M-CIRR: Molecular & Image Retrieval

• M-CIRR is an overloaded abbreviation used both for molecular channel impulse response reconstruction in nanoscale communications and as an instruction-rich multi-modification dataset in image retrieval.
• In molecular communication, it encapsulates statistical channel estimation methods like ML, LSSE, MAP, and LMMSE to infer discrete-time impulse responses from molecule count data.
• In composed image retrieval, M-CIRR extends the base CIRR dataset with detailed, clause-level modifications that enhance alignment between visual entities and textual instructions.

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 (Jamali et al., 2016, Jamali et al., 2015, Lee et al., 4 Dec 2025). 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 (Li et al., 23 Apr 2026, Takeda et al., 26 May 2026). 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	(Jamali et al., 2016)
Molecular communication	Exact molecular channel impulse response for drifted spherical reception	(Lee et al., 4 Dec 2025)
Composed image retrieval	Multi-modification extension of CIRR	(Li et al., 23 Apr 2026)
Multi-turn composed image retrieval	Informal name for the CIRR component of CIRCLED	(Takeda et al., 26 May 2026)

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 (Jamali et al., 2016). 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 (Li et al., 23 Apr 2026).

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 $k$0 (Jamali et al., 2016).

By linearity of expectations, the discrete-time CIR vector is

$k$1

with

$k$2

The continuous-time counterpart is the expected number of received molecules after an instantaneous release at $k$3. For free space, one example is

$k$4

and earlier work expresses the continuous-time CIR as the integral of the expected concentration over the receiver volume $k$5 (Jamali et al., 2015).

The estimation problem is posed over a known pilot. Counts $k$6 are collected and stacked into $k$7, with design vectors

$k$8

Under the assumption of independent Poisson observations,

$k$9

which yields the likelihood

$k$0

This formulation makes M-CIRR a Poisson inverse problem with nonnegativity constraints and finite channel memory (Jamali et al., 2016).

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

$k$1

subject to $k$2, and its stationary condition is

$k$3

In practice, one tests the unconstrained root and, if some components are negative, searches over lower-dimensional faces by fixing zero components (Jamali et al., 2016).

The LSSE estimator matches empirical means through

$k$4

and solves

$k$5

Dropping nonnegativity gives the normal equation

$k$6

Again, if some components are negative, they are set to zero and the reduced system is re-solved (Jamali et al., 2016, Jamali et al., 2015).

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

$k$7

under $k$8. The LMMSE estimator seeks a linear mapping $k$9 that minimizes $s[k]\cdot N^{\rm Tx}$0, yielding

$s[k]\cdot N^{\rm Tx}$1

which depends only on first- and second-order prior statistics (Jamali et al., 2016).

Performance is benchmarked by classical and Bayesian Cramér–Rao bounds. For deterministic $s[k]\cdot N^{\rm Tx}$2, the Fisher information matrix is

$s[k]\cdot N^{\rm Tx}$3

so any unbiased estimator satisfies $s[k]\cdot N^{\rm Tx}$4. With a prior, the Bayesian Fisher information adds the prior term,

$s[k]\cdot N^{\rm Tx}$5

and $s[k]\cdot N^{\rm Tx}$6 (Jamali et al., 2016).

Pilot-sequence design chooses $s[k]\cdot N^{\rm Tx}$7 to reduce estimation error. The LSSE-oriented criterion minimizes an MSE proxy while keeping $s[k]\cdot N^{\rm Tx}$8 well-conditioned; the LMMSE-oriented criterion maximizes

$s[k]\cdot N^{\rm Tx}$9

A simple suboptimal alternative is the ISI-free pilot that emits a $N^{\rm Tx}$0 every $N^{\rm Tx}$1 intervals and zeros otherwise, so that each tap appears in isolation and can be estimated by averaging (Jamali et al., 2016, Jamali et al., 2015).

The practical algorithm is correspondingly modular: choose a training sequence offline, transmit the $N^{\rm Tx}$2-symbol pilot, record counts $N^{\rm Tx}$3, compute $N^{\rm Tx}$4 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 (Jamali et al., 2016).

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 (Lee et al., 4 Dec 2025). Here the receiver has radius $N^{\rm Tx}$5, is located at the origin, the transmitter is at distance $N^{\rm Tx}$6, the drift has speed $N^{\rm Tx}$7, and the drift direction forms angle $N^{\rm Tx}$8 relative to the transmitter–receiver axis.

