AssoCiAm in Hardware & MLLM Evaluation

• AssoCiAm is a cross-domain label defining both a non-volatile in-memory TCAM architecture in hardware and a benchmark for evaluating associative thinking in multimodal models.
• The hardware implementation uses a 2T a-IGZO FG transistor design to enable immediate, energy-efficient pattern matching with enhanced scalability.
• The benchmark challenges creative model evaluation by controlling ambiguity through image-text tasks that measure associative, lateral thinking.

Searching arXiv for the provided AssoCiAm-related papers and recent context. AssoCiAm is a reused research label rather than a single, field-stable term. In hardware research, it denotes “Associative Computational Memory,” a non-volatile, in-memory associative search architecture implemented as a ternary content-addressable memory (TCAM) using two amorphous InGaZnO floating-gate transistors per cell. In multimodal evaluation research, it denotes a benchmark for associative thinking in multimodal LLMs (MLLMs) that is designed to circumvent ambiguity in association tasks. Related work also places the label near two further domains: an associative-memory mechanism for deep clustering, where the term itself is stated not to appear explicitly, and a mould-theoretic formulation of associators expressed through balanced symmetral moulds (Sun et al., 2021, Liu et al., 17 Sep 2025, Saha et al., 2 Jan 2026, Furusho et al., 2023).

1. Terminological scope

The term is best understood through contextual disambiguation. One usage is a device-and-architecture proposal for non-volatile associative search. Another is a benchmark for evaluating creative association in MLLMs. Additional nearby uses connect the label to associative-memory-based clustering and to associators in mould theory. This suggests that “AssoCiAm” functions as a cross-domain label attached to different technical objects rather than as a unified doctrine or framework (Sun et al., 2021, Liu et al., 17 Sep 2025, Saha et al., 2 Jan 2026, Furusho et al., 2023).

Context	Meaning of AssoCiAm	Core construct
3D NAND-compatible hardware	Associative Computational Memory	2T a-IGZO FG TCAM
MLLM evaluation	Benchmark for associative thinking	Image-text ambiguity-controlled benchmark
Deep clustering	Related associative-memory mechanism	AM inside DCAM; term not explicit
Mould theory	Associators in mould formulation	$$\mathsf{GARI}(\mathscr{F})_{\mathsf{as}+\mathsf{bal}}$$

2. AssoCiAm as non-volatile associative computational memory

In the hardware sense, AssoCiAm is a non-volatile, in-memory associative search architecture realized as a TCAM cell with only two a-IGZO floating-gate transistors. The stored ternary symbol $0/1/X$ is encoded as a threshold-voltage state of the two transistors, and query evaluation occurs directly in the memory array by observing matchlines rather than by moving data to a CPU. Each TCAM row has a precharged matchline. The two gates in each cell are driven by complementary search lines $SL$ and $\overline{SL}$. If any cell mismatches, at least one transistor turns ON and discharges the matchline; if all cells match, the matchline remains high. Because storage is non-volatile, search can be performed immediately without refresh, eliminating dynamic power and leakages associated with volatile SRAM-based TCAMs (Sun et al., 2021).

The reported cell topology uses two parallel FG transistors $T0$ and $T1$, with sources tied to the matchline and drains tied to ground. In the test setup, a $3\ \text{M}\Omega$ load resistor connects the matchline to $V_{DD}=1.2\ \text{V}$. The design merges bitline and searchline functionality, and the absence of separate write SRAM and refresh reduces wiring and area relative to 16T-CMOS TCAM. One presented ternary encoding writes “1” as $T0$ high $V_T$, $0/1/X$0 low $0/1/X$1; “0” as $0/1/X$2 low $0/1/X$3, $0/1/X$4 high $0/1/X$5; and “X” as both high $0/1/X$6. The defining requirement is that, for a given search polarity, matches leave both transistors OFF.

The experimental search condition uses $0/1/X$7, $0/1/X$8, and $0/1/X$9 pulses. In simulation, $SL$0 and $SL$1 are used to enlarge the sensing margin by making OFF devices see a more negative gate bias and thereby minimizing residual conduction. This search mechanism makes AssoCiAm a compute-in-memory architecture in the strict sense: pattern matching is executed through device state and matchline dynamics inside the array rather than as an external logic operation (Sun et al., 2021).

3. Device technology, array behavior, and engineering constraints

The hardware implementation is built on a 3D NAND-compatible a-IGZO FG transistor. The device is a bottom-gate TFT with top-contact source and drain, with channel length $SL$2 scaled to $SL$3 and channel width typically $SL$4. The channel is approximately $SL$5 a-IGZO with RMS roughness of approximately $SL$6. The gate stack in the channel region is $SL$7, corresponding to tunneling oxide, floating gate, blocking oxide, and control gate. The thermal budget is $SL$8, which is presented as back-end-of-line compatible and suitable for monolithic 3D integration.

