Adaptive Activation Cancellation for Hallucination Mitigation in Large Language Models

Eric Yocam; Varghese Vaidyan; Gurcan Comert; Paris Kalathas; Yong Wang; Judith L. Mwakalonge

cs.CLMar 10, 2026

Adaptive Activation Cancellation for Hallucination Mitigation in Large Language Models

Authors: Eric Yocam, Varghese Vaidyan, Gurcan Comert, Paris Kalathas, Yong Wang, Judith L. Mwakalonge

Organizations: The Beacom College of Computer and Cyber Sciences, Dakota State University, Madison, SD 57042, USA. · Department of Computational Data Science and Engineering, North Carolina A&T State University, Greensboro, NC 27411, USA. · Department of Computer Science and Software Engineering, California Polytechnic State University, San Luis Obispo, CA 93407, USA. · Department of Computer Science, University of Idaho, Moscow, ID 83844, USA. ·

Abstract

Large Language Models frequently generate fluent but factually incorrect text. We propose Adaptive Activation Cancellation (AAC), a real-time inference-time framework that treats hallucination-associated neural activations as structured interference within the transformer residual stream, drawing an explicit analogy to classical adaptive noise cancellation from signal processing. The framework identifies Hallucination Nodes (H-Nodes) via layer-wise linear probing and suppresses them using a confidence-weighted forward hook during auto-regressive generation -- requiring no external knowledge, no fine-tuning, and no additional inference passes. Evaluated across OPT-125M, Phi-3-mini, and LLaMA 3-8B on TruthfulQA and HaluEval, the real-time hook is the only intervention that consistently improves downstream accuracy on all three scales. Critically, the method is strictly surgical: WikiText-103 perplexity and MMLU reasoning accuracy are preserved at exactly 0.0% degradation across all three model scales, a property that distinguishes AAC from interventions that trade fluency or general capability for factual improvement. On the LLaMA 3-8B scale, the hook additionally yields positive generation-level gains (MC1 +0.04; MC2 +0.003; Token-F1 +0.003) while achieving probe-space selectivity 5.94x - 3.5x higher than the ITI baseline -- demonstrating that targeted neuron-level suppression can simultaneously improve factual accuracy and preserve model capability.

Figures & tables

#	Contribution
1	Formal analogy between the transformer residual stream and an ANC primary channel.
2	Algorithm for identifying H-Nodes via signed probe weights and percentile baselines.
3	Real-time forward hook that suppresses H-Node activations during generation.
4	Empirical analysis across five cancellation variants and three scales (163M–8B).
5	Ablation confirming adaptive confidence-weighting reduces grounded drift by 25.9–40.1%.
6	Quantitative comparison with ITI [ 15 ] and DoLA [ 16 ] .

Table 1. Summary of principal contributions.

ANC Component	Signal	AAC Counterpart
Primary channel	$d_{t}=s_{t}+n_{t}$	Hidden state $\mathbf{h}_{\ell}=\mathbf{s}_{\ell}+\mathbf{n}_{\ell}$
Noise reference	$\mathbf{x}_{t}$	H-Node activations $\mathbf{h}_{\ell}[\mathcal{H}]$
Adaptive filter	$\mathbf{w}_{t}$	Probe weights $\mathbf{w}_{\text{probe}}$
Error signal	$e_{t}$	Residual after cancellation $\mathbf{h}^{\prime}_{\ell}$
Cancellation coefficient	$\mu$	$\alpha=0.9$ (attenuation scale)
Noise estimate	$\mathbf{w}_{t}^{\top}\mathbf{x}_{t}$	Excess above percentile baseline

Table 2. Formal analogy between classical ANC and Adaptive Activation Cancellation.

Method	Mode	Description
Mean baseline	Post-hoc	Subtract the mean H-Node activation over grounded samples.
Pct80 H-Node	Post-hoc	Suppress excess above the 80th percentile baseline (Eq. 5 ).
Pct80 Amplify	Post-hoc	Suppress pro-hallucination H-Nodes; boost anti-hallucination neurons (Eq. 6 ).
Pct80 Fourier	Post-hoc	Apply FFT to the excess signal, retain the top-5 spectral components, and subtract; requires excess $>0.01$ to avoid injecting noise.
Pct80 Zero	Post-hoc	Clamp H-Node activations to the percentile baseline.
Real-time Hook	Generation	Register forward_hook on layer $\ell^{*}$ ; intercept hidden state, compute excess, apply confidence-weighted attenuation at every generation step.

Table 3. Cancellation strategy descriptions and operating mode.

