---
title: Context-based Proposition Maps (CPMs)
url: https://www.emergentmind.com/topics/context-based-proposition-maps-cpms
type: topic
---

# Context-based Proposition Maps (CPMs)

Searching arXiv for the provided paper and related uses of "Context-based Proposition Maps".
Context-based Proposition Maps (CPMs) are structured representations that map context into propositions so that downstream systems can reason over knowledge, protocol behavior, or anatomical relations while retaining constraints that flatter representations often discard. In current arXiv usage, the term appears in three technically distinct forms: PropRAG represents corpus knowledge as contextually rich propositions and traverses proposition paths in a graph for retrieval-augmented generation; "Learn, Check, Test — Security Testing Using Automata Learning and Model Checking" annotates learned Mealy machines with atomic propositions derived from protocol context for model checking; and PrPSeg encodes renal anatomical propositions in a universal proposition matrix that regularizes segmentation under partial labels [2504.18070] [2509.22215] [2402.19286].

## 1. Core idea and scope

In all three formulations, a CPM links contextual evidence to a proposition space that can be consumed by a reasoning or optimization procedure. In PropRAG, the context is textual and relational: propositions preserve conditions, provenance, temporal qualifiers, and $N$-ary relations, and retrieval is reframed as search over proposition paths rather than independent passages. In the security-testing formulation, the context is protocol-specific behavior observed at a communication interface: message types, parameters, status codes, roles, and timing outcomes are mapped to atomic propositions attached to states or temporary internal states. In PrPSeg, the context is anatomical prior knowledge: containment and exclusivity relations among renal structures are encoded as machine-readable propositions that constrain segmentation masks.

| Domain | CPM instantiation | Operational role |
|---|---|---|
| Retrieval-augmented generation | Contextually rich propositions plus proposition-induced graph structure | Multi-step retrieval via beam search and PPR |
| Automata learning and model checking | Conditions that map Mealy input/output context to atomic propositions | Security-property checking on annotated behavioral models |
| Renal pathology segmentation | Universal proposition matrix $M$ | Anatomy loss and partial-label supervision |

The literature suggests that “CPM” is not a single canonical data structure but a shared design principle: preserve context explicitly, represent it propositionally, and make the resulting propositions actionable in a downstream inference pipeline. That family resemblance is strong, but the concrete mathematical objects differ substantially across the three settings.

## 2. Proposition-path CPMs in retrieval-augmented generation

In PropRAG, CPMs are realized by representing corpus knowledge as contextually rich propositions and explicitly traversing multi-proposition paths in a graph whose edges preserve the natural co-occurrence and synonymy structure of entities within those propositions. A proposition $p_i$ has text $t_i$, an entity set $\mathcal{E}(p_i)\subseteq \mathcal{E}$, and an embedding $f(p_i)\in\mathbb{R}^d$; the query is a natural-language question $q$ with embedding $f(q)\in\mathbb{R}^d$. The graph $G=(V,E)$ has $V=V_{\text{entity}}\cup V_{\text{passage}}$. For each proposition $p$, entity clique edges connect every pair of distinct co-occurring entities, passage containment edges connect entities to source passages, and synonymy edges connect entity nodes when $\mathrm{sim}(\mathrm{emb}(e_i),\mathrm{emb}(e_j))\ge \tau_{\text{syn}}$, with $\tau_{\text{syn}}=0.8$ in experiments. This differs fundamentally from triple-based knowledge graphs: instead of fragmenting $N$-ary statements into binary predicates, PropRAG treats each proposition as a hyper-edge implemented as an entity clique, leaving temporal, conditional, and provenance cues accessible through the proposition text [2504.18070].

