How Fine-Tuning Objectives Shape Layer-Wise Information in LLM Hallucination Detection

Kaiyang Wan; Forrest Sheng Bao; Amin Ahmad; Yuxia Wang

cs.AIOct 20, 2026

How Fine-Tuning Objectives Shape Layer-Wise Information in LLM Hallucination Detection

Authors: Kaiyang Wan, Forrest Sheng Bao, Amin Ahmad, Yuxia Wang

Organizations: INSAIT, Sofia University “St. Kliment Ohridski”, Sofia, Bulgaria · PAIR Systems, Inc., Mountain View, California

Abstract

Under matched conditions, classifier-head fine-tuning (cls) paradigm consistently outperforms the LLM-as-Judge sequence-to-sequence (s2s) on binary hallucination detection. We trace this gap to the representations induced by the two objectives. Static analysis shows that cls learns a task-specific discriminative direction, pushing hidden states into a clear two-cluster geometry with stronger class separation and linearly decodable label information. By contrast, s2s relies on the pretrained language-modeling head to make decisions. This minimizes the pressure to reorganize representations, leaving the hidden states much closer to the pretrained backbone. Our analysis also reveals that the cls signal saturates before the final layer, after which upper layers form a plateau of statistically interchangeable and largely redundant representations. However, endpoint redundancy alone does not show whether these layers are dispensable during training. Further training-dynamics analysis reveals 1) a lower pretrained substrate, 2) a narrow upper-middle InfoWindow where task information emerges, and 3) a later-synchronizing upper plateau that mirrors this signal. This mechanism motivates us to remove layers above the InfoWindow. Truncating 31% of the network leaves accuracy unchanged within seed noise on Qwen3-4B and 8B and cuts latency by $1.4\times$ . The same depth fraction, fixed before any analysis of the target models, transfers to Llama-3.2-3B and Mistral-7B-v0.3, and the layer structure behind it reappears there and on QNLI and SST-2. We hope this helps guide efficient task-specific LLM design.

Figures & tables

Backbone	Method	In-distribution			Out-of-distribution			Average
		RAGTruth	ClaimVerify	FactCheck	AggreFact	FaithBench	FavaBench	ID	OOD	All
Qwen3-4B	Base	55.7	70.5	71.0	80.9	24.3	69.4	65.7	58.2	62.0
	S2S LoRA	81.7 $\pm$ 1.1	79.3 $\pm$ 0.5	72.5 $\pm$ 3.0	77.9 $\pm$ 1.1	27.3 $\pm$ 1.5	45.2 $\pm$ 4.2	77.8 $\pm$ 1.0	50.1 $\pm$ 2.1	64.0 $\pm$ 0.9
	CLS LoRA	86.8 $\pm$ 0.1	79.4 $\pm$ 1.0	76.9 $\pm$ 2.1	83.1 $\pm$ 0.3	32.7 $\pm$ 0.3	96.7 $\pm$ 0.5	81.0 $\pm$ 1.0	70.8 $\pm$ 0.4	75.9 $\pm$ 0.5
	$\Delta$ ( cls $-$ s2s )	+5.1 $\pm$ 1.1	+0.1 $\pm$ 1.5	+4.4 $\pm$ 2.0	+5.2 $\pm$ 1.1	+5.4 $\pm$ 1.6	+51.5 $\pm$ 4.0	+3.2 $\pm$ 0.7	+20.7 $\pm$ 2.0	+11.9 $\pm$ 1.2
Qwen3-8B	Base	50.6	69.4	72.2	80.0	21.8	86.6	64.1	62.8	63.4

Table 1: Hallucination detection under matched fine-tuning conditions. Macro-F 1 (%) is reported as mean

\pm

standard deviation over three fine-tuning seeds (42, 43, 44); base is evaluated once. ID, OOD, and All are unweighted averages over the corresponding corpora, computed separately for each seed.

\Delta

reports paired cls

-

s2s differences.

Figure 1: Last-layer hidden-state geometry across paradigms and scales. All variants at a fixed scale share the same cls -defined projection. cls shows the clearest separation along

w^{\mathrm{disc}}

than s2s .

Figure 2: Per-layer direction-aligned Fisher score. cls shows an earlier and stronger rise in separation along the learned decision direction than s2s .

Figure 3: Per-layer probe F 1 and

\mathcal{V}

-information. cls rises sharply in the mid-to-upper layers and saturates near the label-entropy ceiling, whereas s2s improves modestly over base .

