---
title: 'M-CIRR: Molecular & Image Retrieval'
url: https://www.emergentmind.com/topics/m-cirr
type: topic
---

# M-CIRR: Molecular & Image Retrieval

Searching arXiv for the cited papers and closely related entries to ground the article.
M-CIRR is an overloaded research abbreviation whose meaning depends on disciplinary context. In molecular communication, it denotes **Molecular Channel Impulse Response Reconstruction**, namely the problem of estimating the discrete-time impulse response of a diffusive molecular channel from a short pilot burst and molecule-count observations; related molecular-communication work also uses the label for exact analytical channel impulse responses in drifted geometries [1608.03745], [1510.08612], [2512.04858]. In composed image retrieval, M-CIRR denotes an **instruction-rich extension of the CIRR dataset** for multi-modification retrieval, and the term is also used informally for the multi-turn CIRR component of the CIRCLED benchmark [2604.21806], [2605.26734]. The shared acronym therefore names distinct objects: an inference framework and channel model family in nanoscale communication, and a dataset family in multimodal retrieval.

## 1. Terminological scope and disciplinary uses

The acronym appears in multiple arXiv literatures with different expansions and technical roles. This suggests that M-CIRR is not a single standardized term across fields, but a context-dependent shorthand.

| Usage | Meaning | Representative source |
|---|---|---|
| Molecular communication | Molecular Channel Impulse Response Reconstruction | [1608.03745] |
| Molecular communication | Exact molecular channel impulse response for drifted spherical reception | [2512.04858] |
| Composed image retrieval | Multi-modification extension of CIRR | [2604.21806] |
| Multi-turn composed image retrieval | Informal name for the CIRR component of CIRCLED | [2605.26734] |

In the molecular-communication lineage, the core object is the **channel impulse response (CIR)**, defined as the expected number of molecules observed at the receiver over time after the transmitter instantaneously releases molecules. The reconstruction problem is training-based and statistical: a known emission sequence is transmitted, molecule counts are observed, and the unknown CIR vector is estimated [1608.03745]. In the retrieval lineage, the core object is a **triplet dataset** \(\mathcal{T}=\{(x_r^{(i)}, t_m^{(i)}, x_t^{(i)})\}_{i=1}^N\), where the modification text is expanded into explicit, clause-by-clause instructions describing every visual change from reference to target [2604.21806].

## 2. Molecular communication: M-CIRR as channel inference

In diffusive molecular communication, time is discretized into symbol intervals indexed by \(k\). At the start of interval \(k\), the transmitter emits \(s[k]\cdot N^{\rm Tx}\) molecules, with \(N^{\rm Tx}\) fixed and \(0\le s[k]\le 1\). The receiver counts \(r[k]\) molecules. With channel memory \(L\) and background noise, the input-output law is
\[
r[k] = \sum_{\ell=1}^L c_\ell[k] + c_n[k],
\]
where \(c_\ell[k]\sim \mathrm{Poiss}(s[k-\ell+1]\cdot \bar c_\ell)\) and \(c_n[k]\sim \mathrm{Poiss}(\bar c_n)\) [1608.03745].

By linearity of expectations, the discrete-time CIR vector is
\[
\bar{\mathbf c} = [\bar c_1,\dots,\bar c_L,\bar c_n]^T,
\]
with
\[
h[k] = E\{r[k]\} = \sum_{\ell=1}^L s[k-\ell+1]\cdot \bar c_\ell + \bar c_n.
\]
The continuous-time counterpart is the expected number of received molecules after an instantaneous release at \(t=0\). For free space, one example is
\[
h_{\rm free}(t) = N_0\cdot (4\pi D t)^{-3/2}\cdot \exp\!\bigl(-d^2/(4Dt)\bigr),
\]
and earlier work expresses the continuous-time CIR as the integral of the expected concentration over the receiver volume \(V^{\rm rec}\) [1510.08612].

The estimation problem is posed over a known pilot. Counts \(r[L],\dots,r[K]\) are collected and stacked into \(\mathbf r\), with design vectors
\[
\mathbf s_k = [s[k],s[k-1],\dots,s[k-L+1],1]^T.
\]
Under the assumption of independent Poisson observations,
\[
r[k]\sim \mathrm{Poiss}(\bar{\mathbf c}^T\mathbf s_k),
\]
which yields the likelihood
\[
L(\bar{\mathbf c}) = \prod_{k=L}^K \frac{(\bar{\mathbf c}^T\mathbf s_k)^{\,r[k]} e^{-\bar{\mathbf c}^T\mathbf s_k}}{r[k]!}.
\]
This formulation makes M-CIRR a Poisson inverse problem with nonnegativity constraints and finite channel memory [1608.03745].