Retrieval is formulated as search for a proposition path
$$
P=(p_1,p_2,\dots,p_L),
$$
subject to an adjacency constraint through shared or synonymous entities, a length bound $L\le L_{\max}$, and a coherence constraint
$$
\mathrm{sim}(f(p_i),f(p_{i+1}))\ge \tau_{\text{coherence}},
$$
with $\tau_{\text{coherence}}=0.075$ used for pruning. Initial path scoring uses an averaged path embedding,
$$
f_{\text{avg}}(P)=\frac{1}{L}\sum_{i=1}^L f(p_i), \qquad \mathrm{Score}_{\text{avg}}(P,q)=\mathrm{sim}(f_{\text{avg}}(P),f(q)),
$$
and the top $M$ expansions per step are re-ranked by a concatenated-text embedding,
$$
t(P)=t_1\oplus t_2\oplus \dots \oplus t_L,\qquad \mathrm{Score}_{\text{concat}}(P,q)=\mathrm{sim}(f(t(P)),f(q)),
$$
with $M=40$. Beam search initializes from the top $N_{\text{prop}}$ propositions most similar to the query, keeps beam width $B$ partial paths, and stops at $L_{\max}$ or when no valid expansions remain; defaults are $B=5$ and typical $L_{\max}=3$. Two-stage retrieval couples this path search with Personalized PageRank: Stage 1 runs PPR on the full graph with damping $0.75$ to select top $K=100$ passages and induce a focused subgraph, while Stage 2 blends initial and path-derived entity seeds, forms a reset vector with $\lambda_{\text{passage}}=0.05$, and runs a final exploitative PPR on the subgraph with damping $0.45$. The passage ranking uses the standard PPR equation
$$
\pi=\alpha \hat{A}\pi + (1-\alpha)r,
$$
and the top $k_{\text{out}}=5$ passages are selected for answer generation.

The online retrieval process is entirely LLM-free: initial proposition retrieval, both PPR runs, subgraph induction, beam search, and final passage selection do not invoke generative LLMs. LLMs are used offline for entity and proposition extraction with Llama-3.3-Instruct, and post-retrieval for answer generation; NV-Embed-v2 is used for entities, propositions, and passages. Empirically, PropRAG reports zero-shot Recall@5 of $55.3\%$ on PopQA, $93.7\%$ on 2Wiki, $97.0\%$ on HotpotQA, $77.3\%$ on MuSiQue, and $78.1\%$ on NQ, alongside F1 scores including $52.4\%$ on MuSiQue, $74.8\%$ on 2Wiki, $76.3\%$ on HotpotQA, $56.7\%$ on PopQA, and $62.1\%$ on NQ. The reported gains are strongest on multi-hop datasets, and ablations show peak performance at $L_{\max}=3$, robust behavior for $B=3$–$5$, and a slight degradation or no gain when the coherence filter is removed. A concrete MuSiQue path links “Mantua Cathedral is a Roman Catholic cathedral dedicated to Saint Peter,” “St. Peter’s Basilica is located in Vatican City,” and “No successor was appointed to the post of Governor of Vatican City after Marchese Camillo Serafini’s death in 1952,” illustrating how the bridge proposition at depth 2 enables final evidence retrieval and the answer 1952.

## 3. CPMs as semantic annotation for learned protocol models

In the automata-learning and model-checking setting, CPMs bridge a learned Mealy machine and a Kripke-style proposition labeling. The learned behavioral model of the System Under Learning is a Mealy machine
$$
\mathcal{M}=(Q,\Sigma,\Omega,\delta,\lambda,q_0),
$$
with input alphabet $\Sigma$, output alphabet $\Omega$, transition function $\delta:Q\times \Sigma \to Q$, and output function $\lambda:Q\times \Sigma \to \Omega$. The model-checking target is a Kripke structure
$$
\mathcal{K}=(S,S_0,R,\mathcal{L}), \qquad \mathcal{L}:S\to 2^{AP},
$$
and CPMs provide the labeling function by mapping protocol-specific context to atomic propositions. The resulting annotated Mealy machine is formalized as
$$
\mathcal{MK}=(Q,q_0,\Sigma,\Omega,\delta,\lambda,\mathcal{L}),
$$
where states carry proposition labels while Mealy input/output labels are preserved. A CPM is specified by gain conditions $C_g$, loss conditions $C_l$, and temporary/internal conditions $C_\tau$. Each condition is a triplet
$$
c=\langle S_P,S_\Sigma,S_\Omega\rangle,
$$
where $S_P\subseteq P$ is the affected proposition set, $S_\Sigma\subseteq \Sigma$ is the trigger-input set, and $S_\Omega\subseteq \Omega$ is the trigger-output set. Gains add propositions to target states, losses determine when propositions cease to hold, and temporary conditions expose transition outcomes as checkable propositions by splitting a transition and inserting a temporary internal state $\tau$ [2509.22215].