The device metrics are central to the architecture. The reported $SL$9 is $\overline{SL}$0, approximately $\overline{SL}$1 absolute and limited by the measurement tool. The $\overline{SL}$2 ratio is approximately $\overline{SL}$3 at $\overline{SL}$4. The ON current reaches a record $\overline{SL}$5 at $\overline{SL}$6, $\overline{SL}$7, and gate overdrive $\overline{SL}$8, while it is approximately $\overline{SL}$9 at $T0$0 and $T0$1. The subthreshold swing is approximately $T0$2 for long channels and approximately $T0$3 at $T0$4. Effective mobility is $T0$5, and transconductance reaches approximately $T0$6 for long channels and approximately $T0$7 at $T0$8. The memory window is approximately $T0$9 at room temperature and approximately $T1$0 at $T1$1, remaining at least $T1$2 even at $T1$3. Retention is projected at $T1$4 years with memory window approximately $T1$5, and endurance reaches $T1$6 program/erase cycles with the memory window decreasing from $T1$7 to $T1$8. Programming and erasing use approximately $T1$9 and $3\ \text{M}\Omega$0 to $3\ \text{M}\Omega$1 pulses, respectively, with $3\ \text{M}\Omega$2 pulse width sufficient to saturate the threshold shift. The charge-storage relation is given by

$3\ \text{M}\Omega$3

At array level, rows share matchlines and columns share complementary search lines. The dominant search-energy estimate is

$3\ \text{M}\Omega$4

and the mismatch and match timing estimates are

$3\ \text{M}\Omega$5

Using experimentally calibrated models, the design is reported to achieve at least $3\ \text{M}\Omega$6 improvement in array-size scalability and at least $3\ \text{M}\Omega$7 reduction in search energy relative to 16T-CMOS, 2T2R, and 2FeFET TCAMs. The throughput implication is that, although experiments used $3\ \text{M}\Omega$8 search pulses because of the test setup, the ON currents and matchline RC constants support ns-scale search in fabricated arrays. The principal limitations identified are the $3\ \text{M}\Omega$9, $V_{DD}=1.2\ \text{V}$0 write requirement, the need for a negative search bias in simulation, source/drain series resistance at short $V_{DD}=1.2\ \text{V}$1, peripheral-circuit co-design, and the absence of systematic $V_{DD}=1.2\ \text{V}$2 and array-yield studies under cycling and temperature (Sun et al., 2021).

4. AssoCiAm as a benchmark for associative thinking in MLLMs

In the evaluation sense, AssoCiAm is an image-text multimodal benchmark designed to measure associative, creative thinking in MLLMs while explicitly circumventing ambiguity. Association is defined as the process of forming novel connections between seemingly unrelated concepts stored in memory, and the benchmark treats it as foundational to creativity, divergent thinking, and lateral thinking. The benchmark decomposes ambiguity into two forms. Internal ambiguity occurs when the single labeled answer is itself unreasonable or inconsistent with typical human perception. External ambiguity occurs when more than one option is equally valid in principle, but only one is labeled correct. The paper’s cloud-shape example illustrates both: choosing “cat” because a cat can curl into a circle is an instance of internal ambiguity, whereas including several circular distractors such as “donut” and “coin” is an instance of external ambiguity (Liu et al., 17 Sep 2025).

Each AssoCiAm item comprises an image, a multiple-choice textual question, and $V_{DD}=1.2\ \text{V}$3 options with exactly one designated correct answer, with $V_{DD}=1.2\ \text{V}$4. The benchmark contains $V_{DD}=1.2\ \text{V}$5 question-option items derived from $V_{DD}=1.2\ \text{V}$6 high-quality images at $V_{DD}=1.2\ \text{V}$7 resolution, spanning $V_{DD}=1.2\ \text{V}$8 classes. The subtasks are evenly distributed: $V_{DD}=1.2\ \text{V}$9, $T0$0, and $T0$1 each contain $T0$2 items. There is no training or development split; the benchmark is test-only. Text prompts and options are in English. To reduce prompt phrasing randomness, each image is accompanied by three paraphrased questions that preserve the same semantic intent. This design attempts to elicit divergent association visually while keeping scoring compatible with a single-answer multiple-choice protocol.

Construction combines automated and human processes. Masks are extracted from ILSVRC12 images using SAM. Control diffusion models regenerate images guided by these masks, with irrelevant regions overlaid to reduce distractions. CLIP classifies eight regenerated images per mask, and masks are retained when average class probability exceeds $T0$3, a filter intended to ensure representative shapes. Human experts then review the filtered masks, add a small number from Flaticon to cover common contextual associations, and manually filter generated images for consistency, clarity, and completeness. Internal ambiguity is thus handled through representativeness filtering and expert curation, while external ambiguity is deferred to a later distractor-selection procedure (Liu et al., 17 Sep 2025).

