What LLMs Do in the Shadows: Reading Context Reliance from the Unembedding Dark Space

Alexandra Kuleshova; Andrey Volodichev; Daria Kotova; Alexey Zaytsev

cs.CLOct 16, 2026

What LLMs Do in the Shadows: Reading Context Reliance from the Unembedding Dark Space

Authors: Alexandra Kuleshova, Andrey Volodichev, Daria Kotova, Alexey Zaytsev

Organizations: Applied AI Institute

Abstract

Detecting faithfulness hallucinations in large language models (LLMs) is crucial for their safe deployment in high-stakes domains. Recent approaches address this by contrasting contextual against parametric knowledge utilization, yet measuring context reliance typically demands noticeable overhead, such as materializing attention maps, running additional forward passes, or training auxiliary probes. In this work, we show that context reliance can be estimated by projecting hidden states onto the dark space - directions that compose weakly with the unembedding matrix. Beyond its recognized function as an initial-token attention sink, we show that this subspace actively mediates context processing, and that its share of hidden-state norm - termed the Dark Matter Ratio (DMR) - provides a direct score of context utilization. Pairing DMR with a parametric knowledge score yields DarkLens, a lightweight faithfulness detector operating strictly within a single forward pass. Evaluated across three QA and summarization benchmarks on six open-weight LLMs (4B-13B parameters), DarkLens matches or outperforms a wide range of baselines without requiring extra forward passes or probe training.

Figures & tables

Figure 1: Synthetic induction residual swaps on $[x_{1}{\ldots}x_{T},\,x_{1}{\ldots}x_{T}]$ . Each intervention covers a window of four consecutive layers. The first and third panels plot $\Delta$ NLL on the repeated half against the start layer of that window ( dark vs. mid-spectrum control); the dotted line is the average accuracy drop on the second half of the sequence. The second and fourth plot NLL (solid) and accuracy (dashed) against token position at the peak window; the vertical line marks the start of the repeat. Dark swaps impair copying far more than matched mid-spectrum swaps.

Figure 2: Language-modeling NLL under dark vs. mid-spectrum residual ablation. Each intervention covers a window of four consecutive layers; bars show NLL against the first layer of that window. Solid grey: unablated large model; dashed pink: smaller reference model (Qwen3-1.7B and Llama-3.2-1B-Instruct).

			L20–23
Model	Dataset	No ablation	Middle	Dark
Qwen3-8B	NQ-Swap	0.88	0.81	0.22
	ConflictQA	0.98	0.99	0.54
	Cities	1.0	0.83	0.99
	Sports	0.95	0.81	0.93
	Celebrity	0.96	0.80	0.95

Table 1: Share of examples on which the target continuation has strictly lower mean NLL than the alternative. Datasets above the rule are contextual (NQ-Swap, ConflictQA); those below are parametric (Cities, Sports, Celebrity). Middle and dark permute layers 20–23. Dark subspace dimension 205 (last 5% of right singular vectors of

W_{U}

Figure 3: NQ-Swap answer NLL under dark vs. mid-spectrum residual permutation.

\Delta

NLL is relative to the unablated model (positive: less likely). Solid: dark subspace; dashed: matched mid-spectrum control. Context, parametric, and distractor answers are scored separately.

Figure 4: Contextual hit rate under residual enhancement (top) and shrinking (bottom) on layers 20–23. The dark component (solid) or a matched mid-spectrum component (dashed) is scaled by strength

\alpha

; dotted lines are the unablated model. Amplifying the dark subspace raises context use on NQ-Swap and ConflictQA, while shrinking it reduces context use; the mid-spectrum control does not show the same effect.