Backbone	$\bm{L^{\star}}$	$\bm{L^{\star}/L_{\max}}$	$\bm{\Delta_{\mathrm{final}}}$ [95% CI]	Equiv.
Qwen3-4B	L22	61%	$\bm{+0.001}$ $[-0.007,\,+0.008]$	$\checkmark$
Qwen3-8B	L23	64%	$\bm{-0.002}$ $[-0.009,\,+0.005]$	$\checkmark$

Table 2: Bootstrap plateau test for upper-layer saturation in cls .

L^{\star}

is the lowest layer whose probe F 1 is statistically indistinguishable from the final-layer under a 1000-sample bootstrap 95% CI. A confidence interval containing zero indicates readout equivalence to the final layer.

Backbone	$\bm{\ell^{\star}}$	$\bm{\widehat{UI}_{\ell}}$	$\bm{\widehat{\mathcal{S}}}$ (nats)	Fair OOD $\bm{\Delta F_{1}}$ [95% CI, pp]
Qwen3-4B	L18	$\bm{0}$	0.053	$\bm{=}$ -0.2 pp $[-0.6,\,+0.2]^{\,\mathrm{ns}}$
Qwen3-8B	L12	$\bm{0}$	0.049	$\bm{\downarrow}$ 0.9 pp $[-1.7,\,-0.0]$

Table 3: PID-style and fusion tests support upper-plateau homogeneity in cls . At the synergy-optimal layer

\ell^{\star}

, the candidate layer contributes zero unique information once a plateau layer is available, and the synergy budget remains small. Fair OOD

\Delta F_{1}

reports fusion minus single-layer readout, with 10,000-sample paired-bootstrap 95% CI stratified by dataset. Neither cls backbone shows a positive fusion gain, supporting truncation rather than layer fusion.

Figure 4: Per-layer $\mathcal{V}$ -information across the cls training trajectory. The figure plots layers as rows and saved checkpoints as columns, with color indicating $I_{\mathcal{V}}(\ell,t)$ in nats. The lower substrate (L1–L15) remains near its pretrained level, the InfoWindow (L16–L19) shows staggered task-information growth, and the upper plateau (L20–L36) synchronizes later within a tight training-step cluster.

Backbone

Exit position

Verdict

Layer

Saved

RAGTruth

ClaimVerify

FactCheck

AggreFact

FaithBench

FavaBench

Avg.

\bm{\Delta}

Analyzed backbones — exit chosen from the per-model InfoWindow analysis (Section 5.3 )

Qwen3-4B

Full cls (baseline)

—

86.8

\pm

0.1

79.4

\pm

1.0

76.9

\pm

2.1

83.1

\pm

0.3

32.7

\pm

0.3

96.7

\pm

0.5

75.9

\pm

0.5

—

Before InfoWindow

\times

too early

58%

84.9

79.1

74.5

75.7

28.0

52.9

65.9

-9.9 (

n{=}1

)

First plateau layer

\sim

borderline

44%

85.0

\pm

1.3

80.1

\pm

1.5

76.8

\pm

1.2

79.2

\pm

2.4

27.7

\pm

4.5

79.2

\pm

18.8

71.4

\pm

4.3

-4.5

\pm

4.8

InfoWindow + margin

\checkmark

prescribed

31%

86.4

\pm

0.9

80.2

\pm

0.4

76.7

\pm

0.4

82.3

\pm

1.5

32.0

\pm

2.6

94.6

\pm

2.9

75.4

\pm

1.1

-0.5

\pm

1.4

Qwen3-8B

Full cls (baseline)

—

86.8

\pm

0.3

80.0

\pm

0.6

77.1

\pm

1.5

83.7

\pm

0.8

34.2

\pm

4.4

94.1

\pm

2.5

76.0

\pm

0.9

—

Table 4: Early-exit accuracy and retained depth. Macro-F 1 (%) is reported as mean

\pm

standard deviation over three fine-tuning seeds (42, 43, 44), except for L15 (seed 42 only). “Saved” is the fraction of transformer blocks removed;

\Delta

is the six-corpus average difference from the full-depth model, paired by seed. Exit depths are held fixed across runs.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Corpus	Domain	Role	Raw	Deduped	Probe-half	Eval-half	$p(Y{=}1)$
Training corpora ( $95\%$ train / $5\%$ in-training validation)
RAGTruth train	general	train	14,949	—	—	—	0.554
ClaimVerify train	general	train	6,341	—	—	—	0.720
FactCheck-GPT train	general	train	1,537	—	—	—	0.173
Total			22,827	—	—	—	0.487
In-distribution test corpora (probe-half $\to$ analysis; eval-half $\to$ benchmark)

Appendix

Table 5: Per-corpus row counts. Training corpora (top) feed fine-tuning under a

95{/}5

train/validation split. Test corpora (bottom) are deduplicated by

(\text{premise},\text{hypothesis})

fingerprint, then split stratified

50{/}50

into a probe-half for static and dynamic analyses and an eval-half for benchmark evaluation.

p(Y{=}1)

is the positive-class rate measured on the deduplicated corpus.

