SpecFlow: Spectral-Progressive Thought Flow

• Spectral-Progressive Thought Flow (SpecFlow) is a multimodal spatial reasoning framework that uses a fixed-size DCT visual workspace to progressively refine visual details.
• It employs classifier-free guidance and a deterministic ODE solver to align visual evolution with autoregressive textual thoughts, ensuring constant latency and memory usage.
• The framework reduces computation costs by eliminating visual token accumulation, achieving up to 2.1x KV-cache reduction while maintaining competitive reasoning performance.

Spectral-Progressive Thought Flow (SpecFlow) is a lightweight multimodal spatial reasoning framework in which intermediate visual thoughts are represented in a fixed-size discrete cosine space rather than as growing sequences of visual tokens. Its defining premise is that multimodal reasoning often requires preservation of global layout, object relations, and coarse geometry before task-critical fine detail, so visual state should evolve in a coarse-to-fine spectral schedule rather than by unrestricted token accumulation. In the reported formulation, classifier-free guidance aligns this visual evolution with autoregressive textual thoughts, yielding a bounded visual workspace whose updates depend only on the current visual state and accumulated textual trace. The resulting system is designed for long-horizon inference with stable latency and memory usage independent of reasoning depth, and is reported to achieve competitive or superior reasoning performance while reducing computation and KV-cache costs by up to $2.1\times$ (Shen et al., 1 Jun 2026).

1. Problem setting and design rationale

SpecFlow is motivated by a specific systems bottleneck in multimodal spatial reasoning: many interleaved reasoning architectures generate long chains of intermediate textual and visual thoughts, then append them to the context at each hop (Shen et al., 1 Jun 2026). In that regime, sequence length grows monotonically, visual thoughts can be on the order of $\mathcal{O}(10^3)$ tokens, and both attention cost and KV-cache memory increase accordingly. The framework is therefore aimed at reasoning tasks in which the dominant burden is not only cross-modal inference itself, but the cumulative computational cost of storing and attending over prior visual deliberation.

The design claim is that intermediate visual thoughts in spatial reasoning do not generally need pixel-level detail. The representation instead needs to preserve global layout, object relations, coarse geometry, and then later add finer detail when increased spatial precision is required. SpecFlow addresses this by replacing token-accumulating visual thoughts with a fixed-size evolving visual workspace. A plausible implication is that the method treats visual reasoning state as a controlled latent dynamical system rather than as a transcript of visual tokens.

The visual stream is explicitly Markovian in the sense used by the framework: the next visual thought depends on the current visual state and the accumulated text, not on all previous visual thoughts. This is the key formal property that permits overwriting rather than appending visual state. The textual stream remains autoregressive, but the visual workspace is bounded.

2. Spectral representation of the visual workspace

The “spectral-progressive” component of SpecFlow is implemented through a blockwise discrete cosine transform (DCT) representation of visual state (Shen et al., 1 Jun 2026). With block size $b \times b$, the framework uses a forward block DCT $\mathcal{D}_b(\cdot)$ and inverse transform $\mathcal{D}_b^{-1}(\cdot)$. A binary mask $M_t \in \{0,1\}^{b\times b}$ selects the active frequencies at flow time $t$, and the number of active frequencies is

$$m(t) \;=\; \sum_{u=0}^{b-1}\sum_{v=0}^{b-1} M_t(u,v),$$

with $m(t)$ non-decreasing in $t$.

Given an image or visual state $\mathcal{O}(10^3)$0, the frequency-limited reconstruction is

$\mathcal{O}(10^3)$1

where $\mathcal{O}(10^3)$2 is elementwise multiplication, broadcast over blocks. Early in the process, only low-frequency coefficients are active, so the representation is structure-dominant; later, additional bands are activated and the reconstruction becomes sharper and more detailed.

This representation is chosen because the DCT is orthogonal and energy-compacting: low frequencies capture global structure and layout, whereas high frequencies capture fine-grained details and texture. In SpecFlow, this yields a coarse-to-fine visual thought evolution in which global scene structure is stabilized first and higher-frequency detail is introduced only when required. The paper’s terminology frames this as maintaining a compact visual workspace while progressively unlocking spatial precision.

The full multimodal reasoning process alternates between visual and textual thoughts. If $\mathcal{O}(10^3)$3 denotes the visual thought at hop $\mathcal{O}(10^3)$4, $\mathcal{O}(10^3)$5 the textual thought at hop $\mathcal{O}(10^3)$6, $\mathcal{O}(10^3)$7 the initial visual input, and $\mathcal{O}(10^3)$8 the initial textual query, the framework writes

$\mathcal{O}(10^3)$9

$b \times b$0

Within this formulation, the visual state is not a token history but a spectrally constrained workspace.

3. Cosine-space flow matching and textual steering

Instead of autoregressively generating visual tokens, SpecFlow learns a deterministic flow field in cosine space (Shen et al., 1 Jun 2026). Let $b \times b$1 denote the DCT coefficient trajectory. The governing ODE is

$b \times b$2

where $b \times b$3 is the learnable velocity field, $b \times b$4 is the conditioning text, and the masked state is

$b \times b$5

The velocity field therefore operates only on the currently active spectral bands.

