AGSER: Attention-Guided Self-Reflection

• AGSER is a zero-shot hallucination detection technique that partitions input into attentive and non-attentive subqueries to evaluate output consistency.
• It leverages attention-derived token contributions, using pooling methods across layers to reduce computational overhead compared to traditional self-consistency sampling.
• Empirical results show AGSER improves AUC scores over baselines without relying on labeled data or external resources, enhancing practical deployment in critical domains.

Attention-Guided Self-Reflection (AGSER) is a zero-shot hallucination detection technique for LLMs that leverages internal model attention to partition input queries into distinct subcomponents and systematically assesses output consistency. AGSER achieves robust hallucination detection with lower computational overhead compared to traditional self-consistency or resampling approaches, operating without access to labeled hallucination data, external tools, or repeated random sampling (Liu et al., 17 Jan 2025).

1. Motivation and Hallucination Detection Challenge

LLMs often produce outputs that are syntactically plausible yet semantically incorrect—commonly known as hallucinations. This phenomenon severely limits trust in LLMs, especially within domains where factual precision is paramount, such as medicine, law, and finance. Existing hallucination detection mechanisms, like self-consistency sampling, involve generating multiple outputs per query to measure answer agreement, yet they are computationally intensive and may falter when models strongly repeat the same hallucinated content. Zero-shot detectors, which operate without additional training or supervision, are desirable for practical deployment due to their efficiency and independence from specialized labeled datasets (Liu et al., 17 Jan 2025).

AGSER addresses this gap by utilizing attention-derived token importance to create two targeted sub-queries per prompt. It then quantifies how each sub-query supports or fails to reproduce the original output, yielding a principled, attention-based hallucination score without additional fine-tuning or external resources.

2. Attention Contributions and Query Partitioning

AGSER first analyzes the contribution of each input token to the final LLM output via attention mechanisms:

Let the input query be $X = \{x_1, x_2, \dots, x_M\}$. For each self-attention layer $l$ ($1 \leq l \leq L$) with $H$ heads, the attention weights are

$$A^{l,h} = \mathrm{softmax} \left( \frac{(X^{l-1} W_Q^{l,h}) (X^{l-1} W_K^{l,h})^\top}{\sqrt{d_h}} \right)$$

where $d_h = d / H$. The contribution from token $j$ to the position $i$ at layer $l$ is

$$a_{i,j}^{\,l} = \sum_{h=1}^H A_{i,j}^{\,l,h} (x_j^{\,l-1} W_V^{l,h}) W_O^{l,h}$$

The influence of token $l$0 on the final position is summarized as $l$1.

To aggregate across layers, AGSER employs either mean-pooling ($l$2) or max-pooling ($l$3), generating a per-token contribution score $l$4. AGSER then sorts $l$5 in descending order and selects a fraction $l$6 (by default $l$7), designating tokens above threshold $l$8 as “attentive”:

$l$9

This yields two disjoint subsets that jointly sum to the full input token count.

3. Processing Pipeline and Consistency Computation

AGSER’s core mechanism involves three LLM passes:

1. Original pass: The full input $1 \leq l \leq L$0 is fed into the LLM, producing original output $1 \leq l \leq L$1.
2. Attentive pass: The attentive tokens $1 \leq l \leq L$2 are used to generate output $1 \leq l \leq L$3.
3. Non-attentive pass: The non-attentive tokens $1 \leq l \leq L$4 are input to obtain $1 \leq l \leq L$5.

To quantify answer resemblance, AGSER computes consistency scores as the ROUGE-L similarity (or a vector embedding similarity) between each sub-query output and the original answer:

$1 \leq l \leq L$6

where $1 \leq l \leq L$7 denotes a text embedding function.

The final hallucination score is:

$1 \leq l \leq L$8

with $1 \leq l \leq L$9 in typical usage.

This score positions high-confidence, factually grounded outputs (high $H$0, low $H$1) against likely hallucinations (low $H$2, high $H$3), providing a zero-shot, attention-based measurement.

4. Computational Efficiency and Resource Comparison

AGSER offers substantial resource savings compared to self-consistency-based methods. The latter typically require $H$4 full LLM evaluations per query (e.g., $H$5 samples plus the original), resampling the prompt in its entirety each time. In contrast, AGSER executes three total passes (original, attentive, non-attentive) per query, and token throughput is composed only of the aggregate original $H$6 tokens, split disjointly among sub-queries.

Resource savings include:

• Invocation count reduced from $H$7 to 3 (e.g., from 6 to 3 per query for $H$8).
• Token load reduced from $H$9 to approximately $$A^{l,h} = \mathrm{softmax} \left( \frac{(X^{l-1} W_Q^{l,h}) (X^{l-1} W_K^{l,h})^\top}{\sqrt{d_h}} \right)$$0, since no input token is duplicated between the attentive/non-attentive queries.
• Overall compute requirements drop by approximately 50% compared to $$A^{l,h} = \mathrm{softmax} \left( \frac{(X^{l-1} W_Q^{l,h}) (X^{l-1} W_K^{l,h})^\top}{\sqrt{d_h}} \right)$$1 sampling (Liu et al., 17 Jan 2025).