	MS MARCO			CNN/DM
	Mistral-7B	Llama-2-7B	Llama-2-13B	Mistral-7B	Llama-2-7B	Llama-2-13B
Ours	0.77 $\pm$ 0.02	0.71 $\pm$ 0.04	0.72 $\pm$ 0.04	0.66 $\pm$ 0.07	0.58 $\pm$ 0.03	0.59 $\pm$ 0.05
LLM Check	0.49 $\pm$ 0.05	0.47 $\pm$ 0.07	0.51 $\pm$ 0.05	0.54 $\pm$ 0.04	0.50 $\pm$ 0.06	0.54 $\pm$ 0.03
Focus	0.67 $\pm$ 0.06	0.63 $\pm$ 0.03	0.61 $\pm$ 0.03	0.62 $\pm$ 0.04	0.51 $\pm$ 0.09	0.50 $\pm$ 0.06
LN-Entropy	0.65 $\pm$ 0.06	0.63 $\pm$ 0.03	0.62 $\pm$ 0.04	0.65 $\pm$ 0.04	0.63 $\pm$ 0.09	0.56 $\pm$ 0.04
LSC	0.68 $\pm$ 0.04	0.57 $\pm$ 0.03	0.57 $\pm$ 0.05	0.64 $\pm$ 0.04	0.62 $\pm$ 0.09	0.56 $\pm$ 0.05

Table 2: AUROC scores of hallucination detection methods. Bold and underline indicate the best and second-best results, respectively. All results are averaged over 5 runs.

	MS MARCO
Ablation	Mistral-7B	Llama-2-7B	Llama-2-13B
PKS	0.74 $\pm$ 0.01	0.63 $\pm$ 0.04	0.70 $\pm$ 0.02
DMR	0.77 $\pm$ 0.02	0.72 $\pm$ 0.03	0.72 $\pm$ 0.04
Both	0.77 $\pm$ 0.02	0.71 $\pm$ 0.04	0.72 $\pm$ 0.04
	XSum
Ablation	Qwen3-8B	gemma-3-4b	Llama-3.1-8B

Table 3: Ablation of DarkLens components (AUROC score, mean

\pm

std over 5 runs) on MS MARCO and XSum. PKS: parametric knowledge score only; DMR: standalone negative dark matter ratio (

-\operatorname{DMR}

); Both: combined DarkLens detector.

Figure 5: Response-level AUROC of DarkLens as a function of the dark subspace rank (

d_{\mathrm{dark}}

) on MS MARCO (left) and XSum (right).

Figure 6: Response-level AUROC of

-\operatorname{DMR}

as a function of the residual readout layer on MS MARCO.

	MS MARCO		XSum
Subspace	Mistral-7B	Llama-2-7B	Qwen3-8B	gemma-3-4b
Dark	0.74 $\pm$ 0.03	0.72 $\pm$ 0.03	0.71 $\pm$ 0.05	0.67 $\pm$ 0.05
Middle	0.57 $\pm$ 0.05	0.50 $\pm$ 0.05	0.56 $\pm$ 0.08	0.56 $\pm$ 0.05
Random	0.51 $\pm$ 0.04	0.52 $\pm$ 0.05	0.47 $\pm$ 0.09	0.57 $\pm$ 0.05

Table 4: Subspace ablation for

-\operatorname{DMR}

(AUROC, mean

\pm

std over 5 runs). Dark: bottom right singular vectors of

W_{U}

; middle: matched mid-spectrum subspace; random: random subspace of equal rank.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

	Llama-3.1-8B-It		Qwen3-8B
Layers	Rank	Sing. vectors	Rank	Sing. vectors
0–3	1636	1230–2865	1867	1114–2980
4–7	904	1596–2499	782	1657–2438
8–11	1174	1461–2634	205	1945–2149
12–15	954	1571–2524	2706	695–3400
16–19	760	1668–2427	3690	201–3890

Appendix

Table 5: Matched mid-spectrum ranks on WikiText-2. For each layer window, swapping within the indicated middle subspace reproduces the NLL increase of swapping the rank-205 dark subspace (bottom 5% of right singular vectors of

W_{U}

) within

\pm 0.01

Figure 7: Language-modeling NLL under dark residual swaps on the first and last four-layer windows. Blue: first window (layers 0–3); orange: last window (layers 32–35 for Qwen3-8B, 28–31 for Llama-3.1-8B-Instruct); grey: unablated model; pink: smaller reference model (Qwen3-0.6B and Llama-3.2-1B-Instruct). Boundary interventions raise NLL above the unablated model and, on most corpora, to or above the smaller model.

Figure 8: Synthetic induction dark swaps on Qwen3-8B with reduced dark subspace ranks (

d_{\mathrm{dark}}=64

and

128

). Bars show

\Delta

NLL on the repeated half against the start layer of each four-layer window; the dotted line is the average accuracy drop.