## 3. Estimators, bounds, and pilot design in molecular M-CIRR

The non-Bayesian estimators are **maximum likelihood (ML)** and **least sum of square errors (LSSE)**. The ML estimator maximizes
\[
\ell(\bar{\mathbf c}) = \sum_{k=L}^K \bigl[r[k]\ln(\bar{\mathbf c}^T\mathbf s_k)-(\bar{\mathbf c}^T\mathbf s_k)\bigr]
\]
subject to \(\bar{\mathbf c}\ge 0\), and its stationary condition is
\[
\sum_{k=L}^K \Bigl[\frac{r[k]}{\bar{\mathbf c}^T\mathbf s_k}-1\Bigr]\mathbf s_k = 0.
\]
In practice, one tests the unconstrained root and, if some components are negative, searches over lower-dimensional faces by fixing zero components [1608.03745].

The LSSE estimator matches empirical means through
\[
\mathbf r \approx S\,\bar{\mathbf c},
\qquad
S=[\mathbf s_L,\mathbf s_{L+1},\dots,\mathbf s_K]^T,
\]
and solves
\[
\hat{\bar{\mathbf c}}_{\rm LSSE} = \arg\min_{\bar{\mathbf c}\ge 0}\|\mathbf r-S\bar{\mathbf c}\|^2.
\]
Dropping nonnegativity gives the normal equation
\[
(S^TS)\bar{\mathbf c}=S^T\mathbf r
\quad \Rightarrow \quad
\bar{\mathbf c}=(S^TS)^{-1}S^T\mathbf r.
\]
Again, if some components are negative, they are set to zero and the reduced system is re-solved [1608.03745], [1510.08612].

When statistical channel knowledge is available, the framework adds **maximum a posteriori (MAP)** and **linear minimum mean square error (LMMSE)** estimation. The MAP estimator maximizes
\[
\ell(\bar{\mathbf c})+\ln f_c(\bar{\mathbf c})
\]
under \(\bar{\mathbf c}\ge 0\). The LMMSE estimator seeks a linear mapping \(\hat c = F r\) that minimizes \(E\{\|\bar c-Fr\|^2\}\), yielding
\[
F_{\rm LMMSE} = \Phi_{cc}S^T\bigl[S\Phi_{cc}S^T+\mathrm{diag}(S\mu_c)\bigr]^{-1},
\]
which depends only on first- and second-order prior statistics [1608.03745].

Performance is benchmarked by classical and Bayesian Cramér–Rao bounds. For deterministic \(\bar c\), the Fisher information matrix is
\[
I(\bar c)=\sum_{k=L}^K \frac{\mathbf s_k\mathbf s_k^T}{\bar c^T\mathbf s_k},
\]
so any unbiased estimator satisfies \(\mathrm{Cov}(\hat c)\ge I^{-1}\). With a prior, the Bayesian Fisher information adds the prior term,
\[
I_B = I_{\rm classical} + E_c\{-\nabla^2\ln f_c(c)\},
\]
and \(\mathrm{Cov}\ge I_B^{-1}\) [1608.03745].

Pilot-sequence design chooses \(s[1\ldots K]\in\{0,1\}^K\) to reduce estimation error. The LSSE-oriented criterion minimizes an MSE proxy while keeping \(S^TS\) well-conditioned; the LMMSE-oriented criterion maximizes
\[
\mathrm{tr}\!\left\{\Phi_{cc}S^T\left[S\Phi_{cc}S^T+\mathrm{diag}(S\mu_c)\right]^{-1}S\Phi_{cc}\right\}.
\]
A simple suboptimal alternative is the **ISI-free** pilot that emits a \(1\) every \(L+1\) intervals and zeros otherwise, so that each tap appears in isolation and can be estimated by averaging [1608.03745], [1510.08612].

The practical algorithm is correspondingly modular: choose a training sequence offline, transmit the \(K\)-symbol pilot, record counts \(r[L\ldots K]\), compute \(\hat{\bar{\mathbf c}}\) via LSSE, ML, MAP, or LMMSE, and then use the estimate for equalization or detection. The reported trade-off is that ML and MAP give best accuracy at high computational cost, while LSSE and LMMSE are much simpler; the ISI-free sequence plus averaging may be the only feasible choice for nanoscale nodes with very limited processing [1608.03745].

