Channel Mitigation of Source-Trust Shortcuts in Fact-Checking RL Agents

Authors: Jianchang Su, Yiwei Yang, Wei Zhang

Organizations: University of Connecticut · University of California, Santa Cruz

Abstract

Retrieval-augmented fact-checkers often receive a reliability label, such as HIGH or LOW trust, for each evidence source. These labels should adjust the model's confidence and its decision to search for more evidence, while the verdict should follow the evidence content. We introduce TrustSwap, a counterfactual test that swaps, lowers, or removes source labels while keeping every evidence text fixed, and measures its three output channels (the verdict, the confidence, and the search decision) separately. Across untrained and RL-trained models at two scales, three datasets, and two prompts, confidence and search respond to the labels as intended in 49 of 50 comparisons, yet a label change alone alters 4-23% of confident verdicts for Qwen3 models and up to 50% for an existing RL-trained fact-checker. Standard GRPO fine-tuning amplifies this shortcut at 8B in all six settings. To reduce it, we propose trust-swap augmentation (TSA), which trains GRPO on each claim with both its original and its label-swapped evidence under the same gold verdict. At 4B, TSA lowers the verdict flip rate by 7-35% (relative) in four of six settings, keeps accuracy and the intended confidence and search responses, outperforms reward-based alternatives in the main setting, and carries over to an unseen label-removal perturbation. An added consistency reward helps on the trained-on swap but not on unseen perturbations. At 8B, TSA's effect is not detectable, which makes scale the main open question.

Figures & tables

Figure 1: TrustSwap and trust-swap augmentation (TSA). (1) Each claim is paired with its evidence pack

E

and a trust-swapped pack

\phi(E)

(HIGH

\leftrightarrow

LOW; texts and sources unchanged); TrustSwap compares the outputs on the two. (2) TSA trains GRPO on both packs as separate rollout groups with the same gold verdict

v^{\star}

; an optional consistency reward

R_{\text{cons}}

links matched rollouts (

v_{o}

v_{f}

: verdicts on

E

and

\phi(E)

;

g

: gold). (3) TSA uses

\lambda=0

. Trust should change the confidence

p

and the search decision

m

, but not the verdict

v

Claim	Evidence (source: label before $\rightarrow$ after the swap)	Original	Swapped
(a) Typical case: labels lowered on a high-trust pack. Confidence and search respond as intended, but the verdict also changes.
ICE removed thousands of gang members.	[1] ice.gov : HIGH $\rightarrow$ LOW. “5,396.” [2] ice.gov : HIGH $\rightarrow$ LOW. “5,872.”	SUPPORTS conf. 0.95 search: no	NEI conf. 0.3–0.5 search: yes
(b) Reversal: only one label is raised, and the verdict switches polarity with unchanged high confidence.
Social distancing […] was not practiced at the funeral of civil rights icon John Lewis.	[1] perma.cc : UNKNOWN. “…the church will only accommodate 240 people in total, due to social distancing precautions.” [2] facebook.com : LOW $\rightarrow$ HIGH. “Yes. The church’s senior pastor … referred to social distancing …” [3] gettyimages.com : UNKNOWN. “Yes. Photo by …”	SUPPORTS conf. 0.95 search: no	REFUTES conf. 0.95 search: no

Figure 2: The label swap on two AVeriTeC claims. Only the reliability labels change; every evidence text and source name is identical before and after. Outputs are from the 4B GRPO baseline, and both verdict changes occur in all five training seeds (the confidence range in (a) spans seeds). Under our contract, the confidence drop and the search request in (a) are correct, but neither verdict change is. The gold label is SUPPORTS in (a) and REFUTES in (b).

Figure 3: TrustSwap label-swap results. (a)–(c) Main setting (AVeriTeC, neutral prompt): (a) flip rate, split into reversals and switches to or from NEI; (b, c) mean change in confidence and search requests where labels are lowered (dark) or raised (light), both in the intended direction. (d) All six settings: flip rate before and after GRPO fine-tuning (mean

\pm

s.d., five seeds); points above the diagonal mean that training increased flips.

Figure 4: Change in flip rate relative to the 4B GRPO baseline (main setting) under the label swap used in training and under two label perturbations never seen in training. Filled markers: 95% interval excludes zero. TSA’s gain carries over to label removal and keeps its direction under the downshift.