	Llama-3.1-8B-It		Qwen3-8B
Layers	Rank	Sing. vectors	Rank	Sing. vectors
0–3	1112	1492–2603	2161	967–3127
4–7	968	1564–2531	2107	994–3100
8–11	1200	1448–2647	3056	520–3575
12–15	1130	1483–2612	3270	413–3682
16–19	974	1561–2534	3085	505–3589

Appendix

Table 6: Matched mid-spectrum ranks on WikiText-2 (in-sample permutation). For each layer window, permuting within the indicated middle subspace reproduces the NLL increase of permuting the rank-205 dark subspace (bottom 5% of right singular vectors of

W_{U}

) within

\pm 0.01

Model	Type	Dataset	No ablation	L16–19		L24–27
				Middle	Dark	Middle	Dark
Qwen3-8B	Contextual	NQ-Swap	0.88	0.61	0.30	0.92	0.62
	Contextual	ConflictQA	0.98	0.98	0.69	0.99	0.78
	Parametric	Cities	1.0	0.63	0.99	1.0	1.0
		Sports	0.95	0.55	0.92	0.93	0.95
		Celebrity	0.96	0.62	0.93	0.96	0.96

Appendix

Table 7: Same comparison as Table 1 for layers 16–19 and 24–27. Dark subspace dimension 205.

Dataset	no ablation	L16–19	L20–23
NQ-Swap	0.88	0.82	0.76
ConflictQA	0.98	0.98	0.98
Cities	1.00	0.99	1.00
Sports	0.95	0.94	0.96
Celebrity	0.96	0.96	0.96

Appendix

Table 8: Qwen3-8B contextual/factual hit rate under dark residual permutation (dark rank 32). Hit = target continuation has lower mean NLL than the alternative.

Dataset	no ablation	L16–19	L20–23
NQ-Swap	0.88	0.77	0.63
ConflictQA	0.98	0.98	0.97
Cities	1.00	1.00	1.00
Sports	0.95	0.94	0.94
Celebrity	0.96	0.94	0.96

Appendix

Table 9: Qwen3-8B contextual/factual hit rate under dark residual permutation (dark rank 64). Hit = target continuation has lower mean NLL than the alternative.

Dataset	no ablation	L16–19	L20–23
NQ-Swap	0.88	0.60	0.37
ConflictQA	0.98	0.96	0.85
Cities	1.00	1.00	1.00
Sports	0.95	0.92	0.94
Celebrity	0.96	0.94	0.95

Appendix

Table 10: Qwen3-8B contextual/factual hit rate under dark residual permutation (dark rank 128). Hit = target continuation has lower mean NLL than the alternative.

	NQ-Swap	ConflictQA
Question	Who was the actor who played Crocodile Dundee?	Who is the author of How to Stop Worrying and Start Living ?
Context (incorrect answer injected)	Gyanendra of Nepal , AM (born 8 October 1939) is an Australian comedian, actor and television presenter. He was nominated for the Academy Award for Best Original Screenplay and won the Golden Globe Award for Best Actor — Motion Picture Musical or Comedy for his performance as outback adventurer Michael “Crocodile” Dundee in Crocodile Dundee (1986), the first in the Dundee film franchise.	Herbert Croly is indeed the author of How to Stop Worrying and Start Living . This can be confirmed by looking at the book’s cover page, which clearly states “By Herbert Croly” as the author. Additionally, a quick search on reputable book-selling websites such as Amazon or Barnes & Noble will also show that Herbert Croly is credited as the author of this book. Furthermore, Croly is a well-known author and political philosopher, having also written the influential book The Promise of American Life .
Injected incorrect answer	Gyanendra of Nepal	Herbert Croly
Parametric answer	Paul Hogan	Dale Carnegie

Appendix

Table 11: Illustrative knowledge conflicts from NQ-Swap and ConflictQA. In both datasets the provided context has been altered so that it supports an incorrect answer (bolded in the passage).

