BiMind: A Dual-Head Reasoning Model with Attention-Geometry Adapter for Incorrect Information Detection

Zhongxing Zhang; Emily K. Vraga; Jisu Huh; Jaideep Srivastava

cs.CLApr 7, 2026

BiMind: A Dual-Head Reasoning Model with Attention-Geometry Adapter for Incorrect Information Detection

Authors: Zhongxing Zhang, Emily K. Vraga, Jisu Huh, Jaideep Srivastava

Organizations: University of Minnesota, Twin Cities

Abstract

Incorrect information poses significant challenges by disrupting content veracity and integrity, yet most detection approaches struggle to jointly balance textual content verification with external knowledge modification under collapsed attention geometries. To address this issue, we propose a dual-head reasoning framework, BiMind, which disentangles content-internal reasoning from knowledge-augmented reasoning. In BiMind, we introduce three core innovations: (i) an attention geometry adapter that reshapes attention logits via token-conditioned offsets and mitigates attention collapse; (ii) a self-retrieval knowledge mechanism, which constructs an in-domain semantic memory through kNN retrieval and injects retrieved neighbors via feature-wise linear modulation; (iii) the uncertainty-aware fusion strategies, including entropy-gated fusion and a trainable agreement head, stabilized by a symmetric Kullback-Leibler agreement regularizer. To quantify the knowledge contributions, we define a novel metric, Value-of-eXperience (VoX), to measure instance-wise logit gains from knowledge-augmented reasoning. Experiment results on public datasets demonstrate that our BiMind model outperforms advanced detection approaches and provides interpretable diagnostics on when and why knowledge matters.

Figures & tables

Figure 1: Illustrative case of self-correction via knowledge: without knowledge (left), the content-internal head labels Lafufu (counterfeit Labubu) as Labubu; with knowledge (right), the knowledge-augmented head corrects the label as Lafufu.

Figure 2: An illustration of our proposed BiMind framework. (a) Dual-head architecture with a content-internal head (top) and a knowledge-augmented head (bottom). (b) Attention geometry adapter reshapes pre-softmax attention logits via token-conditioned offsets. (c) Self-retrieval knowledge module retrieves top- $k$ neighbors and injects knowledge via FiLM to provide knowledge-augmented representations. (d) Uncertainty-aware fusion combines head logits via an entropy gating, with the VoX metric quantifying the knowledge contributions by comparing head outcomes, where blue and green surfaces in the instance space represent the content and knowledge reasoning heads, respectively.

Figure 3: Illustration of the attention maps with ( $+$ ) and without AGA across shallow (Layer 0) and deep (Layer 9) layers. AGA induces a structured, token-conditioned attention pattern at deeper layers, reshaping attention distribution by injecting global, key-focused inductive bias.

Dataset	CNN		GCN		BERT		HAN		HeteroSGT		BiMind
	Acc	Pre	Acc	Pre	Acc	Pre	Acc	Pre	Acc	Pre	Acc	Pre
MM COVID	0.582±0.035	0.478±0.170	0.717±0.156	0.735±0.236	0.730±0.093	0.727±0.094	0.855±0.005	0.854±0.005	0.915±0.009	0.905±0.011	0.951±0.008	0.950±0.011
ReCOVery	0.658±0.011	0.460±0.104	0.718±0.037	0.691±0.178	0.682±0.030	0.441±0.213	0.722±0.021	0.462±0.197	0.727±0.023	0.731±0.047	0.918±0.013	0.922±0.013
MC Fake	0.825±0.001	0.544±0.156	0.724±0.138	0.516±0.169	0.827±0.006	0.713±0.271	0.825±0.005	0.463±0.098	0.883±0.002	0.812±0.003	0.887±0.005	0.827±0.006
LIAR	0.546±0.019	0.432±0.181	0.487±0.039	0.493±0.047	0.537±0.007	0.513±0.017	0.546±0.025	0.493±0.036	0.581±0.002	0.580±0.003	0.633±0.001	0.637±0.002
Dataset	Rec	F1	Rec	F1	Rec	F1	Rec	F1	Rec	F1	Rec	F1

Table 1: Detection performance on four datasets (best in red , second-best in blue ).

