TriMTR: Training-Free Token Reduction

• TriMTR is a training-free token reduction framework for Vision Mamba that uses the timescale parameter Δt as an importance score to merge low-importance tokens.
• The method preserves token order and handles recurrence by restoring the original sequence, which is critical for the input-dependent state-space dynamics.
• Empirical results show up to 40% FLOPs reduction with minimal accuracy loss, outperforming other training-free, inference-time token compression methods.

Searching arXiv for the primary paper and possible ambiguous uses of “TriMTR”. {"query":"(Ma et al., 18 Jul 2025) TriMTR Training-free Token Reduction for Vision Mamba", "max_results": 5} {"tool":"arxiv_search","query":"(Ma et al., 18 Jul 2025)"} TriMTR denotes, in the Vision Mamba literature, the training-free token-reduction framework introduced in "Training-free Token Reduction for Vision Mamba" (Ma et al., 18 Jul 2025). It is a plug-and-play inference-time method for compressing visual token sequences in Mamba-based vision backbones without retraining or extra parameters. The framework uses a structure-aware importance score derived from Mamba’s selective state-space dynamics, specifically the timescale parameter $\Delta_t$, to rank tokens, merge low-importance tokens into similar retained tokens, and restore original sequence order before subsequent Mamba blocks. Its central premise is that token reduction for Vision Mamba must respect recurrence and order sensitivity rather than reuse attention-based heuristics developed for Vision Transformers (Ma et al., 18 Jul 2025).

1. Motivation and problem setting

Vision Mamba is a selective state space model designed for linear-time sequence modeling. Visual variants such as ViM and VideoMamba scan image patches as ordered sequences and process them through input-dependent SSM dynamics rather than self-attention. In this setting, cutting tokens reduces FLOPs, memory footprint, and latency, which is directly relevant for real-time deployment and higher throughput. The motivation for TriMTR is therefore not merely compression in the abstract, but compression that is compatible with Mamba’s recurrence, input-gated dynamics, and order dependence (Ma et al., 18 Jul 2025).

A central observation is that directly applying Vision Transformer token-reduction methods to Vision Mamba causes substantial degradation. ViT compression methods such as DynamicViT, Token Merging/ToMe, A-ViT, and EViT rely on attention maps or on the $[\mathrm{CLS}]$ token’s attention, whereas Mamba has no Q/K/V and no attention-derived signals. The recurrence is also highly sensitive to token order, so dropping or reordering tokens perturbs state propagation. This is reflected in the reported comparisons at 40\% FLOPs reduction: on ViM-S, EViT drops by $-15.7\%$, PuMer by $-11.4\%$, whereas TriMTR drops by $-5.1\%$; on ViM-B, EViT drops by $-6.0\%$ whereas TriMTR drops by $-1.6\%$ (Ma et al., 18 Jul 2025).

TriMTR’s design therefore addresses two specific failure modes of naïve transfer from ViTs: the lack of an attention-native importance signal, and the disruption of sequence order after compression. The framework treats both as first-order constraints rather than secondary implementation details.

2. State-space formulation and the $\Delta_t$ importance signal

The mathematical basis of TriMTR is the continuous-time and discretized selective state-space formulation used in Vision Mamba. The standard SSM is

$$h'(t) = A h(t) + B x(t), \qquad y(t) = C h(t) + D x(t).$$

With zero-order hold discretization and timescale $\Delta$,

$[\mathrm{CLS}]$0

yielding the discrete SSM

$[\mathrm{CLS}]$1

In Mamba, $[\mathrm{CLS}]$2, $[\mathrm{CLS}]$3, and $[\mathrm{CLS}]$4 are input-dependent:

$[\mathrm{CLS}]$5

With diagonal $[\mathrm{CLS}]$6, the update simplifies elementwise to

$[\mathrm{CLS}]$7

where $[\mathrm{CLS}]$8 and $[\mathrm{CLS}]$9 is the Hadamard product. Dynamic parameters are generated by

$-15.7\%$0

In practice, ViM uses multiple scanning heads $-15.7\%$1, and the layer output is aggregated as

$-15.7\%$2

TriMTR’s key contribution is to treat the timescale parameter $-15.7\%$3 as an importance proxy. The intuition given in the paper is that $-15.7\%$4 acts as an input gate controlling how much the current token contributes relative to past memory. Larger $-15.7\%$5 implies stronger weighting of the current input in the state update $-15.7\%$6. The importance score for layer $-15.7\%$7 is therefore defined by first aggregating across scan heads,