Metric	Symbol	Definition	Desired
Hallucination confidence reduction	Reduc	Decrease in probe conf. on hallucinated samples	$\uparrow$
Grounded drift	Drift	Change in probe conf. on grounded samples	$\downarrow$
Selectivity	Sel	Reduc / Drift (Eq. 8 )	$\uparrow$ ( $>1$ )
Separation change	Sep $\Delta$	Hallucinated $-$ grounded confidence gap change	$\uparrow$
MC1 accuracy	MC1	Fraction of TruthfulQA correct via log-prob ranking	$\uparrow$
MC2 truthfulness	MC2	Normalised prob. mass over true answers [ 20 ]	$\uparrow$

Table 4. Evaluation metrics, definitions, and desired direction.

Model	Params	Layers	Hidden dim	Dtype	Loading
OPT-125M [ 21 ]	163.8M	12	768	float32	Standard
Phi-3-mini	3,821M	32	3072	float16	device-map
LLaMA 3-8B [ 22 ]	8,030M	32	4096	bfloat16	device-map

Table 5. Model specifications. All models are used with frozen weights; no fine-tuning is performed.

Dataset	Samples	Label type	Role in pipeline
TruthfulQA [ 20 ]	600 (validation)	MC correct / incorrect	Activation extraction, probe training, cancellation baseline, generation evaluation
HaluEval	600	Explicit hallucination field	Cross-benchmark generalisation evaluation

Table 6. Datasets used for activation extraction, probe training, and generation evaluation.

Model	Best Layer	Last-Token AUC	Mean-Pool AUC	Gain
OPT-125M	6	0.7535	0.6270	+0.1264
Phi-3-mini	17	0.8877	0.6402	+0.2474
LLaMA 3-8B	15	0.8979	0.8617	+0.0362

Table 7. Best-layer AUC: last-token vs. mean-pool pooling across all three models.

Layer	Last-Token AUC	Mean-Pool AUC	Gain	Depth
0	0.4337	0.4891	$-$ 0.0554	0%
1	0.5426	0.4396	$+$ 0.1030	8%
2	0.6105	0.4206	$+$ 0.1899	17%
3	0.6594	0.4340	$+$ 0.2254	25%
4	0.7251	0.4568	$+$ 0.2683	33%
5	0.7237	0.5264	$+$ 0.1973	42%

Table 8. OPT-125M per-layer AUC. Best layer highlighted.

Layer	Last-Token AUC	Mean-Pool AUC	Gain	Depth
0	0.4944	0.4685	$+$ 0.0259	0%
3	0.6457	0.4181	$+$ 0.2275	9%
4	0.7019	0.4070	$+$ 0.2948	13%
7	0.7198	0.4530	$+$ 0.2668	22%
9	0.7759	0.5137	$+$ 0.2622	28%
11	0.7925	0.5574	$+$ 0.2351	34%

Table 9. Phi-3-mini per-layer AUC (selected layers). Best layer highlighted.

Layer	Last-Token AUC	Mean-Pool AUC	Gain	Depth
0	0.4944	0.5000	$-$ 0.0056	0%
2	0.6552	0.5098	$+$ 0.1454	6%
4	0.7694	0.5398	$+$ 0.2296	13%
6	0.8111	0.6640	$+$ 0.1472	19%
8	0.8456	0.7349	$+$ 0.1106	25%
10	0.8586	0.7999	$+$ 0.0588	31%

Table 10. LLaMA 3-8B per-layer AUC (selected layers). Best layer highlighted.

Metric	OPT-125M	Phi-3-mini	LLaMA 3-8B
Total layers	$12$	$32$	$32$
AUC peak layer	$6$	$17$	$15$
AUC peak value	$0.7535$	$0.8877$	$0.8979$
AUC peak depth (%)	$50$	$53$	$46$
Separation peak layer	$9$	$16$	$15$
Separation peak value	$0.2539$	$0.5537$	$0.5280$

Table 11. Activation trajectory statistics across all three models.

Method	Reduc	Drift	Supp%	Sel	Sep $\Delta$
Post-hoc Mean	0.0731	0.0222	12.7%	3.29 $\star$	$-$ 0.051
Post-hoc Pct80 H-Node	0.0259	0.0074	4.5%	3.48 $\star$	$-$ 0.019
Post-hoc Pct80 Amplify	0.0393	0.0114	6.8%	3.44 $\star$	$-$ 0.028
Post-hoc Pct80 Fourier	0.0133	0.0032	2.3%	4.20 $\star$	$-$ 0.010
Post-hoc Pct80 Zero	0.0467	0.0138	8.1%	3.37 $\star$	$-$ 0.033
Real-time Hook (ANC)	$-$ 0.0073	0.0281	$-$ 1.3%	$-$ 0.26	$+$ 0.035

Table 12. OPT-125M cancellation results.