5. Hybrid ambiguity circumvention, evaluation protocol, and empirical findings

The benchmark’s central methodological contribution is a hybrid computational pipeline that integrates rule-based constraints, representation learning, stochastic optimization, and expert review. External ambiguity is handled by constructing an undirected complete graph $T0$4 over classes or masks, with edge weights $T0$5 defined by DINO-v2 shape-level similarity computed on binary masks. For a subgraph $T0$6 containing the correct answer $T0$7, the average answer-to-distractor similarity is

$T0$8

and the optimization target is

$T0$9

Minimizing $V_T$0 makes distractors dissimilar from the answer, while the variance term regularizes similarity structure among distractors to avoid trivial elimination strategies. The search is performed with a genetic algorithm using binary-string encoding, population size $V_T$1, tournament selection, single-point crossover with probability $V_T$2, bit-flip mutation at $V_T$3, and constraint handling by flipping bits to satisfy the required option count. A reported $V_T$4 analysis shows that, for $V_T$5 id $V_T$6, increasing $V_T$7 from $V_T$8 to $V_T$9 raises $0/1/X$00 by $0/1/X$01 and decreases $0/1/X$02 by $0/1/X$03, while for $0/1/X$04 id $0/1/X$05, $0/1/X$06 increases by $0/1/X$07 and $0/1/X$08 decreases by $0/1/X$09 (Liu et al., 17 Sep 2025).

Performance is measured primarily by top-1 accuracy for each $0/1/X$10 task and by the weighted average

$0/1/X$11

The full-dataset random baseline is $0/1/X$12 for $0/1/X$13, $0/1/X$14 for $0/1/X$15, $0/1/X$16 for $0/1/X$17, and weighted average $0/1/X$18. The best reported model is LLaVA-OneVision-7B with weighted average $0/1/X$19, followed closely by LLaVA-OneVision-72B with $0/1/X$20, Gemini-1.5-pro with $0/1/X$21, and Qwen-VL-Max with $0/1/X$22. GPT-4o-mini reaches $0/1/X$23. Human experts score $0/1/X$24 across all subtasks. The top two models obtain $0/1/X$25 and $0/1/X$26 on $0/1/X$27, respectively, showing a consistent decline as the number of options increases.

The empirical analysis argues that ambiguity materially degrades evaluation quality. Replacing correct answers with unreasonable ones or replacing distractors with shape-similar ambiguous distractors drives accuracy toward chance, described as random-like behavior. With random distractor selection, $0/1/X$28 of question-option sets are externally ambiguous overall, broken down as $0/1/X$29 for $0/1/X$30 options, $0/1/X$31 for $0/1/X$32, and $0/1/X$33 for $0/1/X$34; with the optimization algorithm, the corresponding figure is $0/1/X$35. Association scores also correlate positively with MMMU cognition scores: Pearson $0/1/X$36 exceeds $0/1/X$37 across tasks, with three correlations above $0/1/X$38. At the same time, the paper notes several limits: the benchmark is confined to shape-based association, uses English prompts, enforces a single correct answer despite the diversity of genuine associations, and does not report Cronbach’s alpha, test-retest reliability, bootstrap confidence intervals, or formal significance tests beyond confidence intervals around regression lines (Liu et al., 17 Sep 2025).

6. Adjacent usages and conceptual boundaries

A related but distinct use of associative memory appears in deep clustering. The paper “Deep Clustering with Associative Memories” states explicitly that the term “AssoCiAm” does not appear in the paper. Its method, DCAM, inserts an associative memory into the latent space of an autoencoder and defines the energy

$0/1/X$39

with retrieval dynamics

$0/1/X$40

The single training objective is

$0/1/X$41

If “AssoCiAm” is intended as shorthand for an associative-memory-based deep clustering approach, the paper identifies the associative-memory component inside DCAM as the relevant mechanism, but not as a named AssoCiAm system (Saha et al., 2 Jan 2026).

A further mathematical usage attaches the label to associators in mould theory. There, the relevant object is $0/1/X$42, a set of balanced symmetral moulds forming a mould-theoretic reformulation of Drinfeld’s associator set. The balance condition is written

$0/1/X$43

the mould pentagon relation is

$0/1/X$44

and the paper proves

$0/1/X$45

together with the embedding

$0/1/X$46

In the formal power series case, the identification

$0/1/X$47

links the mould-theoretic construction to classical associators and, via the double-shuffle side, to $0/1/X$48 (Furusho et al., 2023).

Taken together, these usages delimit the term’s present research profile. In hardware, AssoCiAm names a specific 2T non-volatile TCAM architecture. In multimodal evaluation, it names a benchmark engineered to separate associative ability from annotation ambiguity. In deep clustering and mould theory, the label is adjacent to, or directly attached to, technically different associative-memory and associator constructions. The coexistence of these meanings suggests that any serious use of the term requires explicit field-level qualification.