The Temporal Tug-of-War: Visualizing and Detecting RAG Conflicts in Diffusion Models via Trajectory Variance

Sravan Karthick T; Pranav Darshan; Pranav A; Minal Moharir; Ivan P. Yamshchikov

cs.CLSep 16, 2026

The Temporal Tug-of-War: Visualizing and Detecting RAG Conflicts in Diffusion Models via Trajectory Variance

Authors: Sravan Karthick T, Pranav Darshan, Pranav A, Minal Moharir, Ivan P. Yamshchikov

Organizations: Department of Computer Science and Engineering, R.V. College of Engineering, India · CAIRO, Technical University of Applied Sciences Würzburg-Schweinfurt, Germany

Abstract

Retrieval-Augmented Generation (RAG) introduces a specific failure mode in discrete diffusion language models: when retrieved context contradicts parametric knowledge, the iterative denoising process becomes a visible battleground between competing knowledge sources. We identify temporal semantic divergence as an observable for detecting these conflicts and introduce the Trajectory Variance Score (TVS), a simple and interpretable measure of this divergence. TVS computes the mean pairwise cosine distance of answer embeddings across independent stochastic denoising trajectories, capturing the temporal tug of war between parametric and contextual attractors. Requiring as few as two parallel inference runs, TVS is computationally lightweight. Across four diverse datasets (Synthetic, SciQ, PopQA, and CounterFact), a simple Logistic Regression classifier using TVS achieves $70.10\%$ accuracy and $0.7647$ AUROC on LLaDA. On Dream 7B, increasing the number of trajectories from two to five improves accuracy from $63.91\%$ to $69.62\%$ . More complex sequential models provide only marginal improvements over the linear classifier. Evaluation across LLaDA and Dream 7B demonstrates that conflict-induced trajectory dynamics and their key properties transfer across distinct diffusion architectures.

Figures & tables

Figure 1: Complete evaluation pipeline: facts undergo parametric memory filtering, conflict contexts are injected, $N=2$ independent denoising trajectories are executed, the mean pairwise cosine distance (TVS) is computed at each of $T=50$ timesteps to produce a feature vector, and the resulting vector is classified by a Logistic Regression detector.

Figure 2: Average TVS across the 50 step denoising trajectory for LLaDA (Clean: blue solid; RAG Conflict: red dashed; shaded = SEM). Three phases are visible: Zone 1 ( $t\in[0,15]$ ): Early Exploration—both conditions exhibit high dispersion. Zone 2 ( $t\in[15,28]$ ): Mid Step Friction—the clean curve decays while the conflict curve remains elevated (gap $\approx 0.15$ ). Zone 3 ( $t\in[28,50]$ ): Delayed Crystallization—both curves collapse, but a residual gap persists ( $\Delta\approx 0.07$ ).

Figure 3: Average TVS across the 50 step denoising trajectory for Dream 7B ( $N=2$ seeds; Clean: blue solid; RAG Conflict: red dashed; shaded = SEM). The same three phase structure is observed, but with a narrower clean versus conflict gap, consistent with Dream’s smoother remasking dynamics based on entropy.

Figure 4: Detailed TVS dynamics for LLaDA with $\pm$ std shading ( $N=2$ seeds). The conflict curve (red) starts at $\approx 0.51$ versus clean (green) at $\approx 0.39$ , maintaining a consistent gap through $t\approx 28$ before both converge—with conflict retaining a higher residual dispersion ( $\approx 0.07$ versus $\approx 0.01$ ).

Method	LLaDA Acc (%)	LLaDA AUROC	Dream Acc (%)	Dream AUROC
TraceDet*	$71.34\pm 0.66$	$0.7865\pm 0.0078$	$64.56\pm 0.56$	$0.6876\pm 0.0056$
TVS-LR (Ours)	$70.10\pm 1.27$	$0.7647\pm 0.0125$	$63.29\pm 1.21$	$0.6697\pm 0.0071$

Table 1: Overall conflict detection performance (mean

\pm

std over 5 random splits). TVS achieves competitive accuracy on both models. The

\sim

7% gap on Dream likely reflects its smoother remasking dynamics based on entropy. *TraceDet codebase is not publicly available; we benchmark against an unverified best effort replication for context.

Metric	LLaDA (%)	Dream (%)
Precision	$71.55\pm 1.97$	$65.19\pm 1.05$
Recall	$66.86\pm 0.91$	$56.95\pm 2.80$
F1-Score	$69.11\pm 0.90$	$60.77\pm 1.89$

Table 2: Classification metrics for Logistic Regression (mean

\pm

std over 5 random splits). Dream’s lower recall (

56.95\%

) indicates that its smoother trajectories make conflict samples harder to distinguish from clean ones.

Model	Synth.	SciQ	PopQA	CounterFact
LLaDA	$76.92\pm 5.11$	$73.74\pm 1.53$	$70.34\pm 2.41$	$61.54\pm 4.20$
Dream	$64.17\pm 5.12$	$65.71\pm 1.82$	$58.76\pm 4.37$	$54.69\pm 3.50$

Table 3: Detection accuracy (%) by domain (LR classifier, mean

\pm

std over 5 splits). The ordering of difficulty across datasets (Synth.

>

SciQ

>

PopQA

>

CounterFact) is preserved across both architectures.