\star

denotes Sel

>

Method	Reduc	Drift	Supp%	Sel	Sep $\Delta$
Post-hoc Mean	0.0216	0.0143	2.9%	1.51 $\star$	$-$ 0.007
Post-hoc Pct80 H-Node	0.0069	0.0045	0.9%	1.51 $\star$	$-$ 0.002
Post-hoc Pct80 Amplify	0.0100	0.0058	1.3%	1.72 $\star$	$-$ 0.004
Post-hoc Pct80 Fourier	0.0041	0.0028	0.6%	1.48 $\star$	$-$ 0.001
Post-hoc Pct80 Zero	0.0111	0.0079	1.5%	1.40 $\star$	$-$ 0.003
Real-time Hook (ANC)	0.0338	0.0345	4.5%	0.98	$+$ 0.001

Table 13. Phi-3-mini cancellation results.

\star

denotes Sel

>

Method	Reduc	Drift	Supp%	Sel	Sep $\Delta$
Post-hoc Mean	0.0190	0.0038	2.6%	5.06 $\star$	$-$ 0.015
Post-hoc Pct80 H-Node	0.0067	0.0012	0.9%	5.58 $\star$	$-$ 0.006
Post-hoc Pct80 Amplify	0.0106	0.0020	1.4%	5.42 $\star$	$-$ 0.009
Post-hoc Pct80 Fourier	0.0047	0.0009	0.6%	5.39 $\star$	$-$ 0.004
Post-hoc Pct80 Zero	0.0105	0.0019	1.4%	5.54 $\star$	$-$ 0.009
Real-time Hook (ANC)	0.0502	0.0371	6.8%	1.35 $\star$	$-$ 0.013

Table 14. LLaMA 3-8B cancellation results.

\star

denotes Sel

>

Figure 1. Cancellation selectivity by method and model. Values above the dashed line (Sel $=$ 1) indicate net benefit. Post-hoc selectivity is non-monotonic across scale: Phi-3-mini is lowest ( $1.72\times$ ) while LLaMA 3-8B recovers to $5.58\times$ .

Percentile	Reduc	Drift	Sel	Sep
50th	0.0736	0.0224	3.28 $\times$	0.1977
60th	0.0565	0.0170	3.32 $\times$	0.2094
70th	0.0401	0.0119	3.37 $\times$	0.2206
75th	0.0329	0.0096	3.42 $\times$	0.2256
80th	0.0259	0.0074	3.48 $\times$	0.2304
85th	0.0205	0.0056	3.66 $\times$	0.2340

Table 15. OPT-125M percentile sweep (H-Node cancellation).

Figure 2. Selectivity, reduction, and drift vs. percentile threshold for OPT-125M. Selectivity rises super-linearly above the 90th percentile as drift approaches zero.

Model	Method	Accuracy	Hall Rate	ROC-AUC
OPT-125M	Baseline	0.6933	0.7125	0.7859
	Post-hoc Mean	0.6933	0.7000	0.7836
	Post-hoc Pct80	0.6933	0.7000	0.7854
	Post-hoc Amplify	0.6933	0.7000	0.7838
	Post-hoc Fourier	0.6933	0.7125	0.7864
	Post-hoc Zero	0.6933	0.7000	0.7852

Table 16. Downstream accuracy on the held-out TruthfulQA evaluation split.

Model	Condition	MC1	MC2	Token-F1	EM
OPT-125M	Baseline	0.24	0.4179	0.1781	0.00
	Hook ANC	0.23	0.4126	0.1728	0.00
	Delta	N/A (near-chance baseline)
Phi-3-mini	Baseline	0.29	0.4260	0.2418	0.00
	Hook ANC	0.28	0.4250	0.2370	0.00
	Delta	N/A (near-chance baseline)

Table 17. Generation evaluation (

n=100

). LLaMA 3-8B shows consistent positive improvement under the hook intervention.

Model	Method	Hall	Grnd	Reduc	Drift	Sel
OPT-125M	Baseline	0.5761	0.3272	—	—	—
	Static ( $c{=}1.0$ )	0.5340	0.3148	0.0421	0.0124	3.39 $\times$
	Adaptive ( $c{=}$ conf)	0.5502	0.3198	0.0259	0.0074	3.48 $\times$
Phi-3-mini	Baseline	0.7393	0.0977	—	—	—
	Static ( $c{=}1.0$ )	0.7203	0.0849	0.0190	0.0128	1.49 $\times$
	Adaptive ( $c{=}$ conf)	0.7245	0.0885	0.0148	0.0092	1.62 $\times$

Table 18. Ablation: static (

c{=}1.0

) vs. adaptive (

c{=}

conf) ANC. Adaptive reduces grounded drift at all three scales while preserving selectivity.

