Paper: arXiv 2512.05156
Authors: Igor Halperin
Abstract
Evaluating faithfulness of Large Language Models (LLMs) to a given task is a complex challenge. We propose two new unsupervised metrics for faithfulness evaluation using insights from information theory and thermodynamics. Our approach treats an LLM as a bipartite information engine where hidden layers act as a Maxwell demon controlling transformations of context $C $ into answer \(A\) via prompt \(Q\). We model Question-Context-Answer (QCA) triplets as probability distributions over shared topics. Topic transformations from \(C\) to \(Q\) and \(A\) are modeled as transition matrices \({"\bf Q"}\) and \({"\bf A"}\) encoding the query goal and actual result, respectively. Our semantic faithfulness (SF) metric quantifies faithfulness for any given QCA triplet by the Kullback-Leibler (KL) divergence between these matrices. Both matrices are inferred simultaneously via convex optimization of this KL divergence, and the final SF metric is obtained by mapping the minimal divergence onto the unit interval [“0,1”], where higher scores indicate greater faithfulness. Furthermore, we propose a thermodynamics-based semantic entropy production (SEP) metric in answer generation, and show that high faithfulness generally implies low entropy production. The SF and SEP metrics can be used jointly or separately for LLM evaluation and hallucination control. We demonstrate our framework on LLM summarization of corporate SEC 10-K filings.
Complexity vs Empirical Score
- Math Complexity: 7.5/10
- Empirical Rigor: 6.0/10
- Quadrant: Holy Grail — high math complexity, high empirical rigor
Why this score: The paper employs advanced mathematical concepts including KL divergence, convex optimization, and thermodynamic entropy production, placing it in the high math complexity range. However, its empirical demonstration is limited to 10 SEC 10-K examples without code, backtesting, or statistical validation, placing it just above the threshold for high empirical rigor.
Research Flowchart
flowchart TD A["Research Goal:<br>Evaluate LLM Faithfulness<br>without Supervision"] --> B["Data & Input:<br>Question-Context-Answer QCA triplets<br>from SEC 10-K Summarization"] B --> C["Methodology:<br>Model QCA as probability distributions<br>and transition matrices via convex optimization"] C --> D["Computational Processes:<br>1. Semantic Faithfulness SF<br>via KL Divergence mapping to [0,1]<br>2. Semantic Entropy Production SEP<br>thermodynamic analysis]<br>D --> E['Key Outcomes:<br>SF & SEP metrics enable<br>unsupervised hallucination control<br>High Faithfulness correlates with Low Entropy"]
Evidence Card
What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card
Proposes
The paper proposes two new unsupervised metrics for evaluating LLM faithfulness: the Semantic Faithfulness (SF) metric based on KL divergence between topic transition matrices, and the Semantic Entropy Production (SEP) metric derived from stochastic thermodynamics.
Source sentence
We propose two new unsupervised metrics for faithfulness evaluation using insights from information theory and thermodynamics. Our semantic faithfulness (SF) metric quantifies faithfulness for any given QCA triplet by the Kullback-Leibler (KL) divergence between these matrices.
Assumes 6
- The LLM is treated as a black box with no access to internal activations, log‑probs, or other internal information. model
Source sentence
The black-box setting of our formulation implies that we do not have access to additional information such as internal activations, log-probs etc.
- All questions Qk are constructed as semantically equivalent paraphrases of an initial question Q. data
Source sentence
we assume a more general setting where all questions Qk are constructed as semantically equivalent paraphrases of a given initial question Q. As discussed in [9], prompt paraphrasing is useful to enrich the data and extract more meaningful topic representations.
- The only available data for analysis are the Question‑Context‑Answer triplets (Qk, C, Ak). data
Source sentence
In what follows, we assume that triplets (Qk, C, Ak) constitute the only available data for our analysis.
- Topic distributions over shared topics can be represented as probability distributions and modeled with transition matrices Q and A. model
Source sentence
We model Question-Context-Answer (QCA) triplets as probability distributions over shared topics. Topic transformations from $C$ to $Q$ and $A$ are modeled as transition matrices ${\bf Q}$ and ${\bf A}$ encoding the query goal and actual result, respectively.