Dataset	Example question	Answer
Cities	Is the city of Tokyo in Japan?	Yes
	Is the city of Tokyo in Pakistan?	No
Celebrities	Is Leonardo DiCaprio an Academy Award-winning actor?	Yes
	Is Rihanna from Brazil?	No
Sports	Was Michael Jordan a famous basketball player?	Yes
	Was Usain Bolt a famous swimmer?	No

Appendix

Table 12: Example yes/no questions from the Cities, Celebrities, and Sports factual probes. Each dataset mixes true and false statements; the model is asked to answer Yes or No .

Model	Total	Hal. Rate (%)
Mistral-7B	749	39.4
Llama-2-7B	746	34.9
Llama-2-13B	730	32.3
Llama-3.1-8B	650	28.9
Qwen3-8B	750	24.3
gemma-3-4b	650	31.3

Appendix

Table 13: Hallucination statistics for XSum (annotations from Kuleshova et al. (2026) ).

LLM Hallucination Detection

Large language model hallucination detection identifies unsupported, incorrect, or fabricated claims in model outputs and traces.

See all 198 papers

Rethinking Faithfulness in LLMs: A Pairwise Context-Sensitive Perspective
Oct 6, 2026Zizhuo Zhang, Xiong Peng, Jingwei Sun +3Contextual Faithfulness LLM Evaluation
TMLR Group, Department of Computer Science, Hong Kong Baptist University · Huawei Noah’s Ark Lab.
Hallucinations as Orthogonal Noise: Inference-Time Manifold Alignment via Dynamic Contextual Orthogonalization
Jun 2, 2026Mingkuan Zhao, Wentao Hu, Tianchen Huang +6LLM Hallucination Mitigation Hallucination in Language Models
Xi’an Jiaotong University · Xingchen AGI Lab, China Telecom AI Technology (Beijing) Co., Ltd. · University of Science and Technology of China +2
Context-Fidelity Boosting: Enhancing Faithful Generation through Watermark-Inspired Decoding
Apr 24, 2026Weixu Zhang, Fanghua Ye, Qiang Gao +7Contextual Faithfulness Language Model Decoding
Hunyuan AI Digital Human, Tencent · McGill University · Mila - Quebec AI Institute +4

Oct 6, 2026J/K move · Enter open · S save

Rethinking Faithfulness in LLMs: A Pairwise Context-Sensitive Perspective

Zizhuo Zhang, Xiong Peng, Jingwei Sun, Rong Yao+2

TMLR Group, Department of Computer Science, Hong Kong Baptist University · Huawei Noah’s Ark Lab.

Contextual Faithfulness LLM Evaluation Hallucination in Language Models LLM Abstention

Open paper AskPDF arXiv

Large language models (LLMs) are expected to answer questions faithfully based on the provided context, abstaining when the context information is insufficient to answer the questions. Existing faithfulness evaluations typically assess each question-context instance in isolation; however, such instance-level evaluation fails to capture a fundamental requirement of faithful behavior: the ability to adapt model responses to changes in available contexts. In particular, a model should provide correct answers when sufficient evidence is present and abstain when it is not. In this work, we propose a Pairwise Faithfulness Benchmark (PFaithBench) that evaluates whether a model can switch between answering and abstaining for the same question under supporting versus non-supporting contexts. Our evaluations across thirty-nine models with seven model families demonstrate that faithfulness fundamentally involves a trade-off between answering and abstaining, and that most current models exhibit a strong bias toward answering, with most faithfulness errors arising from over-answering, i.e., models tend to fabricate a response even when the provided context is insufficient. We further conduct a series of studies on faithfulness training under different data constructions. Our results show that training outcomes are highly sensitive to the specific composition of answering and abstaining data. Constructing answering and abstaining data from mismatched sources can cause models to rely on dataset-specific shortcuts rather than actual context sufficiency. Moreover, increasing answer-supervised data improves answering performance but exacerbates over-answering, while increasing abstaining data reduces hallucination but leads to over-abstention. The code and data are released at https://github.com/tmlr-group/PFaithBench.

Abstract

LLM Hallucination Detection

Explore similar work

Rethinking Faithfulness in LLMs: A Pairwise Context-Sensitive Perspective

Hallucinations as Orthogonal Noise: Inference-Time Manifold Alignment via Dynamic Contextual Orthogonalization

Context-Fidelity Boosting: Enhancing Faithful Generation through Watermark-Inspired Decoding