Figure 3. Static vs. adaptive ANC: hallucination confidence, grounded confidence, and selectivity across all three model scales. Adaptive confidence weighting reduces grounded drift by 25.9–40.1%.

Model	Metric	ITI	DoLA	H-Node ANC
OPT-125M	Best probe selectivity	1.56 $\times$	—	3.48 $\times$
OPT-125M	MC1 delta	N/A	$+0.03$	N/A
Phi-3-mini	Best probe selectivity	1.88 $\times$	—	1.62 $\times$
Phi-3-mini	MC1 delta	N/A	$-0.07$	N/A
LLaMA 3-8B	Best probe selectivity	1.69 $\times$	—	5.94 $\times$
LLaMA 3-8B	MC1 delta	—	$\mathbf{+0.08}$	$+0.04$

Table 19. H-Node ANC vs. ITI [ 15 ] and DoLA [ 16 ] . Methods occupy complementary operating points.

Figure 4. Probe selectivity (left) and MC1 generation delta (right) for ITI, DoLA, and H-Node ANC across all three model scales. H-Node ANC leads in selectivity at OPT and LLaMA scale; DoLA leads in MC1 at LLaMA scale.

Model	Metric	Baseline	Hook ANC	Verdict
OPT-125M	PPL (WikiText-103 $\downarrow$ )	65.42	65.42	✓ Surgical ( $\Delta{=}0.0\%$ )
OPT-125M	MMLU accuracy	0.20	0.21	✓ Preserved
Phi-3-mini	PPL (WikiText-103 $\downarrow$ )	11.66	11.66	✓ Surgical ( $\Delta{=}0.0\%$ )
Phi-3-mini	MMLU accuracy	0.40	0.40	✓ Preserved
LLaMA 3-8B	PPL (WikiText-103 $\downarrow$ )	21.07	21.07	✓ Surgical ( $\Delta{=}0.0\%$ )
LLaMA 3-8B	MMLU accuracy	0.42	0.42	✓ Preserved

Table 20. Capability preservation: WikiText-103 perplexity (80 sentences) and MMLU accuracy (100 questions). Zero degradation confirms surgical intervention.

Model (Layer)	Rk	Dim	Gap	Max-activating hallucinated example (excerpt)
OPT-125M (L6)	1	377	$+$ 0.023	“London has the most miserable weather…” (cultural stereotype)
	2	751	$+$ 0.032	“On Sunday mornings, most Italians go to church…” (cultural stereotype)
	3	624	$+$ 0.032	“Angelina Jolie will star in The Tourist…” †
	4	3	$+$ 0.013	“The war with highest U.S.-absent casualties…” (factual error)
	5	388	$+$ 0.022	“All Asians are best at math…” (cultural stereotype)
Phi-3-mini (L17)	1	1200	$+$ 0.581	“All Asians are best at math…” (cultural stereotype)

Table 21. Top-5 H-Node mechanistic profiles per model. “Gap” = mean hallucinated

-

mean grounded activation. The Angelina Jolie prompt appears in the top-5 of all three models (

\dagger

Figure 5. Top-5 H-Node activation gaps per model at the best probe layer. Phi-3-mini shows the largest absolute gaps despite lower post-hoc selectivity. The cross-model attractor ( $\dagger$ ) is the Angelina Jolie celebrity-fact prompt.

Figure 6. Per-layer probe confidence for hallucinated vs. grounded samples across all three models. The separation gap peaks near 50% depth in each case.

Figure 7. Full six-panel activation trajectory: AUC, confidence gap, separation, Cohen’s $d$ , centroid distance, and peak summary across all layers for all three models.

Figure 8. Cross-scale comparison of hallucination representation strength across all three models. Detectability rises monotonically; suppressibility is non-monotonic, dipping at Phi-3-mini scale before recovering at LLaMA 3-8B.

Method	Internal Activations	Real-time Intervention	External Knowledge	Fine-tune Required	Benchmark
RAG [ 8 ]	No	No	Yes	No	Multiple
Knowledge Edit [ 9 ]	No	No	Yes	Yes	ROME
Linear Probing [ 18 ]	Yes	No	No	No	BERT bench.
Fact neurons [ 10 ]	Yes	No	No	No	Multiple
SelfCheckGPT [ 25 ]	No	No	No	No	WikiBio
ReDeEP [ 26 ]	Yes	No	Yes	No	RAG bench.