The molecule position satisfies the Itô SDE

$N^{\rm Tx}$9

and Girsanov’s theorem is used to transform the no-drift hitting-time law into the drifted one. If $0\le s[k]\le 1$0, the drifted joint density obeys

$0\le s[k]\le 1$1

The CIR is obtained by marginalizing over the sphere,

$0\le s[k]\le 1$2

A key step is a surface-integral identity involving Legendre polynomials $0\le s[k]\le 1$3 and modified Bessel functions $0\le s[k]\le 1$4. The resulting exact closed form is

$0\le s[k]\le 1$5

with

$0\le s[k]\le 1$6

The peak time $0\le s[k]\le 1$7 is characterized by the exact condition $0\le s[k]\le 1$8, and once $0\le s[k]\le 1$9 is found numerically, the peak amplitude is $r[k]$0 (Lee et al., 4 Dec 2025).

This formulation is presented as exact for arbitrary $r[k]$1, 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 (Li et al., 23 Apr 2026). It is defined as

$r[k]$2

where $r[k]$3 is a reference image, $r[k]$4 is the target image, and $r[k]$5 is a multi-modification text (MMT) that decomposes, in an explicit clause-by-clause manner, every visual change needed to transform $r[k]$6 into $r[k]$7.

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 (Li et al., 23 Apr 2026).

The resulting corpus contains $r[k]$8 training and $r[k]$9 validation triplets. Modification length ranges from 35 to 468 tokens with mean $L$0. 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 $L$1 contains 1 200 queries whose negative candidates are visually confusable (Li et al., 23 Apr 2026).

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 (Li et al., 23 Apr 2026).

Relative to the original CIRR modification texts of 2–50 tokens with mean $L$2, 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 (Li et al., 23 Apr 2026).

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

M-CIRR follows CIRR’s standard recall protocol on validation. The principal metric is

$L$3

where $L$4 is the rank of the true target for query $L$5. For the fine-discrimination subset, the reported metrics are $L$6 for $L$7, and the overall score is

$L$8

Representative results with BLIP backbones report Candidate at $L$9, $r[k] = \sum_{\ell=1}^L c_\ell[k] + c_n[k],$0, $r[k] = \sum_{\ell=1}^L c_\ell[k] + c_n[k],$1, $r[k] = \sum_{\ell=1}^L c_\ell[k] + c_n[k],$2, $r[k] = \sum_{\ell=1}^L c_\ell[k] + c_n[k],$3, and $r[k] = \sum_{\ell=1}^L c_\ell[k] + c_n[k],$4, while TEMA reports $r[k] = \sum_{\ell=1}^L c_\ell[k] + c_n[k],$5, $r[k] = \sum_{\ell=1}^L c_\ell[k] + c_n[k],$6, $r[k] = \sum_{\ell=1}^L c_\ell[k] + c_n[k],$7, $r[k] = \sum_{\ell=1}^L c_\ell[k] + c_n[k],$8, $r[k] = \sum_{\ell=1}^L c_\ell[k] + c_n[k],$9, and $$c_\ell[k]\sim \mathrm{Poiss}(s[k-\ell+1]\cdot \bar c_\ell)$$0 (Li et al., 23 Apr 2026).

The scoring function is written as

$$c_\ell[k]\sim \mathrm{Poiss}(s[k-\ell+1]\cdot \bar c_\ell)$$1

where $$c_\ell[k]\sim \mathrm{Poiss}(s[k-\ell+1]\cdot \bar c_\ell)$$2 and $$c_\ell[k]\sim \mathrm{Poiss}(s[k-\ell+1]\cdot \bar c_\ell)$$3 are the entity-aware embeddings of the reference image and MMT, and $$c_\ell[k]\sim \mathrm{Poiss}(s[k-\ell+1]\cdot \bar c_\ell)$$4 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 (Li et al., 23 Apr 2026).

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 (Takeda et al., 26 May 2026). CIRCLED differs from the single-turn M-CIRR benchmark in both object and protocol. It builds sessions of the form

$$c_\ell[k]\sim \mathrm{Poiss}(s[k-\ell+1]\cdot \bar c_\ell)$$5

uses a CIReVL-based retrieval pipeline, and filters sessions by retrieval success, multi-turn validity, $$c_\ell[k]\sim \mathrm{Poiss}(s[k-\ell+1]\cdot \bar c_\ell)$$6-consistency with $$c_\ell[k]\sim \mathrm{Poiss}(s[k-\ell+1]\cdot \bar c_\ell)$$7, and $$c_\ell[k]\sim \mathrm{Poiss}(s[k-\ell+1]\cdot \bar c_\ell)$$8-diversity with $$c_\ell[k]\sim \mathrm{Poiss}(s[k-\ell+1]\cdot \bar c_\ell)$$9. The complete dataset contains 22 608 sessions across nine subsets, with CIRR$k$00 containing 6 874 sessions and CIRR$k$01 containing 959 (Takeda et al., 26 May 2026).

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.