Dataset	Head/Mode	Acc	F1	Pre	Rec	VoX (mean / pos%)	Gate Mean (% $<$ 0.3 / % $>$ 0.7)
ReCOVery	$\text{Head}_{0}$	85.22	85.15	85.10	85.22	–	–
	$\text{Head}_{E}$	87.19	86.93	86.98	87.19	+0.47 / 84.24%	0.04 (100.00% / 0.00%)
	Fused	87.19	86.93	86.98	87.19	–	0.04 (100.00% / 0.00%)
MC Fake	$\text{Head}_{0}$	86.50	84.85	84.79	86.50	–	–
	$\text{Head}_{E}$	87.35	86.90	86.63	87.35	-0.08 / 39.91%	0.22 (81.24% / 0.00%)
	Fused	87.42	86.82	86.52	87.42	–	0.22 (81.24% / 0.00%)

Table 2: Cross-dataset VoX results. We report mean VoX value and percentage of samples with positive gain (pos%); we also show mean gate value and routing mass below/above thresholds, with %

<

0.3 indicating the gate leaned strongly toward knowledge head

\text{Head}_{E}

, and %

>

0.7 indicating it leaned strongly toward content head

\text{Head}_{0}

Model Variant	Acc	Pre	Rec	F1
Full BiMind model (w/ knowledge)	0.897	0.895	0.897	0.895
Baseline (content w/o knowledge)	0.852	0.849	0.852	0.848
– Attention geometry adapter	0.872	0.870	0.872	0.870
– Knowledge retrieval	0.847	0.847	0.847	0.847
– Gated fusion	0.862	0.861	0.862	0.861
– Trainable agreement head	0.867	0.868	0.867	0.867

Table 3: Ablation study on ReCOVery dataset. One module is removed at a time from the full BiMind model.

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Model	Model Size	Trainable Params	Ratio
BiMind-Custom (Tiny)	3.5M	3.5M	100%
BiMind-RoBERTa (Medium)	125M	3.3M	2.6%
BiMind-LLaMA (Large)	7B	44.4M	$<$ 1%

Appendix

Table 4: Parameter efficiency across BiMind variants.

Dataset	# Label 0	# Label 1	# Total	Avg. Length (words)
MM COVID	1,888	1,162	3,048	25
RoCOVery	605	1,294	1,899	500
LIAR	2,507	2,053	4,560	17
MC Fake	2,671	12,621	15,292	300

Appendix

Table 5: Statistics of the datasets used in our experiments.

Dataset	Vocab Alignment (%)	Max Sim.	Mean Sim.	Flesch (helps)
LIAR	76.55	0.6603±0.0924	0.6077±0.0843	37.94
MM COVID	75.60	0.7586±0.1228	0.6870±0.1082	31.68
MC Fake	79.91	0.8176±0.1084	0.7751±0.1082	-280.47
ReCOVery	81.68	0.8056±0.1071	0.7567±0.1026	-585.00

Appendix

Table 6: Statistical comparison of knowledge attributes across datasets. Vocabulary alignment measures lexical intersection between inputs and the knowledge base. Retrieval relevance is reported as cosine similarity (mean

\pm

std). Flesch Reading Ease for the helps category reflects the linguistic complexity of retrieved knowledge.

Metric	Content Head	Knowledge Head	Fused
Accuracy	68.2%	61.5% (-6.7%)	70.1% (+1.9%)
F1-Score	0.65	0.58	0.68

Appendix

Table 7: Empirical results on fusion strategy.

Knowledge Impact	% of Test Set	Interpretation
Helps (wrong $\rightarrow$ correct)	$\sim$ 15%	Genuine knowledge benefit
Hurts (correct $\rightarrow$ wrong)	$\sim$ 8%	Knowledge introduces noise
Neutral / Both Wrong	$\sim$ 77%	No leaked advantage

Appendix

Table 8: Knowledge impact on data leakage.

Model	Dataset	Head/Mode	Acc	F1	Pre	Rec
LLaMA-7B	ReCOVery	$\text{Head}_{0}$	91.13 $\pm$ 2.34	91.03 $\pm$ 2.59	91.35 $\pm$ 2.32	91.13 $\pm$ 2.34
		$\text{Head}_{E}$	91.04 $\pm$ 1.18	91.09 $\pm$ 1.21	91.53 $\pm$ 1.44	91.04 $\pm$ 1.18
		Fused	91.82 $\pm$ 1.29	91.86 $\pm$ 1.27	92.20 $\pm$ 1.26	91.82 $\pm$ 1.29
	MM COVID	$\text{Head}_{0}$	94.69 $\pm$ 1.01	94.70 $\pm$ 1.01	94.46 $\pm$ 1.45	94.67 $\pm$ 0.99
		$\text{Head}_{E}$	94.80 $\pm$ 0.77	94.81 $\pm$ 0.76	94.48 $\pm$ 1.21	94.61 $\pm$ 0.64
		Fused	95.12 $\pm$ 0.78	95.12 $\pm$ 0.78	94.98 $\pm$ 1.06	95.08 $\pm$ 0.86

Appendix

Table 9: Extended performance comparison (mean

\pm

std, in %, across 10 runs, best in bold ).