Model	Clean Context	RAG Conflict	$\Delta$ (Elevation)
LLaDA 8B	$17.84$	$22.46\pm 0.42$	$+4.62$
Dream 7B	$2.66$	$3.71\pm 0.30$	$+1.05$

Table 4: Entity Flip Rate across architectures under clean and conflicting RAG contexts. Both models exhibit a statistically significant elevation in semantic flips when subjected to a conflict. LLaDA’s confidence-based remasking results in significantly higher baseline volatility and conflict oscillation compared to Dream’s smoother entropy-based approach.

Classifier	LLaDA Acc (%)	LLaDA AUROC	Dream Acc (%)	Dream AUROC
Logistic Regression	$71.05\pm 0.72$	$0.7701\pm 0.0098$	$63.40\pm 1.30$	$0.6972\pm 0.0072$
LSTM (Unidirectional)	$71.75\pm 0.54$	$0.7895\pm 0.0061$	$62.87\pm 1.14$	$0.6713\pm 0.0069$
Bidirectional LSTM + Attention	$71.93\pm 0.71$	$0.7906\pm 0.0062$	$64.67\pm 0.70$	$0.6977\pm 0.0069$

Table 5: Architecture comparison over 5 random splits. On both models, all confidence intervals overlap substantially. The gap between LR and the Bidirectional LSTM is

0.88\%

on LLaDA and

1.27\%

on Dream, confirming that sequential modeling adds negligible benefit across architectures.

$N$ Seeds	LLaDA Acc (%)	LLaDA AUROC	Dream Acc (%)	Dream AUROC
2	$70.76\pm 0.62$	$0.7725\pm 0.0082$	$63.91\pm 1.17$	$0.6893\pm 0.0119$
3	$71.75\pm 1.06$	$0.7806\pm 0.0084$	$67.12\pm 0.95$	$0.7261\pm 0.0131$
5	$71.23\pm 1.00$	$0.7851\pm 0.0074$	$69.62\pm 0.51$	$0.7589\pm 0.0134$

Table 6: Effect of seed count on detection performance (LR classifier, 5 random splits). LLaDA’s accuracy is stable across seed counts (

\Delta=+0.47\%

), while Dream benefits substantially from additional seeds (

\Delta=+5.71\%

), suggesting the optimal seed count is dependent on architecture.

Knowledge Conflicts in Language Models

Knowledge conflicts in language models concern how models reconcile stored knowledge with conflicting context, retrieved evidence, tools, or memory.

See all 142 papers

ConflictRAG: Detecting and Resolving Knowledge Conflicts in Retrieval Augmented Generation
May 17, 2026Chenyu Wang, Yueyuan Li, Yingmin Liu +1Retrieval-Augmented Generation Knowledge Conflicts in Language Models
School of Computer and Artificial Intelligence, Zhengzhou University, Zhengzhou, China · College of Environment and Ecology, Taiyuan University of Technology, Taiyuan, China · Zhejiang University, Hangzhou, China
EvoTrustRAG: Evolution-Aware Conflict Attribution and Evidence Handling for Reliable Retrieval-Augmented Generation
Aug 8, 2026Xi Nie, Hongwei Li, Shenghao Wu +3Retrieval-Augmented Generation Knowledge Conflicts in Language Models
School of Computer Science and Engineering, University of Electronic Science and Technology of China · Institute of Information Engineering, Chinese Academy of Sciences
Does RAG Know When Retrieval Is Wrong? Diagnosing Context Compliance under Knowledge Conflict
May 14, 2026Yihang Chen, Pin Qian, Su Wang +4Retrieval-Augmented Generation Knowledge Conflicts in Language Models
1Georgia Institute of Technology · 2Carnegie Mellon University · University of California San Diego

May 17, 2026J/K move · Enter open · S save

ConflictRAG: Detecting and Resolving Knowledge Conflicts in Retrieval Augmented Generation

Chenyu Wang, Yueyuan Li, Yingmin Liu, Yang Shu

School of Computer and Artificial Intelligence, Zhengzhou University, Zhengzhou, China · College of Environment and Ecology, Taiyuan University of Technology, Taiyuan, China · Zhejiang University, Hangzhou, China

Retrieval-Augmented Generation Knowledge Conflicts in Language Models RAG Evaluation

Open paper AskPDF arXiv

Retrieval-Augmented Generation (RAG) systems implicitly assume mutual consistency among retrieved documents -- an assumption that frequently fails in practice. We present ConflictRAG, a conflict-aware RAG framework that detects, classifies, and resolves knowledge conflicts prior to answer generation. The framework introduces three contributions: (1) a two-stage conflict detection module combining a lightweight embedding-based MLP classifier with selective LLM refinement, reducing API costs by 62% while maintaining 90.8% detection accuracy; (2) an Entropy-TOPSIS framework for data-driven source credibility assessment, improving selection accuracy by 7.1% over manual heuristics; and (3) a Conflict-Aware RAG Score (CARS) for diagnostic evaluation of conflict-handling capabilities. Experiments on three benchmarks against six baselines demonstrate 88.7% conflict-detection F1 and consistent 5.3--6.1% correctness gains over the strongest conflict-aware baseline, with the pipeline transferring effectively across backbone LLMs.

Abstract

Knowledge Conflicts in Language Models

Explore similar work

ConflictRAG: Detecting and Resolving Knowledge Conflicts in Retrieval Augmented Generation

EvoTrustRAG: Evolution-Aware Conflict Attribution and Evidence Handling for Reliable Retrieval-Augmented Generation

Does RAG Know When Retrieval Is Wrong? Diagnosing Context Compliance under Knowledge Conflict