## 4. Exact analytical M-CIRR for drifted spherical receivers

A later molecular-communication use of the acronym concerns the **exact molecular channel impulse response** for a fully absorbing spherical receiver in three dimensions under uniform drift with arbitrary direction [2512.04858]. Here the receiver has radius \(R\), is located at the origin, the transmitter is at distance \(r_0\), the drift has speed \(v\), and the drift direction forms angle \(\theta\) relative to the transmitter–receiver axis.

The molecule position satisfies the Itô SDE
\[
d\mathbf{X}_t=\mathbf{v}\,dt+\sigma\,d\mathbf{B}_t,\qquad \sigma^2=2D,
\]
and Girsanov’s theorem is used to transform the no-drift hitting-time law into the drifted one. If \(T=\inf\{t>0:|\mathbf X_t|=R\}\), the drifted joint density obeys
\[
f^{(v)}(t,y)
=
\exp\!\Bigl(
\tfrac{\mathbf v\cdot(y-\mathbf x_0)}{\sigma^2}
-\tfrac{|\mathbf v|^2 t}{2\sigma^2}
\Bigr)
f_0(t,y).
\]
The CIR is obtained by marginalizing over the sphere,
\[
h(t;v,\theta,r_0,R)=\int_{|y|=R} f^{(v)}(t,y)\,dS(y).
\]

A key step is a surface-integral identity involving Legendre polynomials \(P_m(\cos\theta)\) and modified Bessel functions \(I_{m+1/2}\). The resulting exact closed form is
\[
h(t;v,\theta,r_0,R)
=
\exp\!\Bigl(
-\tfrac{|v|r_0\cos\theta}{2D}
-\tfrac{|v|^2 t}{4D}
\Bigr)
\frac{R}{r_0}
\sum_{m=0}^{\infty}
\alpha_m
\int_0^\infty
\frac{\lambda\,
Z_{m+\tfrac12}\!\bigl(\tfrac{r_0}{R},\tfrac{\lambda R}{\sqrt{2D}}\bigr)}
{J_{m+\tfrac12}^2\!\bigl(\tfrac{\lambda R}{\sqrt{2D}}\bigr)+
Y_{m+\tfrac12}^2\!\bigl(\tfrac{\lambda R}{\sqrt{2D}}\bigr)}
e^{-D\lambda^2 t}\,d\lambda,
\]
with
\[
\alpha_m=(2m+1)\,I_{m+\tfrac12}\!\bigl(\tfrac{|v|R}{2D}\bigr)\,P_m(\cos\theta).
\]
The peak time \(t_{\rm peak}\) is characterized by the exact condition \(\frac{d}{dt}h(t)\vert_{t=t_{\rm peak}}=0\), and once \(t_{\rm peak}\) is found numerically, the peak amplitude is \(h_{\max}=h(t_{\rm peak};v,\theta,r_0,R)\) [2512.04858].

This formulation is presented as exact for arbitrary \(\theta\in[0,\pi]\), in contrast to prior off-axis approximations that project the flow onto the transmitter–receiver line or use one-dimensional formulas irrespective of drift direction. A plausible implication is that this analytical M-CIRR complements statistical reconstruction: the former supplies a high-fidelity forward model, while the latter estimates unknown discrete-time taps from data.

## 5. M-CIRR in composed image retrieval: multi-modification benchmark

In composed image retrieval, M-CIRR is an instruction-rich extension of the original CIRR dataset, tailored for **multi-modification composed image retrieval** [2604.21806]. It is defined as
\[
\mathcal{T}=\bigl\{(x_r^{(i)},\,t_m^{(i)},\,x_t^{(i)})\bigr\}_{i=1}^{N},
\]
where \(x_r\) is a reference image, \(x_t\) is the target image, and \(t_m\) is a **multi-modification text (MMT)** that decomposes, in an explicit clause-by-clause manner, every visual change needed to transform \(x_r\) into \(x_t\).

The dataset is constructed from the standard CIRR benchmark, which contains **21 552 real-world images with 28 225 training and 4 181 validation triplets**. Each original short modification text is replaced through a three-stage pipeline. First, **Automatic Generation** uses Llama 3.2 with a carefully crafted prompt to “faithfully expand” the original description into a sequence of fine-grained, numbered clauses. Second, **Hallucination Filtering** uses GPT-4o’s hallucination detector to flag and remove any clause referring to objects or attributes not present in the reference or target image. Third, **Human Quality Check** is performed by a team of ten annotators under four criteria—Consistency, Accuracy, Diversity, and Quality—and then a final **reference-grounding** filter removes any clause that does not explicitly compare reference to target [2604.21806].