The labeling algorithm first initializes all state labels to $\emptyset$, then applies gains on transitions, propagates propositions through loss-aware reachability until a fixpoint, and finally converts transition-defined outcomes into state labels via $\tau$ states. For a transition $q \xrightarrow{\sigma/\omega} q'$, if a temporary condition matches, the transition is replaced by
$$
q \xrightarrow{\sigma/\tau} \tau,\qquad \tau \xrightarrow{\epsilon/\omega} q',
$$
with
$$
\mathcal{L}(\tau):=\mathcal{L}(q)\cup S_P.
$$
Temporary propositions hold only in the inserted state. This makes LTL and CTL properties over transition outcomes expressible as state-label properties without discarding the original Mealy semantics. The annotated model is translated into Rebeca as a composition
$$
\mathcal{M}:=r_{\text{sys}} \, || \, r_{\text{env}},
$$
where the system actor stores the current learned state $q$ and AP booleans, and the environment actor stores temporary propositions from $C_\tau$, resets temporaries via a `req` method, and issues inputs non-deterministically. The paper further verifies code faithfulness by generating the full state space, collapsing the resulting LTS to a Mealy machine, and checking bisimilarity with the original learned model.

The generic security properties are protocol-agnostic and instantiated through CPM-provided APs such as `AUTH`, `PROT`, `ACCESSOK`, `PRIV`, `CRIT`, and `INVKEYOK`. Reported formulas include
$$
\square(\neg \mathrm{AUTH}\land \mathrm{PROT}\rightarrow \neg \mathrm{ACCESSOK}),
$$
for authentication,
$$
\square(\mathrm{PROT}\rightarrow \neg \mathrm{UREADOK}),
$$
for confidentiality, and
$$
\square(\neg \mathrm{INVKEYOK}),
$$
for key validity. The case studies use NFC/eMRTD and UDS. For eMRTD, gains include `AUTH` on input `BAC` with output `9000`, `DF` and `PROT` on `DF*` with `9000`, `EF` on `EF*` with `9000`, and `CRIT` on `EF_DG2, EF_DG3` with `9000`; temporaries include `UREADOK` and `READOK` on `RD_BIN` with `9000` and `SREADOK` and `READOK` on `SRD_BIN` with `9000`. For UDS, gains include `AUTH` on `SAWithKey` with positive response `67`, `EXT` on `ExtendedDiagnosticSession` with `5003`, and `PROG` on `ProgrammingSession` with `5002`; temporaries include `INVKEYOK, WRONGKEYOK` on `SAWithWrongKey` with `67`, `PROT` and `ACCESSOK, UACCESSOK` on `CheckASWBit` with `71`, `CRIT` on `RequestDownload` with `74`, and `UREADOK` on `Read*` with `62`. Reported checking results state that on two Austrian passports, authentication, confidentiality on protected EF, privilege levels, and key validity all held, whereas on the tested UDS ECU, authentication held, privilege levels held in practice, confidentiality was generally not applicable because UDS on CAN is unencrypted, and key validity was violated because `SAWithWrongKey` was accepted.

## 4. CPMs as anatomical priors in panoramic renal pathology segmentation

In PrPSeg, a CPM is a structured, machine-readable representation of class-level propositions that encode anatomical context, including part-of or containment, mutual exclusivity, and optionally adjacency, proximity, co-occurrence, and hierarchical relationships. The reported instantiation uses a universal proposition matrix
$$
M\in \mathbb{Z}^{K\times K},
$$
with $K=8$ in the main study, to encode strong affirmative and negatory relations inspired by Aristotelian categorical propositions. Matrix entries are interpreted as
$$
M(i,j)=+1 \text{ if } i\subseteq j,\qquad
M(i,j)=-1 \text{ if } i\supseteq j,
$$
$$
M(i,j)=2 \text{ if } i\cap j=\varnothing,\qquad
M(i,j)=0 \text{ otherwise}.
$$
The eight modeled classes span three scales: regions (`cortex`, `medulla`), functional units (`glomerular capsule`, `glomerular tuft`, `distal tubule`, `proximal tubule`), and cells (`podocytes`, `mesangial cells`). Encoded relations include `Cortex ⊇ {Cap., Tuft, DT, PT}`, `Tuft ⊇ {Pod., Mes.}`, `Cap. ⊇ Tuft`, and `Medulla ∩ Cortex = ∅`. Although the general proposition tensor can support multiple relation types,
$$
\mathbf{P}\in \mathbb{R}^{K\times K\times R},
$$
with $R=3$ in the current instantiation, the reported implementation focuses on containment and exclusivity because these are presented as the most reliable relations for pathology masks across scales [2402.19286].