Show 2 more
- The convex optimization problem for jointly inferring Q and A has a unique global optimum and converges reliably. statistical
Source sentence
Our proposed faithfulness metric, which we call the Semantic Faithfulness (SF) score and denote as FS, is defined in terms of the minimal value of this KL divergence (4), obtained by jointly optimizing over A and Q subject to all constraints: FS := 1 1 + Dmin . Importantly, both optimization problems are convex with guaranteed convergence to global optima, as established by the Csisz´ar-Tusn´ady alternating
- Embedding and clustering of text into a fixed number of semantic topics (N=23) adequately captures the semantic structure needed for the metrics. other
Source sentence
Clustering: The joint embedding space was clustered into N = 23 semantic topics using the Upper-Bounded Deterministic Information Bottleneck (UDIB) algorithm [10] in a single clustering pass.
Data
- NVIDIA 2024 Annual Report Risk Factors (Form 10‑K) section text-news public
Source sentence
The context for all QCA triplets was extracted from NVIDIA Corpora-tion’s fiscal year 2024 Annual Report (Form 10-K), specifically the comprehensive Risk Factors section (Item 1A).
Costs
Not applicable
Finds 5
- Both groups of questions show similar mean Semantic Faithfulness (Group A FS = 0.509, Group B FS = 0.503), indicating question type alone does not strongly determine faithfulness. statistical
Source sentence
Both groups exhibit similar mean Semantic Faith-fulness (Group A: FS = 0.509, Group B: FS = 0.503), indicating that question type alone does not strongly determine faithfulness in this dataset.
- Individual triplet FS values range from 0.472 to 0.577, suggesting specific question‑context pairings matter more than broad categories. statistical
Source sentence
However, individual triplet variation within groups (overall FS range: [0.472, 0.577]) suggests that specific question-context pairings matter more than broad question categories.
- Positive correlation between question entropy H(Q) and SF (Pearson r = 0.695, p = 0.026). statistical
Source sentence
We observe a positive correlation between H(Q) and FS (Pearson r = 0.695, p = 0.026), see Fig. 2.
- Mean Semantic Entropy Production across all triplets is 0.287 bits (range 0.076–0.771 bits), with Group A showing higher SEP (0.359 bits) than Group B (0.215 bits). statistical figure from elsewhere in the paper
Source sentence
Group A exhibits higher SEP (mean = 0.359 bits) than Group B (mean = 0.215 bits), consistent with Group A’s higher system entropy change.
- Correlation between FS and SEP is negative (r = ‑0.612, p = 0.060). statistical
Source sentence
The correlation between FS and SEP is r = −0.612 (p = 0.060), see Fig. 3.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 18 related papers). Coverage is partial, so this is neither replication nor refutation.
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.
Cited References (17)
- Topic Identification in LLM Input-Output Pairs through the Lens of Information Bottleneck (2025) — link
- Detecting hallucinations in large language models using semantic entropy (2024) — link
- Judging LLM-as-a-judge with MT-Bench and Chatbot Arena (2023) — link
- Energy and information flows in autonomous systems (2022) — link
- Unified framework for the entropy production and the stochastic interaction based on information geometry (2018) — link
- Thermodynamics from Information (2018) — link
- Information geometry and its applications (2012) — link
- Stochastic thermodynamics, fluctuation theorems and molecular machines (2012) — link
and 9 more
- Fluctuation Relations: A Pedagogical Overview (2012) — link
- Entropy production along a stochastic trajectory and an integral fluctuation theorem. (2005) — link
- Information Theory and Statistics: A Tutorial (2004) — link
- ROUGE: A Package for Automatic Evaluation of Summaries (2004)
- Bleu: a Method for Automatic Evaluation of Machine Translation (2002) — link
- Computation of channel capacity and rate-distortion functions (1972) — link
- Information Theory: 1948-1998 - Guest Editorial (1998)
- “An Algorithm to Compute the Capacity of Discrete Memoryless Channels” (1972)
- “Prompt-Response Semantic Divergence Metrics for Faithfulness Hal-lucination Detection in Large Language Models”
0 of 17 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.