Backbone	$\bm{L}$	$\bm{D}$	Exits $\ell_{\mathrm{exit}}$
Qwen3-4B	36	2560	15, 20, 25
Qwen3-8B	36	4096	15, 20, 25
Llama-3.2-3B	28	3072	17, 20
Mistral-7B-v0.3	32	4096	19, 23

Appendix

Table 6: Per-backbone constants for early-exit runs.

L

is the number of transformer blocks;

D

is the hidden width; “Exits” lists the

\ell_{\mathrm{exit}}

values evaluated in Table 4 .

Hyperparameter	Default (Qwen3, Llama)	Mistral-7B-v0.3
Optimizer	AdamW	AdamW
Learning rate	$1\times 10^{-4}$	$\bm{6\times 10^{-5}}$
Weight decay	$0.01$	$\bm{0.1}$
Warmup ratio (linear)	$0.03$	$\bm{0.05}$
LR scheduler	linear decay	linear decay
Number of epochs	$3$	$3$

Appendix

Table 7: Optimization hyperparameters for early-exit runs. “Default” covers Qwen3-4B, Qwen3-8B, and Llama-3.2-3B. The Mistral column changes only the listed optimization entries. The effective batch size is per-device batch size

\times

gradient-accumulation steps

\times

world size; all runs use one GPU, so the effective batch is

2\times 4\times 1=8

Backbone	Seed	S2S ID	CLS ID	$\Delta$ ID	S2S OOD	CLS OOD	$\Delta$ OOD	S2S All	CLS All	$\Delta$ All
Qwen3-4B	42	78.09	80.24	2.14	52.43	71.27	18.84	65.26	75.75	10.49
Qwen3-4B	43	78.85	82.36	3.51	47.34	70.75	23.41	63.10	76.56	13.46
Qwen3-4B	44	76.52	80.40	3.87	50.54	70.38	19.84	63.53	75.39	11.86
Qwen3-8B	42	77.97	80.33	2.36	66.68	74.07	7.39	72.33	77.20	4.87
Qwen3-8B	43	79.02	81.47	2.46	60.22	68.56	8.34	69.62	75.02	5.40
Qwen3-8B	44	78.12	82.10	3.98	66.54	69.34	2.80	72.33	75.72	3.39

Appendix

Table 8: Individual full-depth runs. Macro-F 1 (%) and paired differences in percentage points. Within each group,

\Delta

is cls

-

s2s .

Backbone	Exit	Seed	ID	OOD	All	$\Delta$ ID	$\Delta$ OOD	$\Delta$ All
Qwen3-4B	L15	42	79.52	52.19	65.85	-0.72	-19.08	-9.90
Qwen3-4B	L20	42	81.14	67.05	74.09	0.91	-4.22	-1.66
Qwen3-4B	L20	43	80.17	50.31	65.24	-2.19	-20.44	-11.32
Qwen3-4B	L20	44	80.72	68.76	74.74	0.33	-1.63	-0.65
Qwen3-4B	L25	42	81.40	72.11	76.76	1.17	0.84	1.00
Qwen3-4B	L25	43	81.01	67.42	74.22	-1.35	-3.33	-2.34

Appendix

Table 9: Individual early-exit runs. Macro-F 1 (%) and differences from the same-seed full-depth cls model. Exit depths are held fixed across seeds.

Variant	Qwen3-4B	Qwen3-8B	Llama-3.2-3B	Mistral-7B
s2s , generation (paper protocol)	65.2	72.3	60.2	67.7
(a) s2s , Fixed verbalizer-ensemble LLR	65.3	71.6	59.5	67.7
(c) s2s , Generation + trainable LM head	68.0	71.6	61.7	69.5
(b) s2s , LLR + per-corpus label-shift calibration	74.0	76.9	73.5	75.5
cls , fixed $0.5$ threshold (no tuning)	75.7	77.2	69.2	71.7
cls , same label-shift calibration as (b)	79.3	79.5	76.6	77.7

Appendix

Table 10: Alternative s2s readouts and matched threshold calibration. Six-corpus mean macro-F 1 (%), using fine-tuning seed 42. The verbalizer ensemble is fixed, without selection on the eval half. Per-corpus thresholds are fitted separately for each paradigm on labeled probe-half examples from that corpus. LM-head adaptation uses generation and verdict parsing.