Training follows a flow-matching objective with paired endpoints $b \times b$6 and $b \times b$7, transformed into cosine space as

$b \times b$8

The interpolation is linear,

$b \times b$9

with constant target velocity

$\mathcal{D}_b(\cdot)$0

The training loss is

$\mathcal{D}_b(\cdot)$1

This formulation gives the visual update the structure of transport from noise toward a target visual state under a progressively expanding spectral mask.

Textual thoughts steer the flow by classifier-free guidance (CFG). With probability $\mathcal{D}_b(\cdot)$2, the condition is replaced by a null token $\mathcal{D}_b(\cdot)$3 during training. At inference, the guided velocity is

$\mathcal{D}_b(\cdot)$4

which the paper summarizes conceptually as

$\mathcal{D}_b(\cdot)$5

The guided ODE is then

$\mathcal{D}_b(\cdot)$6

This mechanism is central to the framework’s claim that visual state evolution can be aligned with linguistic intent without expanding the multimodal context. The unconditional component supplies generic prior dynamics; the conditional residual supplies text-attributable direction; and the guidance scale $\mathcal{D}_b(\cdot)$7 controls the strength of textual steering.

At inference, a fixed-step solver, typically Euler, is used. If $\mathcal{D}_b(\cdot)$8 is the coefficient state at step $\mathcal{D}_b(\cdot)$9 and $\mathcal{D}_b^{-1}(\cdot)$0, then

$\mathcal{D}_b^{-1}(\cdot)$1

After $\mathcal{D}_b^{-1}(\cdot)$2 steps, the updated visual state is reconstructed by

$\mathcal{D}_b^{-1}(\cdot)$3

4. Hop-wise inference loop, architecture, and computational profile

The hop-wise SpecFlow pipeline consists of five stages (Shen et al., 1 Jun 2026). First, the conditioning text is built as

$\mathcal{D}_b^{-1}(\cdot)$4

Second, the visual workspace is initialized from the previous hop,

$\mathcal{D}_b^{-1}(\cdot)$5

Third, a fixed-step ODE solver is run: at each step the coefficients are masked, unconditional and conditional velocities are computed, CFG combines them, and the state is updated by Euler integration. Fourth, the updated visual thought is decoded as $\mathcal{D}_b^{-1}(\cdot)$6. Fifth, the next textual thought is generated autoregressively,

$\mathcal{D}_b^{-1}(\cdot)$7

The crucial systems property is that the visual workspace is overwritten at each hop. Earlier visual thoughts are not retained in the context. As a result, memory and compute scale primarily with the textual history plus a constant-size visual workspace, rather than with an accumulated history of visual tokens.

The reported architecture couples two modules: a flow-matching visual-thought generator, implemented in VAE latent space and then transformed to cosine space, and an autoregressive text module. The default backbone setup uses Qwen3-VL-8B for autoregressive reasoning and Qwen-Image-Edit-2509 / 20B MMDiT for the flow-based visual generator. Stronger autoregressive baselines with Qwen3-VL-32B are also reported for model-scale control.

The computational rationale has five parts. There is no visual token accumulation; visual state size is fixed; early computation is frequency-limited because only low-frequency coefficients are active; deterministic flow with few solver steps avoids long stochastic sampling chains; and CFG steers the update internally rather than by enlarging the prompt. The paper reports that on compositional spatial reasoning tasks SpecFlow reduces KV-cache to about $\mathcal{D}_b^{-1}(\cdot)$8–$\mathcal{D}_b^{-1}(\cdot)$9 GB, compared with $M_t \in \{0,1\}^{b\times b}$0–$M_t \in \{0,1\}^{b\times b}$1 GB for VoCoT, corresponding to about $M_t \in \{0,1\}^{b\times b}$2–$M_t \in \{0,1\}^{b\times b}$3 reduction. On dynamic spatial decision-making tasks, it reports about $M_t \in \{0,1\}^{b\times b}$4 KV-cache reduction relative to MVoT, remaining around $M_t \in \{0,1\}^{b\times b}$5–$M_t \in \{0,1\}^{b\times b}$6 GB across tasks.

5. Training configuration, empirical results, and ablations

For SpecFlow flow-matching training, the reported setup uses an MSE velocity objective, CFG with condition dropout, Euler ODE inference, fixed $M_t \in \{0,1\}^{b\times b}$7 inference steps, guidance scale $M_t \in \{0,1\}^{b\times b}$8, time steps sampled from a logit-normal distribution, LoRA rank $M_t \in \{0,1\}^{b\times b}$9, $t$0 epochs, and learning rate $t$1 (Shen et al., 1 Jun 2026). The autoregressive baselines are trained by supervised fine-tuning for $t$2 epochs with learning rate $t$3 and LoRA rank $t$4, and by GRPO for $t$5 epoch with learning rate $t$6, rollout size $t$7, and KL coefficient $t$8.

Evaluation covers compositional multimodal spatial reasoning and dynamic spatial decision-making. The spatial benchmarks are VSR, V-Star, EmbSpatial, and Winoground. The dynamic tasks are Maze, MiniBehavior, and FrozenLake. The reported headline results on the compositional benchmarks are as follows.

