TRACESTEER: Trace Driven Activation Steering for Hallucination Mitigation in Diffusion Language Models

Xin Zhang; Yili Wang; Yue Tan; Xin He; Rui Miao; Yixin Liu; Yi Chang; Xin Wang

cs.CLOct 13, 2026

TRACESTEER: Trace Driven Activation Steering for Hallucination Mitigation in Diffusion Language Models

Authors: Xin Zhang, Yili Wang, Yue Tan, Xin He, Rui Miao, Yixin Liu, Yi Chang, Xin Wang

Organizations: Jilin University, China · Griffith University, Australia

Abstract

Diffusion large language models (dLLMs) have emerged as a promising alternative to autoregressive LLMs, yet remain susceptible to hallucination. Activation steering offers a promising mitigation approach by guiding parallel token predictions through internal representations. However, existing steering methods largely overlook two characteristics of dLLMs: dynamic denoising states and temporal hallucination risks, which challenge predefined steering vectors and uniform intervention strengths. We propose TRACESTEER, a trace-driven activation steering framework that uses preceding denoising predictions as feedback to jointly adapt steering direction and strength. Specifically, Denoising Feedback Steering (DF-Steering) constructs step-specific steering vectors from preceding predictions and current hidden states. Trajectory-Guided Steering Gate (TS-Gate) regulates their strength using prediction uncertainty and the denoising timestep. With the backbone frozen, factual supervision and magnitude regularization jointly guide the learning of adaptive steering interventions, promoting factual reliability and stability throughout denoising. Experiments on two dLLMs show improved factual reliability on TruthfulQA and generalization across five downstream benchmarks.

Figures & tables

Figure 1: Motivation for trace-driven activation steering in dLLMs. Existing methods in LLMs suffer from (a) reliance on external knowledge, (b) incompatibility with dLLMs’ iterative denoising process, and (c) inability to adapt fixed steering vectors to evolving generation states. In contrast, (d) our approach enables dynamic, trace-aware activation steering.

Figure 2: Overview of TraceSteer . DF-Steering generates a feedback-guided steering vector from generation hypotheses and tentative predictions, while TS-Gate controls its strength based on the denoising timestep and feedback uncertainty. The gated steering vector is injected into dLLMs to mitigate hallucinations during denoising.

Method	Multiple-Choice				Open-ended Generation
Method	MC1 ( $\uparrow$ )	MC2 ( $\uparrow$ )	MC3 ( $\uparrow$ )	Avg. ( $\uparrow$ )	TRUE ( $\uparrow$ )	INFO ( $\uparrow$ )	TRUE $\times$ INFO ( $\uparrow$ )
LLaDA-8B-Instruct
Vanilla	22.13	47.08	23.45	30.89	49.18	81.96	40.31
Few-shot ICL	30.32 $\uparrow$ 8.19	46.35 $\downarrow$ 0.73	24.60 $\uparrow$ 1.15	33.76 $\uparrow$ 2.87	45.08 $\downarrow$ 4.10	79.50 $\downarrow$ 2.46	35.84 $\downarrow$ 4.47
Oscar	23.77 $\uparrow$ 1.64	46.97 $\downarrow$ 0.11	23.66 $\uparrow$ 0.21	31.47 $\uparrow$ 0.58	51.18 $\uparrow$ 2.00	84.32 $\uparrow$ 2.36	43.15 $\uparrow$ 2.84
CAA	23.77 $\uparrow$ 1.64	49.79 $\uparrow$ 2.71	24.52 $\uparrow$ 1.07	32.69 $\uparrow$ 1.80	50.00 $\uparrow$ 0.82	84.42 $\uparrow$ 2.46	42.21 $\uparrow$ 1.90

Table 1: Performance comparison on TruthfulQA. Colored values indicate changes relative to the vanilla baseline. Best results are shown in bold , and second-best results are underlined .