5. Empirical Evaluation and Comparative Results

AGSER was evaluated across four LLMs (Llama2-7B, Llama2-13B, Llama3-8B, Qwen2.5-14B) using three zero-shot benchmarks: Books (author/year lookup), Movies (cast lists), and Global Country Information (GCI). No hallucination labels were available in training; all detection was done zero-shot.

Detection performance was measured by Area Under ROC Curve (AUC) for hallucination score $$A^{l,h} = \mathrm{softmax} \left( \frac{(X^{l-1} W_Q^{l,h}) (X^{l-1} W_K^{l,h})^\top}{\sqrt{d_h}} \right)$$2:

Model	Dataset	AGSER AUC	Best Baseline (InterrogateLLM)	Δ (AGSER gain)
Llama2-7B	Books	0.859	0.819	+0.040
Llama2-7B	Movies	0.935	0.891	+0.044
Llama2-7B	GCI	0.974	0.961	+0.013
Llama2-13B	Avg	—	—	+0.028
Llama3-8B	Avg	—	—	+0.009
Qwen2.5-14B	Avg	—	—	+0.067

The average improvement over InterrogateLLM is +3.6 AUC points on Llama2-7B, with comparable gains for larger models (Liu et al., 17 Jan 2025).

Ablation studies reveal that full AGSER is necessary for highest performance: “attentive only” drops AUC by ~1%, while “non-attentive only” reduces AUC to approximately 0.57. Pooling attention across all layers achieves highest reliability (AUC ~0.89), while using only beginning, middle, or last layer attention yields lower AUCs.

6. Qualitative Analysis, Case Examples, and Limitations

AGSER’s qualitative behavior demonstrates correct flagging of both factual and hallucinated outputs. When presented with a factual query (e.g., “Who wrote Dreamcatcher, what year?”), the attentive sub-query suffices to reproduce the correct answer (high $$A^{l,h} = \mathrm{softmax} \left( \frac{(X^{l-1} W_Q^{l,h}) (X^{l-1} W_K^{l,h})^\top}{\sqrt{d_h}} \right)$$3), while the non-attentive query yields off-topic content (low $$A^{l,h} = \mathrm{softmax} \left( \frac{(X^{l-1} W_Q^{l,h}) (X^{l-1} W_K^{l,h})^\top}{\sqrt{d_h}} \right)$$4), resulting in a high $$A^{l,h} = \mathrm{softmax} \left( \frac{(X^{l-1} W_Q^{l,h}) (X^{l-1} W_K^{l,h})^\top}{\sqrt{d_h}} \right)$$5 (factual). In contrast, for a hallucinated answer (e.g., “Who wrote Final Stand, what year?” with a confident but incorrect answer), both sub-queries yield divergent or consistently incorrect outputs, shrinking the difference $$A^{l,h} = \mathrm{softmax} \left( \frac{(X^{l-1} W_Q^{l,h}) (X^{l-1} W_K^{l,h})^\top}{\sqrt{d_h}} \right)$$6 and properly flagging a hallucination.

However, AGSER can underperform on very brief prompts, where partitioned queries lose semantic coherence. Attention scores may not reliably separate tokens in short inputs, and outputs can become arbitrarily unstable, leading to spurious hallucination flags. AGSER is incompatible with closed-weight API models that do not expose attention tensors, unless proxy mechanisms become available.

7. Algorithm Summary

The AGSER procedure, as precisely reported, is:

1. Obtain LLM output $$A^{l,h} = \mathrm{softmax} \left( \frac{(X^{l-1} W_Q^{l,h}) (X^{l-1} W_K^{l,h})^\top}{\sqrt{d_h}} \right)$$7 for the full query.
2. Compute per-token contributions $$A^{l,h} = \mathrm{softmax} \left( \frac{(X^{l-1} W_Q^{l,h}) (X^{l-1} W_K^{l,h})^\top}{\sqrt{d_h}} \right)$$8 at each layer; pool across layers to produce $$A^{l,h} = \mathrm{softmax} \left( \frac{(X^{l-1} W_Q^{l,h}) (X^{l-1} W_K^{l,h})^\top}{\sqrt{d_h}} \right)$$9.
3. Threshold to select attentive/non-attentive token subsets based on top-$d_h = d / H$0 $d_h = d / H$1 values.
4. Obtain LLM outputs $d_h = d / H$2 for each sub-query.
5. Compute consistencies $d_h = d / H$3 with the original output.
6. Calculate and return hallucination score $d_h = d / H$4, typically with $d_h = d / H$5.

This process is executed per query without reliance on supervised data, grounding its decision solely in attention-derived partitioning and answer consistency (Liu et al., 17 Jan 2025).

---

AGSER establishes an efficient, attention-guided paradigm for zero-shot hallucination detection, leveraging internal model signals rather than expensive external ensembling or retraining. It sets a resource benchmark for future zero-shot detection methodologies in LLMs.