	s2s on FavaBench			Gap, all six		Gap, excl. FavaBench
Backbone	parse fail	pred. neg / pos	verdict	$\Delta$ [95% CI]	$p$	$\Delta$ [95% CI]	$p$
Qwen3-4B	0.0%	33% / 67%	over-positive	+10.5 [+9.1, +11.8]	$<$ 0.001	+3.3 [+1.8, +4.9]	$<$ 0.001
Qwen3-8B	0.0%	69% / 31%	miscalibrated	+4.9 [+3.5, +6.2]	$<$ 0.001	+3.6 [+2.0, +5.1]	$<$ 0.001
Llama-3.2-3B	0.0%	11% / 89%	collapse to positive	+9.0 [+7.6, +10.3]	$<$ 0.001	+2.9 [+1.3, +4.4]	$<$ 0.001
Mistral-7B-v0.3	0.0%	49% / 51%	miscalibrated	+4.0 [+2.7, +5.3]	$<$ 0.001	+2.2 [+0.7, +3.7]	0.004

Appendix

Table 11: FavaBench predictions and sensitivity to excluding the corpus. All results use fine-tuning seed 42. Left: invalid-verdict rate and predicted class proportions on FavaBench, whose positive-label rate is 20%. Right: paired cls

-

s2s differences in mean macro-F 1 points, with 95% percentile intervals from 10,000 within-corpus bootstrap resamples. These intervals describe example-sampling uncertainty for fixed models.

Figure 5: Per-layer linear-probe F 1 on all four backbones. base , s2s , and cls probes on the in-domain probe half (mean of three probe seeds, bands show one standard deviation across probe seeds), plotted against relative depth

\ell/L

. Every backbone shows the flat pretrained substrate, a sharp rise, and a plateau, with cls above s2s above base ; the dotted line marks the cls saturation layer

L^{\star}_{1\%}

Backbone	$\bm{L}$	base	s2s	cls	$\bm{L^{\star}_{1\%}}$	$\bm{L^{\star}_{1\%}/L}$	cls $-$ s2s (same probe family)
Qwen3-4B	36	0.798	0.835	0.988	L23	0.64	+0.153
Qwen3-8B	36	0.809	0.850	0.984	L21	0.58	+0.134
Llama-3.2-3B	28	0.742	0.807	0.935	L12	0.43	+0.128
Mistral-7B-v0.3	32	0.756	0.848	0.986	L18	0.56	+0.138

Appendix

Table 12: Linear-probe accuracy and saturation depth across backbones. Scores are the maximum over layers of mean probe macro-F 1 across three probe seeds, using the original fine-tuning runs.

L^{\star}_{1\%}

is the earliest layer within 0.01 of that maximum. The final column subtracts the best-layer scores of separately fitted probes with the same architecture and training protocol. These are descriptive probe comparisons, not early-exit selection results.

Figure 6: The layer structure replicates on QNLI and SST-2. Per-layer probe F 1 for Qwen3-4B fine-tuned on hallucination detection, QNLI, and SST-2 with the same cls / s2s recipe. The shape and the location of

L^{\star}_{1\%}

are shared; the cls

-

s2s gap shrinks on the two easier tasks because the pretrained backbone already decodes them well.

Task (Qwen3-4B)	base	s2s	cls	cls $-$ s2s	$\bm{L^{\star}_{1\%}}$ ( cls )	$\bm{L^{\star}_{1\%}/L}$
Hallucination detection	0.798	0.835	0.988	+0.153	L23	0.64
QNLI (entailment)	0.895	0.924	0.947	+0.023	L22	0.61
SST-2 (sentiment)	0.940	0.956	0.959	+0.002	L18	0.50

Appendix

Table 13: Best-layer probe scores and saturation depths on three tasks. QNLI and SST-2 use a single fine-tuning seed and the logistic-regression protocol described in Appendix F.5 ; hallucination detection uses the three-probe-seed protocol in Appendix F.4 . Differences are computed before rounding.

L^{\star}_{1\%}

summarizes early saturation and does not test statistical equivalence.

GPU	Model	Full depth	L25 ( $-31\%$ blocks)	L20 ( $-44\%$ blocks)
H200 (datacenter)	Qwen3-4B	47.4 ms	33.0 ms ( 1.44 $\times$ )	26.5 ms ( 1.79 $\times$ )
H200 (datacenter)	Qwen3-8B	71.8 ms	50.1 ms ( 1.43 $\times$ )	40.1 ms ( 1.79 $\times$ )
RTX A6000 (constrained)	Qwen3-4B	220.6 ms	155.8 ms ( 1.42 $\times$ )	125.0 ms ( 1.76 $\times$ )

Appendix

Table 14: Forward-pass latency of physically truncated backbones. Mean time per forward pass at batch size 1 and sequence length 2,048, using bf16 and SDPA. Models use pretrained backbone weights and a two-class head, without task adapters. Parentheses give speedup over full depth. These measurements isolate the effect of reducing depth; they are not end-to-end detector timings.