The resulting corpus contains \(N=28{,}225\) training and \(4{,}181\) validation triplets. Modification length ranges from **35 to 468 tokens** with mean \(\bar L=319.4\). Each MMT contains on average **4.2 clauses**, with **20%** having two clauses, **45%** three–four clauses, and **35%** five or more. Annotators identified **15 400 unique entity mentions**, and each query addresses on average **3.1 distinct entities**. A specialized **fine-discrimination subset** \(R_{\text{subset}}\) contains **1 200 queries** whose negative candidates are visually confusable [2604.21806].

The dataset schema stores each example as JSON lines with fields including `ref_id`, `tgt_id`, `mmt`, `tokens`, `clauses`, and `entities`. The `clauses` field is an array of objects, each containing the raw clause string and the list of entities it modifies, while the ordering of `clauses` preserves the original instruction sequence [2604.21806].

Relative to the original CIRR modification texts of **2–50 tokens** with mean \(\bar L=12.8\), M-CIRR expands each description into **35–468 tokens** with explicit coverage of **3–5 entities on average**. The stated motivation is to address **Insufficient Entity Coverage** and **Clause–Entity Misalignment** [2604.21806].

## 6. Evaluation, baselines, and relation to multi-turn CIRR extensions

M-CIRR follows CIRR’s standard recall protocol on validation. The principal metric is
\[
\mathrm{R}@K=\frac{1}{N}\sum_{i=1}^{N}\mathbf 1\{r_i\le K\},
\]
where \(r_i\) is the rank of the true target for query \(i\). For the fine-discrimination subset, the reported metrics are \(\mathrm{R}_{\mathrm{subset}@k}\) for \(k=1,2\), and the overall score is
\[
\mathrm{Avg}=\tfrac12\bigl(\mathrm{R}@5+\mathrm{R}_{\mathrm{subset}@1}\bigr).
\]
Representative results with BLIP backbones report **Candidate** at \(R@1=42.03\), \(R@5=75.92\), \(R@10=84.61\), \(R_{\mathrm{subset}@1}=69.58\), \(R_{\mathrm{subset}@2}=83.84\), and \(\mathrm{Avg}=72.75\), while **TEMA** reports \(R@1=45.29\), \(R@5=79.46\), \(R@10=88.17\), \(R_{\mathrm{subset}@1}=72.05\), \(R_{\mathrm{subset}@2}=86.52\), and \(\mathrm{Avg}=75.76\) [2604.21806].

The scoring function is written as
\[
s(x)=\cos\bigl(\mathrm{Combiner}(E_r,E_m),\,E_t(x)\bigr),
\]
where \(E_r\) and \(E_m\) are the entity-aware embeddings of the reference image and MMT, and \(E_t(x)\) is the target image embedding. The benchmark is therefore designed not merely for long-text retrieval, but specifically for explicit alignment between instruction clauses and visual entities [2604.21806].

A related but distinct development is **CIRCLED**, a multi-turn CIR dataset that extends FashionIQ, CIRR, and CIRCO and is described as being “sometimes referred to informally as ‘M-CIRR’” for its CIRR component [2605.26734]. CIRCLED differs from the single-turn M-CIRR benchmark in both object and protocol. It builds sessions of the form
\[
(I_1,T_1)\to (I_2,T_2)\to \cdots \to (I_L,T_L)\to Z,
\]
uses a CIReVL-based retrieval pipeline, and filters sessions by retrieval success, multi-turn validity, \(\varepsilon\)-consistency with \(\varepsilon=30\), and \(\tau\)-diversity with \(\tau=0.8\). The complete dataset contains **22 608** sessions across nine subsets, with **CIRR\(_{\rm train}\)** containing **6 874** sessions and **CIRR\(_{\rm val}\)** containing **959** [2605.26734].

This distinction matters because the two retrieval usages of M-CIRR are structurally different. The dataset of TEMA is a **single-turn multi-modification** benchmark with dense clause-level supervision, whereas the CIRCLED usage is **multi-turn** and models iterative refinement through dialogue history. A plausible implication is that the first stresses clause–entity grounding within one query, while the second stresses consistency and information accretion across turns.

Source: https://www.emergentmind.com/topics/m-cirr