$-15.7\%$8

and then averaging over channels,

$-15.7\%$9

This score is explicitly Mamba-native: it exists without extra computation, is coupled to the recurrence, and reflects input-gated contributions to downstream states. Empirically, it outperforms alternatives such as $-11.4\%$0, $-11.4\%$1, hidden-state magnitude $-11.4\%$2, and $-11.4\%$3 similarity. At 40\% FLOPs reduction, ViM-S reaches $-11.4\%$4 top-1 with $-11.4\%$5, compared with $-11.4\%$6 for $-11.4\%$7, $-11.4\%$8 for $-11.4\%$9, $-5.1\%$0 for $-5.1\%$1, and $-5.1\%$2 for $-5.1\%$3; on ViM-B, the corresponding values are $-5.1\%$4, $-5.1\%$5, $-5.1\%$6, $-5.1\%$7, and $-5.1\%$8 (Ma et al., 18 Jul 2025).

3. Reduction procedure and order preservation

TriMTR inserts an MTR module after each Mamba block. At layer $-5.1\%$9, it extracts $-6.0\%$0 from the block, computes $-6.0\%$1, sorts tokens in descending order of importance, partitions them into Keep, Target, and Source groups, merges Source into similar Target tokens, and then restores original order before the next Mamba block. The framework is training-free, uses fixed ratios, and introduces no new learnable parameters (Ma et al., 18 Jul 2025).

Let $-6.0\%$2 be the desired reduction ratio. TriMTR sets

• Keep $-6.0\%$3,
• Source $-6.0\%$4,
• Target $-6.0\%$5.

Source tokens are merged into Target tokens by bipartite soft matching, so the retained set has size

$-6.0\%$6

No tokens are simply skipped in the preferred configuration; the method uses merging to minimize information loss. After merging, the retained tokens are re-sorted according to their original indices before entering the next Mamba block. This restoration step is essential because the recurrent update is order-sensitive.

The paper’s algorithmic description is:

1. Sort $-6.0\%$7 by $-6.0\%$8 in descending order.
2. Compute Target size $-6.0\%$9.
3. Partition into Keep $-1.6\%$0, Target $-1.6\%$1, and Source $-1.6\%$2.
4. Merge Source into Target via bipartite soft matching: for each $-1.6\%$3, find the top-1 similar $-1.6\%$4, add edge $-1.6\%$5, then merge by the mean of vectors connected by the edge.
5. Concatenate $-1.6\%$6 and $-1.6\%$7, then reorder by original indices:

$-1.6\%$8

The same reduction is applied to hidden state and residual paths to ensure consistency. For multi-head or bi-directional scanning, $-1.6\%$9 is summed across scan heads before channel averaging. In batching, importance computation and grouping are performed per sample. If tokens originate from $\Delta_t$0D patches, the implementation maintains the mapping between sequence index and spatial coordinates so that order restoration and positional indexing remain consistent (Ma et al., 18 Jul 2025).

4. Computational properties and deployment characteristics

The computational model used in the paper is linear in sequence length. For sequence length $\Delta_t$1, channel dimension $\Delta_t$2, and $\Delta_t$3 scan heads, the Mamba block has

$\Delta_t$4

With token reduction to $\Delta_t$5,

$\Delta_t$6

Memory likewise scales with $\Delta_t$7, so token reduction yields proportional savings. Reported results include approximately 40\% FLOPs reduction on ViM-B with only a $\Delta_t$8 ImageNet top-1 drop, from $\Delta_t$9 to $$h'(t) = A h(t) + B x(t), \qquad y(t) = C h(t) + D x(t).$$0, without retraining. ViM-S at the same reduction drops from $$h'(t) = A h(t) + B x(t), \qquad y(t) = C h(t) + D x(t).$$1 to $$h'(t) = A h(t) + B x(t), \qquad y(t) = C h(t) + D x(t).$$2, and VideoMamba-B drops from $$h'(t) = A h(t) + B x(t), \qquad y(t) = C h(t) + D x(t).$$3 to $$h'(t) = A h(t) + B x(t), \qquad y(t) = C h(t) + D x(t).$$4. Throughput increases monotonically with the reduction ratio $$h'(t) = A h(t) + B x(t), \qquad y(t) = C h(t) + D x(t).$$5, reflecting the linear dependence on sequence length (Ma et al., 18 Jul 2025).