The training objective couples ordinary segmentation supervision with an anatomy loss derived from the proposition matrix. For a labeled class $i$ with ground-truth mask $Y_i\in\{0,1\}^{H\times W}$ and predicted probability maps $\hat{Y}_j\in [0,1]^{H\times W}$, the soft Dice loss is
$$
\mathrm{DCE}(A,B)=1-\frac{2\sum_p A_pB_p+\epsilon}{\sum_p A_p+\sum_p B_p+\epsilon}.
$$
PrPSeg defines relational penalties
$$
\phi_{\supset}(i,j)=\mathrm{DCE}(1-Y_i,\hat{Y}_j),
$$
for the constraint $i\supseteq j$,
$$
\phi_{\subset}(i,j)=\mathrm{DCE}(Y_i,\hat{Y}_j),
$$
for $i\subseteq j$, and
$$
\phi_{\perp}(i,j)=\mathrm{DCE}(Y_i,\hat{Y}_j),
$$
for exclusivity. The anatomy loss for class $i$ is
$$
\mathcal{L}_{\text{anat}(i)}=\sum_{j\neq i}\Big(
\mathbb{I}_{\supset}(i,j)\phi_{\supset}(i,j)+
\mathbb{I}_{\subset}(i,j)\phi_{\subset}(i,j)+
\mathbb{I}_{\perp}(i,j)\phi_{\perp}(i,j)
\Big),
$$
and the total objective is
$$
\mathcal{L}(i)=\mathcal{L}_{\text{seg}(i)}+\lambda_{\text{upl}}\mathcal{L}_{\text{anat}(i)},
$$
with $\lambda_{\text{upl}}=0.1$ in the experiments. Training proceeds in two phases: 50 epochs with $\mathcal{L}_{\text{seg}}$ only, then joint optimization with the anatomy loss.

The architecture is a token-based dynamic head single-network design built on a Residual U-Net encoder-decoder. Class-aware tokens $T_c\in \mathbb{R}^{K\times d}$ and scale-aware tokens $T_s\in \mathbb{R}^{M\times d}$, with $M=4$ magnifications $(5\times,10\times,20\times,40\times)$, condition the encoder and a lightweight dynamic head. The controller generates parameters for a 3-layer dynamic head with channels $(8,8,2)$ and 162 parameters total. CPMs do not gate the feature path in the reported implementation; instead, they modulate optimization through $\mathcal{L}_{\text{anat}}$, which supplies “virtual supervision” for unlabeled classes in partially labeled images. Reported datasets include PAS-stained region patches at $5\times$ and $1024\times 1024$, NEPTUNE functional-unit ROIs at $40\times$ resized to $256\times 256$, and PAS cell images at $20\times$ and $512\times 512$, with a $6{:}1{:}3$ train/validation/test split at patient level. PrPSeg reports average Dice $78.66$, with class-wise Dice of `Medulla 72.38`, `Cortex 72.64`, `DT 72.45`, `PT 85.27`, `Cap. 94.23`, `Tuft 94.40`, `Pod. 70.98`, and `Mes. 66.96`; the corresponding average for Omni-Seg is $76.66`. An ablation reports `Omni-Seg → +Token-based dynamic head: 76.66 → 77.89`, `Swin-UNETR → +Token-based dynamic head: 62.38 → 70.97`, and `PrPSeg (Token-based dynamic head + universal proposition learning via anatomy loss): 78.66`. The appendix further states that with seven added classes, superior performance is maintained by updating only tokens and $M$.