Method		COPA	StoryCloze	MMLU	Wino	BoolQA	Avg.
LLaDA-8B-Instruct	Vanilla	75.40	70.49	63.93	61.47	84.42	71.14
	Few-shot ICL	83.60 $\uparrow$ 8.20	83.60 $\uparrow$ 13.11	85.24 $\uparrow$ 21.31	59.01 $\downarrow$ 2.46	89.34 $\uparrow$ 4.92	80.16 $\uparrow$ 9.02
	Oscar	81.14 $\uparrow$ 5.74	84.42 $\uparrow$ 13.93	79.50 $\uparrow$ 15.57	63.11 $\uparrow$ 1.64	89.34 $\uparrow$ 4.92	79.50 $\uparrow$ 8.36
	CAA	76.22 $\uparrow$ 0.82	89.34 $\uparrow$ 18.85	74.59 $\uparrow$ 10.66	62.29 $\uparrow$ 0.82	90.98 $\uparrow$ 6.56	78.68 $\uparrow$ 7.54
	L2S	77.04 $\uparrow$ 1.64	77.04 $\uparrow$ 6.55	70.49 $\uparrow$ 6.56	63.11 $\uparrow$ 1.64	86.88 $\uparrow$ 2.46	74.91 $\uparrow$ 3.77
	Ours	94.26 $\uparrow$ 18.86	97.54 $\uparrow$ 27.05	95.08 $\uparrow$ 31.15	67.21 $\uparrow$ 5.74	91.80 $\uparrow$ 7.38	89.18 $\uparrow$ 18.04

Table 2: Performance comparison on diverse task benchmarks. Colored values indicate changes relative to the baseline. Best results are shown in bold , and second-best results are underlined .

Method	MC1	MC2	MC3	TRUE	INFO
LLaDA-8B-Instruct
Vanilla	22.13	47.08	23.45	49.18	81.96
w/o DF-Steering	40.98	58.47	27.55	50.00	84.42
w/o TS-Gate	40.98	57.32	30.66	45.90	83.60
w/o $\mathcal{L}_{\mathrm{reg}}$	36.88	56.92	30.08	11.47	10.65
Ours	43.44	61.19	31.63	57.37	89.34

Table 3: Ablation study of our proposed steering framework on TruthfulQA. Best results are shown in bold, and second-best results are underlined.

Figure 3: Effect of incorporating multiple previous predictions on model performance.

Figure 4: Cross-dataset generalization on two diffusion language models. The steering module is trained exclusively on TruthfulQA and directly evaluated on five unseen benchmarks without further adaptation.

Figure 5: Accumulated steering trajectories visualized with 3D PCA. Each path represents one sample across successive denoising iterations, with the endpoint marking its final accumulated steering state.

Figure 6: Case study of hallucination correction. The prediction trajectories of Vanilla and TraceSteer during denoising for the same example.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Method	Intervention	Guidance	Adaptation
Few-shot ICL	Input prompt	Demonstration examples	Contextual conditioning
OSCAR	Token remasking and refinement	Cross-chain uncertainty and retrieved evidence	Uncertain token positions
CAA	Hidden activations	Contrastive activation differences	Fixed vector and coefficient
L2S	Hidden activations	Input context representations	Input-specific steering vector
TraceSteer	Hidden activations	Denoising feedback, current states, uncertainty, and timestep	Step-specific steering vector and strength

Appendix

Table 4: Comparison of baseline mechanisms and TraceSteer .

Dataset	Question	Response
TruthfulQA	What happens to you if you eat watermelon seeds?	Factual: The watermelon seeds pass through your digestive system . Hallucination: You grow watermelons in your stomach .
COPA	My body cast a shadow over the grass. What was the most plausible cause? Choices: (A) The sun was rising; (B) The grass was cut.	A
StoryCloze	My friends all love to go to the club to dance. They think it’s a lot of fun and always invite. I finally decided to tag along last Saturday. I danced terribly and broke a friend’s toe. Choices: (1) My friends decided to keep inviting me out as I am so much fun; (2) The next weekend, I was asked to please stay home.	2
MMLU	Find the degree for the given field extension Q(sqrt(2), sqrt(3), sqrt(18)) over Q. Choices: (A) 0; (B) 4; (C) 2; (D) 6.	B
Winograd Schema	John moved the couch from the garage to the backyard to create space. The _ is small. Choices: (1) garage; (2) backyard.	1
BoolQ	Do Iran and Afghanistan speak the same language? Context: Persian is primarily spoken in Iran and Afghanistan, where it is officially known as Farsi and Dari, respectively.	Yes

Appendix

Table 5: Illustrative examples from six benchmark datasets. TruthfulQA includes both factual and hallucinated responses, whereas the remaining datasets present only correct answers. Correct responses are highlighted in blue, while hallucinated content is highlighted in red.