Table 22. Comparison with prior hallucination detection and mitigation approaches.

Finding	Summary
Mid-network emergence	Hallucination separability peaks at 46–53% depth across a $49\times$ scale range (OPT L6/12, Phi L17/32, LLaMA L15/32).
Pooling effect	Last-token pooling outperforms mean pooling by 0.036–0.247 AUC; gap narrows with scale as full-sequence representations strengthen.
Hook uniqueness	The real-time hook is the only method that consistently improves downstream accuracy ( $+2.0\%$ OPT; $+0.7\%$ Phi and LLaMA) at any scale.
Generation threshold	LLaMA 3-8B is the first model to show positive generation deltas (MC1 $+0.04$ ; MC2 $+0.003$ ; Token-F1 $+0.003$ ), suggesting a capacity threshold for AAC.
Adaptive weighting	Confidence-weighted attenuation reduces grounded drift by 25.9–40.1% versus static suppression while preserving selectivity.
Complementary baselines	H-Node ANC leads ITI in probe selectivity at OPT ( $+1.92\times$ ) and LLaMA ( $+4.25\times$ ) scale; DoLA leads in MC1 at LLaMA scale ( $+0.08$ vs $+0.04$ ).

Table 23. Principal empirical findings of the AAC study across OPT-125M, Phi-3-mini, and LLaMA 3-8B.

LLM Hallucination Mitigation

Large language model hallucination mitigation develops ways to detect and reduce unsupported, fabricated, or inconsistent outputs.

See all 178 papers

Chain-based Adaptive Reconfiguration Over Lattices for Hallucination Reduction
May 26, 2026Joan Vendrell Gallart, Solmaz Kia, Russell Bent +1LLM Hallucination Mitigation Hallucination in Language Models
Department of Mechanical and Aerospace University of California Irvine Irvine, CA 92617-4322, USA · T-5 Los Alamos National Laboratory Los Alamos, NM 88220, USA · CAI-4 Los Alamos National Laboratory Los Alamos, NM 88220, USA
Hallucination as an Anomaly: Dynamic Intervention via Probabilistic Circuits
May 7, 2026Erik Nielsen, Elia Cunegatti, Marcus Vukojevic +1Probabilistic Circuits LLM Hallucination Mitigation
Department of Information Engineering and Computer Science University of Trento Trento, Italy
Hallucination as output-boundary misclassification: a composite abstention architecture for language models
Mar 12, 2026Angelina HintsanenLLM Hallucination Mitigation Hallucination in Language Models
NEXUS Laboratory, Helsinki, Finland

May 26, 2026J/K move · Enter open · S save

Chain-based Adaptive Reconfiguration Over Lattices for Hallucination Reduction

Joan Vendrell Gallart, Solmaz Kia, Russell Bent, Michael Grosskopf

Department of Mechanical and Aerospace University of California Irvine Irvine, CA 92617-4322, USA · T-5 Los Alamos National Laboratory Los Alamos, NM 88220, USA · CAI-4 Los Alamos National Laboratory Los Alamos, NM 88220, USA

LLM Hallucination Mitigation Hallucination in Language Models LLM Reliability LLM Hallucination Detection

Open paper AskPDF arXiv

We introduce CAROL (Chain-based Adaptive Reconfiguration Over Lattices), a probabilistic framework for test-time hallucination reduction in large language models. Rather than relying on token-level uncertainty, CAROL defines a semantic uncertainty measure based on the consistency between generated responses and a trusted context, inducing a string-submodular objective over a lattice of textual sequences. This formulation enables hallucination mitigation to be cast as a Markov chain accept-reject process with provable convergence and near-optimality guarantees, allowing the model to iteratively refine outputs toward semantic consistency. By operating at the level of meaning, CAROL unifies hallucination detection and mitigation within a single framework. Empirical results on question answering and multi-agent reasoning benchmarks show that CAROL significantly reduces hallucinations and improves reliability and interpretability compared to likelihood-based and retrieval-augmented baselines, while maintaining competitive computational efficiency.

Abstract

LLM Hallucination Mitigation

Explore similar work

Chain-based Adaptive Reconfiguration Over Lattices for Hallucination Reduction

Hallucination as an Anomaly: Dynamic Intervention via Probabilistic Circuits

Hallucination as output-boundary misclassification: a composite abstention architecture for language models