Benchmark	SpecFlow	Reported comparator
VSR	70.14%	VoCoT 68.88%
V-Star	61.28%	VoCoT 59.87%
EmbSpatial	67.79%	SparseVLM 63.89%
Winoground	70.47%	VoCoT 70.09%

On dynamic spatial decision-making, SpecFlow is reported as best across all Maze grid sizes, with $t$9 at size $$m(t) \;=\; \sum_{u=0}^{b-1}\sum_{v=0}^{b-1} M_t(u,v),$$0. On MiniBehavior it reports $$m(t) \;=\; \sum_{u=0}^{b-1}\sum_{v=0}^{b-1} M_t(u,v),$$1 at size $$m(t) \;=\; \sum_{u=0}^{b-1}\sum_{v=0}^{b-1} M_t(u,v),$$2, described as remaining robust at long horizons. On FrozenLake it is reported to outperform naive Qwen3-VL, SFT, GRPO, and image-only diffusion reasoning. Relative to DiffThinker, the reported average success improves from $$m(t) \;=\; \sum_{u=0}^{b-1}\sum_{v=0}^{b-1} M_t(u,v),$$3 to $$m(t) \;=\; \sum_{u=0}^{b-1}\sum_{v=0}^{b-1} M_t(u,v),$$4.

The ablation studies isolate the spectral schedule, guidance scale, solver depth, and representation choice. Progressive schedules outperform a fixed low-frequency-only schedule, and a cosine progressive schedule often gives the best latency/accuracy trade-off. Performance improves as CFG scale increases, peaking around $$m(t) \;=\; \sum_{u=0}^{b-1}\sum_{v=0}^{b-1} M_t(u,v),$$5; too little guidance under-aligns visual updates, while too much can over-constrain them. Accuracy improves as ODE steps increase but saturates around $$m(t) \;=\; \sum_{u=0}^{b-1}\sum_{v=0}^{b-1} M_t(u,v),$$6–$$m(t) \;=\; \sum_{u=0}^{b-1}\sum_{v=0}^{b-1} M_t(u,v),$$7 steps, after which latency rises with diminishing returns. A comparison among spectral-only, VAE-only, and spectral+VAE variants reports that the combined version is best, being much faster than VAE-only and more accurate than spectral-only.

These findings support the reported interpretation that the spectral schedule is not merely a compression device but a control mechanism for reasoning dynamics. The article’s experimental claims are therefore tied to both efficiency and trajectory shaping: the model is intended to retain enough low-frequency information for global reasoning while deferring costly detail synthesis until later in the flow.

6. Distinctions, related uses, and scope

The name “SpecFlow” can invite confusion with earlier mathematical usages of “spectral flow,” but the multimodal framework is conceptually distinct. In operator theory, spectral flow denotes a classical notion of functional analysis and differential geometry; inside the essential spectrum, one paper identifies the singular spectral shift function $$m(t) \;=\; \sum_{u=0}^{b-1}\sum_{v=0}^{b-1} M_t(u,v),$$8 as an integer-valued extension of spectral flow for almost every spectral parameter, and relates it to total resonance index and Pushnitski $$m(t) \;=\; \sum_{u=0}^{b-1}\sum_{v=0}^{b-1} M_t(u,v),$$9-invariants (Azamov, 2014). That theory concerns self-adjoint operator paths, resonance points, and singular spectral data, not multimodal visual reasoning.

There is, however, a broader conceptual affinity with work on compositional spectral propagation in operator networks. A recent theory of spectral propagation for compositional operator systems develops three invariants—operadic spectrum $m(t)$0, spectral derivatives $m(t)$1, and interaction residue $m(t)$2—and formalizes how local spectral information is carried, transformed, and augmented across paths and interfaces in a network (Chang, 4 Jun 2026). A plausible implication is that both frameworks treat spectral content as something that evolves progressively through composition rather than as a static descriptor of isolated states, although the domains and formal objects are different.

A related but separate use of progressive flow ideas appears in spectral compressive imaging, where a trajectory-controllable unfolding network for CASSI reconstruction enforces smooth, continuous optimization paths from noisy initial estimates to high-quality reconstructions and reports stage-wise gradual refinement rather than abrupt jumps (Wang et al., 15 Sep 2025). That work addresses inverse reconstruction rather than multimodal reasoning, but it illustrates the more general appeal of explicitly controlled trajectories.

Within its own scope, SpecFlow suggests a shift from token-heavy visual chain-of-thought to structured continuous visual state. The reported trade-offs include dependence on the choice of block size $m(t)$3, guidance scale $m(t)$4, and number of ODE steps $m(t)$5, and very aggressive low-frequency restriction can lose task-critical details. The strongest evidence is reported on spatial reasoning and decision-making tasks. Even with those caveats, the framework defines a precise and technically distinctive interpretation of “spectral-progressive thought”: a fixed-size visual workspace in block-DCT space, updated by guided flow matching, progressively refined from low-frequency structure to higher-frequency detail, and coupled to autoregressive textual reasoning without visual-context growth (Shen et al., 1 Jun 2026).