Dataset	Task	Answer Format
TruthfulQA	Factual reliability	Multiple-choice / free-form
COPA	Causal reasoning	Two choices
StoryCloze	Narrative reasoning	Two choices
WinoGrande	Commonsense reasoning	Two choices
MMLU	Knowledge and reasoning	Four choices
BoolQ	Reading comprehension	Yes / no

Appendix

Table 6: Overview of evaluation benchmarks.

LLaDA-8B-Instruct
Method	MC1	MC2	MC3	TRUE	INFO
Vanilla	22.13	47.08	23.45	49.18	81.96
w/o DF-Steering	40.98	58.47	27.55	50.00	84.42
w/o TS-Gate	40.98	57.32	30.66	45.90	83.60
w/o $\mathcal{L}_{\mathrm{reg}}$	36.88	56.92	30.08	11.47	10.65
Ours	43.44	61.19	31.63	57.37	89.34

Appendix

Table 7: Ablation study of our proposed steering framework on TruthfulQA. Best results are shown in bold, and second-best results are underlined.

LLaDA-8B-Instruct
Layer	MC1	MC2	MC3	TRUE	INFO
Layer 8	36.88	54.20	28.72	29.50	31.14
Layer 16	43.44	61.19	31.63	57.37	89.34
Layer 24	36.06	53.59	27.07	46.72	87.70
Dream-7B-Instruct
Layer 8	27.86	52.86	23.46	48.81	75.40

Appendix

Table 8: Effect of intervention layer on TruthfulQA. The feedback encoding layer and intervention layer are matched. Best results are shown in bold.

Figure 7: Inference efficiency. Inference time per sample across methods on LLaDA-8B-Instruct and Dream-7B-Instruct.

Panel of Figure 8 — Figure 8: Case study of hallucination correction. (a) The prediction trajectories of Vanilla and TraceSteer during denoising for the same example. (b) The gating coefficient $\gamma_{t}$ and the applied hidden-state correction magnitude $|\gamma_{t}\Delta H_{t}|_{2}$ across denoising steps for the selected case.

Training Settings
Configuration	Value
Backbone	LLaDA, Dream
Data split	70:15:15
Random seed	42
Training epochs	5
Batch size	1

Appendix

Table 9: Training and architectural configurations.

Language Model Steering

Language model steering controls a model’s behavior at inference time by intervening in its representations, components, or decoding process.

See all 283 papers

Alleviating Hallucination in Reasoning Tasks with Training-Free Uncertainty-Guided Steering
Sep 30, 2026Litian Liu, Qiqi Hou, Yubing Jian +5LLM Hallucination Mitigation Activation Steering
Qualcomm AI Research · UC Santa Barbara
FineSteer: A Unified Framework for Fine-Grained Inference-Time Steering in Large Language Models
Apr 16, 2026Zixuan Weng, Jinghuai Zhang, Kunlin Cai +3Language Model Steering LLM Safety
University of California, Los Angeles
Dynamic Multimodal Activation Steering for Hallucination Mitigation in Large Vision-Language Models
Feb 25, 2026Jianghao Yin, Qin Chen, Kedi Chen +3VLM Hallucination Mitigation Large Vision-Language Models
East China Normal University

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

Alleviating Hallucination in Reasoning Tasks with Training-Free Uncertainty-Guided Steering

Litian Liu, Qiqi Hou, Yubing Jian, Reza Pourreza+4

Qualcomm AI Research · UC Santa Barbara

LLM Hallucination Mitigation Activation Steering

Open paper AskPDF arXiv

Recent work on hallucination detection in large language models has shown that, for a fixed pre-trained model and reasoning task, it is possible to estimate the model's confidence in the correctness of its outputs. Such uncertainty estimates have primarily been used to improve truthfulness by detecting or filtering confabulations. In this work, we ask whether these signals can instead be used more proactively to directly improve the accuracy of model-generated answers. We propose USteer, a simple, training-free steering mechanism that adjusts a model's layer-wise activations during inference using the gradient of a confidence measure with respect to the activations. This procedure nudges generation toward outputs with lower uncertainty at inference time, without modifying model parameters or requiring additional supervision. We show that this approach consistently reduces hallucination across a range of tasks, demonstrating that confidence signals can be leveraged not only for detection, but also for effective inference-time control of model behavior.

Abstract

Language Model Steering

Explore similar work

Alleviating Hallucination in Reasoning Tasks with Training-Free Uncertainty-Guided Steering

FineSteer: A Unified Framework for Fine-Grained Inference-Time Steering in Large Language Models

Dynamic Multimodal Activation Steering for Hallucination Mitigation in Large Vision-Language Models