## 5. Formal patterns, distinctions, and common misconceptions

The most important distinction is that CPMs are not tied to a single representation family. In PropRAG, the central object is a proposition-induced graph whose proposition semantics are carried by text embeddings, entity cliques, synonymy edges, and path scoring; a proposition is not necessarily a graph node. In the security-testing formulation, the central object is a rule system over Mealy transitions that computes a labeling function $\mathcal{L}:Q\to 2^P$ and introduces $\tau$ states to expose transition outcomes. In PrPSeg, the central object is a signed relation matrix $M$ that converts anatomical priors into overlap-based regularization. Treating CPMs as synonymous with triple-based knowledge graphs, with state labeling alone, or with matrix priors alone would therefore be inaccurate [2504.18070] [2509.22215] [2402.19286].

A related misconception is that CPMs are merely descriptive annotations. In all three papers, the propositions are operational. PropRAG uses them for beam search, coherence filtering, and PPR-based ranking; the security-testing approach uses them to instantiate generic LTL properties and generate counterexample-driven tests; PrPSeg uses them to impose anatomy loss and exploit partially labeled data. Another misconception is that context is ancillary. Each paper treats context preservation as the main technical reason to use propositions rather than flatter alternatives: PropRAG contrasts proposition text and entity cliques with triple fragmentation and context collapse; the security-testing work makes protocol semantics checkable by embedding status-code and message context into AP labels; and PrPSeg turns anatomical containment and exclusivity into computational constraints rather than leaving classes independent.

The literature also suggests a common structural template despite the domain differences. First, define a proposition vocabulary appropriate to the domain. Second, specify how contextual evidence activates, propagates, or scores propositions. Third, integrate the resulting propositional structure into a downstream objective: retrieval and ranking, model checking, or segmentation optimization. This suggests that CPMs are best understood as context-preserving proposition interfaces between raw domain evidence and formal downstream procedures.

## 6. Robustness, limitations, and future directions

Each CPM formulation has domain-specific failure modes. In PropRAG, reported failures include spurious paths, where high local similarity chains unrelated propositions, semantic drift on longer paths, and incomplete propositions when entity extraction misses key entities. The paper mitigates these issues with the coherence threshold $\tau_{\text{coherence}}$, a best-performing cap of $L_{\max}=3$, balanced seeding through $\max(s'_{\text{init}},s'_{\text{beam}})$, and a two-stage exploration/exploitation schedule using PPR damping $0.75$ and $0.45$; it also presents non-parametric retrieval as a mechanism for avoiding catastrophic forgetting, since updates can be made by adding new propositions and edges offline without retraining the reader [2504.18070].

In the security-testing setting, the trade-offs are different. CPM design must reflect protocol semantics accurately; otherwise, gains, losses, or temporaries may mislabel the model. The insertion of temporary $\tau$ states slightly increases state space, confidentiality properties may be inapplicable for unencrypted bus protocols such as UDS over CAN, and some roles such as higher-level security may be hypothetical if not supported by the System Under Learning. The reported outlook includes broader adapters for more protocols, additional generic properties, timing-aware learning and CPMs for timed properties, automated CPM and property generation from threat models, and dynamic input alphabet extension via mutation [2509.22215].

In PrPSeg, the central risks arise from prior mis-specification rather than graph drift or labeling semantics. Incorrect containment or exclusivity relations can bias training; conflicting rules or cycles in the hierarchy reduce consistency; domain shift across stains or scanners can degrade performance; and adjacency or proximity constraints require reliable boundaries before they should be introduced as soft penalties. The reported mitigation strategies are expert review, transitivity checks, stain augmentation, scale tokens, and optional reweighting of CPM penalties. The adaptation guideline is direct: for other anatomies, create a new class list, encode strong containment and exclusivity first, add adjacency or proximity only when annotations support it, and update the proposition matrix and class tokens without changing the backbone [2402.19286].

Taken together, these works show CPMs as a general strategy for preserving domain context in propositional form, but not as a single standardized formalism. Their common contribution is methodological rather than representational uniformity: context is encoded explicitly, propositionally, and in a form that can drive search, verification, or learning.

Source: https://www.emergentmind.com/topics/context-based-proposition-maps-cpms