The integration recipe is correspondingly simple. The method is hooked after each Mamba block, using the already computed gating path $$h'(t) = A h(t) + B x(t), \qquad y(t) = C h(t) + D x(t).$$6. Default reduction ratios are presented as performance-focused $$h'(t) = A h(t) + B x(t), \qquad y(t) = C h(t) + D x(t).$$7, balanced $$h'(t) = A h(t) + B x(t), \qquad y(t) = C h(t) + D x(t).$$8, and speed-focused $$h'(t) = A h(t) + B x(t), \qquad y(t) = C h(t) + D x(t).$$9. The framework is applied uniformly across layers by default, with no layer-specific tuning required. A held-out validation set is used to sanity-check top-1 accuracy and profile FLOPs or latency. This suggests that TriMTR is intended as an inference-time systems component rather than a retraining-based architectural redesign (Ma et al., 18 Jul 2025).

5. Empirical results and ablation structure

The empirical evaluation spans ImageNet-1K classification on ViM and VideoMamba backbones, ablations on importance indicators and reduction operations, visualization, and throughput profiling. Across these settings, TriMTR consistently outperforms training-free baselines adapted from ViT token reduction, including EViT, PuMer, UTR, and HSA (Ma et al., 18 Jul 2025).

Backbone	Baseline top-1	TriMTR at 40\% FLOPs reduction
ViM-S	80.5\%	75.4\%
ViM-B	81.9\%	80.3\%
VideoM-S	81.2\%	76.6\%
VideoM-B	82.7\%	80.5\%

At the same 40\% reduction point, the comparator gaps are large. On ViM-S, EViT reaches $\Delta$0, PuMer $\Delta$1, UTR $\Delta$2, and HSA $\Delta$3. On ViM-B, EViT reaches $\Delta$4, PuMer $\Delta$5, UTR $\Delta$6, and HSA $\Delta$7. On VideoM-S, EViT reaches $\Delta$8, PuMer $\Delta$9, UTR $[\mathrm{CLS}]$00, and HSA $[\mathrm{CLS}]$01. On VideoM-B, EViT reaches $[\mathrm{CLS}]$02, PuMer $[\mathrm{CLS}]$03, UTR $[\mathrm{CLS}]$04, and HSA $[\mathrm{CLS}]$05.

The ablations are structurally important because they justify the two main design choices. First, $[\mathrm{CLS}]$06 is the best importance indicator among the candidates evaluated. Second, merging is preferable to pruning and competitive with or better than hybrid reduction. At 40\% reduction on ViM-S, merging yields $[\mathrm{CLS}]$07, hybrid $[\mathrm{CLS}]$08, and pruning $[\mathrm{CLS}]$09; on ViM-B, merging and hybrid both yield $[\mathrm{CLS}]$10, while pruning yields $[\mathrm{CLS}]$11. The visualization analysis further reports that Keep tokens align with high-response CAM regions, that Target plus Keep primarily cover foreground objects, and that filtered Source tokens largely correspond to background or irrelevant areas. Competing methods such as UTR and HSA may drop some foreground tokens. Throughput increases monotonically with the reduction ratio, so the evaluation presents a direct accuracy–speed trade-off rather than a single operating point (Ma et al., 18 Jul 2025).

6. Limitations, extensions, and terminological scope

The paper identifies several limitations. Aggressive compression eventually hurts accuracy; the gate-based score may underestimate tokens whose long-range influence manifests later in deeper layers. Tasks that depend heavily on fine-grained temporal or spatial order, such as very fine segmentation or tracking, may be more sensitive. Domain shift can also weaken $[\mathrm{CLS}]$12-based importance estimates, and the paper notes that light calibration or combining $[\mathrm{CLS}]$13 and $[\mathrm{CLS}]$14 could improve robustness. Proposed extensions include multi-modal Mamba, input-adaptive runtime policies for $[\mathrm{CLS}]$15, light tuning or calibration heads, and spatially local or layer-wise schedules for dense tasks (Ma et al., 18 Jul 2025).

The term itself is not unique across the arXiv literature. In unrelated contexts, closely related or identical shorthand refers to "Targeted Row-wise Iterative Metric-driven pruning" for extreme LLM sparsity (Beck et al., 22 May 2025), "TrimR" for verifier-based training-free thinking compression in large reasoning models (Lin et al., 22 May 2025), tract-specific MTR derived from dual-encoded magnetization transfer and diffusion imaging with COMMIT (Leppert et al., 2023), "MedTrim" for meta-entity-driven triplet mining in medical vision-language alignment (Ozturk et al., 22 Apr 2025), and a two-round triangle-counting algorithm under edge local differential privacy (Guo et al., 9 Jul 2025). Additional hardware-oriented mappings connect the label to hierarchical triple-modular redundancy (0902.0241) and to three-terminal thyristor RAM (Lee et al., 2024). This suggests that TriMTR is field-dependent nomenclature rather than a universally standardized acronym, and in current computer-vision usage it is most specifically associated with training-free token reduction for Vision Mamba (Ma et al., 18 Jul 2025).