Fine-Tuning Language Models

See all 46 papers

May 25, 2026

Automatic Layer Selection for Hallucination Detection

Recent studies on hallucination detection have shown that hallucination-related signals are more strongly encoded in intermediate layers than in the final layer of large language models (LLMs). Although a growing body of work has sought to exploit this property for hallucination detection, how to automate the selection of high-performing layers remains underexplored, and principled methods for this purpose are still lacking. To address this gap, we first propose several hypotheses for why such signals emerge in intermediate layers and evaluate corresponding criteria for automatic layer selection across diverse LLM architectures, scales, and tasks, covering both question answering and summarization hallucination detection benchmarks. However, we find that none of these criteria consistently delivers satisfactory performance. We therefore propose a new selection criterion, First Effective Peak of Intrinsic Dimension (FEPoID), which consistently identify optimal or near-optimal layers and outperforms both the aforementioned criteria and existing hallucination detection baselines. FEPoID is training-free and incurs negligible computational overhead. In addition, we study the generation behaviors of LLMs and introduce a simple yet effective truncation strategy, which further amplifies hallucination-related signals and substantially improves overall detection performance. Code is publicly available at https://github.com/DesoloYw/Automatic-Layer-Selection-for-Hallucination-Detection.git

Xinpeng Wang, William X. Cao, Andrew Gordon Wilson +1

University of Virginia · 2Walrus Security · 3New York University.

Hallucination in Language Models LLM Hallucination Detection

Ask PDF

May 31, 2026

TriLens: Per-Layer Logit-Lens Entropy for White-Box Hallucination Detection

When a language model hallucinates, the final answer is wrong, but the mistake is not necessarily invisible inside the model. Different internal pathways may remain uncertain, disagree in how quickly they sharpen, or commit to competing continuations before the output is produced. We introduce TriLens, a white-box detector that turns this intuition into a compact representation: at every layer, it reads the multi-head self-attention output, the feed-forward output, and the residual stream through the model's own logit lens, then records only the entropy of each readout. The resulting 3L-dimensional trajectory describes how certainty forms across depth and across modules, without storing high-dimensional hidden states or sampling multiple generations. This simple signal yields a strong detector across instruction-tuned LLMs and QA benchmarks, and our analyses show that the three module-wise entropy trajectories provide complementary evidence. TriLens suggests that hallucination detection can benefit from tracking how internal computation settles, not only what the final layer predicts.

Bohan Yang, Yijun Gong, Zhi Zhang +3

Binjiang Institute of Zhejiang University · Beijing Normal-Hong Kong Baptist University · Great Bay University +1

Language Model Error Detection LLM Reliability LLM Hallucination Detection LLM Uncertainty Estimation

Ask PDF

Sep 4, 2025

Cross-Layer Attention Probing for Fine-Grained Hallucination Detection

With the large-scale adoption of Large Language Models (LLMs) in various applications, there is a growing reliability concern due to their tendency to generate inaccurate text, i.e. hallucinations. In this work, we propose Cross-Layer Attention Probing (CLAP), a novel activation probing technique for hallucination detection, which processes the LLM activations across the entire residual stream as a joint sequence. Our empirical evaluations using five LLMs and three tasks show that CLAP improves hallucination detection compared to baselines on both greedy decoded responses as well as responses sampled at higher temperatures, thus enabling fine-grained detection, i.e. the ability to disambiguate hallucinations and non-hallucinations among different sampled responses to a given prompt. This allows us to propose a detect-then-mitigate strategy using CLAP to reduce hallucinations and improve LLM reliability compared to direct mitigation approaches. Finally, we show that CLAP maintains high reliability even when applied out-of-distribution. Code available at: https://github.com/itsmemala/CLAP

Malavika Suresh, Rahaf Aljundi, Ikechukwu Nkisi-Orji +1

Robert Gordon University, Aberdeen, United Kingdom · Toyota Motor Europe, Brussels, Belgium

LLM Hallucination Detection LLM Reliability Language Model Probing LLM Hallucination Mitigation

Ask PDF

Abstract

Fine-Tuning Language Models

Explore similar work

Automatic Layer Selection for Hallucination Detection

TriLens: Per-Layer Logit-Lens Entropy for White-Box Hallucination Detection

Cross-Layer Attention Probing for Fine-Grained Hallucination Detection