Dataset	Model	Acc	Pre	Rec	F1	Training	Testing	Graph
MM COVID	HeteroSGT	0.915±0.009	0.905±0.011	0.883±0.013	0.893±0.011	55.11	–	13.62
MM COVID	BiMind	0.902±0.116	0.902±0.110	0.898±0.142	0.900±0.132	15.19	0.08	–
ReCOVery	HeteroSGT	0.727±0.023	0.731±0.047	0.585±0.036	0.571±0.049	21.94	–	9.03
ReCOVery	BiMind	0.879±0.017	0.862±0.028	0.843±0.203	0.854±0.208	14.99	0.10	–
MC Fake	HeteroSGT	0.883±0.002	0.812±0.003	0.762±0.002	0.783±0.003	478.53	–	40.04
MC Fake	BiMind	0.887±0.051	0.827±0.016	0.700±0.099	0.798±0.109	153.01	0.45	–

Appendix

Table 10: Performance and runtime comparison between BiMind and HeteroSGT across four benchmark datasets. Performance is reported as mean

\pm

std. Runtime is measured in seconds per run.

Method	Accuracy	F1	Improvement
kNN (k=5)	83.74%	83.54%	-
kNN-BERT (k=5)	83.74%	83.54%	-
Standard RAG (k=3)	83.25%	83.00%	-0.49%
BiMind (Ours)	87.70%	87.90%	+3.96%

Appendix

Table 11: Comparison with retrieval-based baselines on ReCOVery.

Figure 4: An illustration of knowledge impacts at the instance level. Blue curve $=$ baseline (no knowledge), red dashed curve $=$ with knowledge, shaded regions indicate gain (green) or loss (orange).

Figure 5: Layer 9 attention head specialization with and without AGA. Left: Baseline Transformer without AGA. Right: Transformer with AGA.

Fake News Detection

Fake news detection identifies false or misleading claims by analyzing content, context, evidence, and how information spreads.

See all 98 papers

Memory-First Fact-Checking: A Knowledge-Graph-Grounded Multi-Agent System for Misinformation Detection
Aug 30, 2026Amelia Petrenciuc, Alexandru Lecu, Adrian GrozaAutomated Fact-Checking Fake News Detection
CheckRLM: Effective Knowledge-Thought Coherence Checking in Retrieval-Augmented Reasoning
Jul 2, 2026Dingling Xu, Ruobing Wang, Qingfei Zhao +8Retrieval-Augmented Reasoning LLM Reliability
School of Artificial Intelligence, Beijing Normal University, Beijing 100875, China · Beijing Key Laboratory of Artificial Intelligence for Education, Beijing 100875, China · Institute of Information Engineering, Chinese Academy of Sciences +3
MM-VeriAgent: Learning to Use Extensive Tools to Verify Multimodal Misinformation with Reinforcement Learning
Sep 25, 2026Peipei Li, Shuhan Xia, Shengyang Liu +2Multimodal Deception Detection Multimodal Tool Use
Beijing University of Posts and Telecommunications · Minzu University of China · NLPR, Institute of automation, Chinese academy of science, Chinese Academy of Sciences

Aug 30, 2026J/K move · Enter open · S save

Memory-First Fact-Checking: A Knowledge-Graph-Grounded Multi-Agent System for Misinformation Detection

Amelia Petrenciuc, Alexandru Lecu, Adrian Groza

Automated Fact-Checking Fake News Detection Multi-Agent Systems

Open paper AskPDF arXiv

This paper introduces a hybrid fact-checking framework that integrates Knowledge Graph-based semantic memory with adversarial multi-agent reasoning for explainable misinformation detection. The proposed system follows a memory-first, web-fallback architecture, in which input claims are initially evaluated against a dual-index Knowledge Graph through Sentence-BERT-based semantic retrieval and Natural Language Inference. When the evidence retrieved from the graph is insufficient to support a reliable decision, the framework collects information from trusted web sources and assesses it using an adversarial tribunal composed of support, contradiction, and judging agents. A graph-aware confidence mechanism combines semantic similarity, NLI confidence, and structural graph evidence to determine whether internal knowledge is sufficient, thereby reducing unnecessary web retrieval. Following verification, validated information is transformed into structured triples and incorporated into the Knowledge Graph, supporting the incremental expansion of the system's semantic memory. Experimental evaluation on a curated COVID-19 misinformation benchmark demonstrates that the proposed framework achieves an accuracy of 97.4% and a macro-averaged F1-score of 92.6% on resolved claims, outperforming a Llama~~3.3~~70B baseline, which obtains an accuracy of 87.7% and a macro-averaged F1-score of 86.3%.

Abstract

Fake News Detection

Explore similar work

Memory-First Fact-Checking: A Knowledge-Graph-Grounded Multi-Agent System for Misinformation Detection

CheckRLM: Effective Knowledge-Thought Coherence Checking in Retrieval-Augmented Reasoning

MM-VeriAgent: Learning to Use Extensive Tools to Verify Multimodal Misinformation with Reinforcement Learning