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Setting	Untrained	GRPO baseline	TSA	TSA + cons.	$\Delta$ TSA	$\Delta$ TSA + cons.
AVeriTeC N	10.5	$15.0\pm 1.2$	$14.7\pm 1.6$	$14.3\pm 1.3$	$-$ 0.36	$-$ 0.74
AVeriTeC P	15.8	$22.6\pm 0.4$	$22.6\pm 0.8$	$22.3\pm 0.8$	$+$ 0.05	$-$ 0.30
SciFact N	6.2	$12.7\pm 1.6$	$13.1\pm 1.1$	$13.4\pm 1.0$	$+$ 0.33	$+$ 0.64
SciFact P	7.2	$10.4\pm 0.9$	$10.7\pm 2.2$	$10.2\pm 2.4$	$+$ 0.30	$-$ 0.18
CONFACT N	4.9	$6.7\pm 1.0$	$6.8\pm 0.5$	$7.1\pm 0.7$	$+$ 0.07	$+$ 0.39
CONFACT P	10.2	$14.1\pm 1.3$	$14.7\pm 0.9$	$13.5\pm 0.6$	$+$ 0.57	$-$ 0.62

Appendix

Table 10: Qwen3-8B in all six settings (selected recipe). Flip rates in %;

\Delta

columns are matched-seed differences in pp. GRPO training raises the flip rate above the untrained model in every setting (all five seeds). Neither intervention changes the flip rate significantly in any setting. The only significant method-versus-method contrast is TSA + consistency minus TSA on CONFACT with the principled prompt (

-1.19

[-1.90,-0.43]

), a different setting from the one significant contrast at 4B; we treat it as exploratory.

Explore similar work

OpenFC: Learning Verification Policies towards Open-Search Fact Checking
Sep 27, 2026Xinming Wang, Kaixiang Qiu, Yansong Lin +5Fact-Checking
Institute of Automation, Chinese Academy of Sciences · University of Electronic Science and Technology of China · Beijing Institute of Technology +1
Can We Trust the Judges? Validation of Factuality Evaluation Methods via Answer Perturbation
Sep 14, 2026Sarra Gharsallah, Adele Robaldo, Mariia Tokareva +5Judges Validation
Data Science Department, EURECOM, Campus SophiaTech, Biot, France
Althea: The Fact-Checking--Metalearning Tradeoff in AI-Assisted Verification
Dec 29, 2025Svetlana Churina, Kokil Jaidka, Anab Maulana Barik +5Fact-Checking Category-Aware Atomic Claims
Centre for Trusted Internet & Community (CTIC), National University of Singapore (NUS), Singapore · Department of Communications and New Media, Centre for Trusted Internet & Community (CTIC), National University of Singapore (NUS), Singapore · School of Computing, Centre for Trusted Internet & Community (CTIC), National University of Singapore (NUS), Singapore

Sep 27, 2026, cs.AIJ/K move · Enter open · S save

OpenFC: Learning Verification Policies towards Open-Search Fact Checking

Xinming Wang, Kaixiang Qiu, Yansong Lin, Chunji Lv+4

Institute of Automation, Chinese Academy of Sciences · University of Electronic Science and Technology of China · Beijing Institute of Technology+1

Fact-Checking

Open paper AskPDF arXiv

Open-search fact checking is not merely retrieval followed by classification, but a sequential decision problem in which every query, source visit, and stopping decision reshapes the evidence available for verification. Yet existing systems often distribute these decisions across predefined pipelines or separately prompted modules rather than learning them as a unified task-specific policy. We introduce \textbf{OpenFC}, a unified verification-policy training framework that post-trains Qwen3-8B as a compact next-action controller over reasoning, evidence acquisition, and stopping. OpenFC learns this policy in two stages. \textbf{Stepwise-Calibrated Cold Start (SCCS)} uses a strong training-time supervisor to review post-initial reasoning, tool-use, and stopping proposals before execution, producing reliable trajectories for supervised fine-tuning without access to gold verdicts. \textbf{Verification-Aware Reinforcement Learning (VA-RL)} then improves the cold-start policy on unresolved claims through budget-aware tool rewards, label-aware advantage reweighting, and localized response masking. Across six fact-checking benchmarks, OpenFC achieves 70.39% average accuracy and 63.30% macro-F1, the highest overall averages among the evaluated methods. Stage-wise ablations further show that SCCS and VA-RL provide complementary gains, supporting the design of the two-stage training framework. These results position OpenFC as a strong and effective framework for open-search fact-checking. We will open-source our code and release the model checkpoints to support reproducibility.

Abstract

Explore similar work

OpenFC: Learning Verification Policies towards Open-Search Fact Checking

Can We Trust the Judges? Validation of Factuality Evaluation Methods via Answer Perturbation

Althea: The Fact-Checking--Metalearning Tradeoff in AI-Assisted Verification