FILT3R in Vision & Data Systems

• FILT3R is a term for two distinct adaptive filtering methods: one refines latent states in streaming 3D reconstruction while the other evaluates approximate membership filters.
• In 3D reconstruction, FILT3R replaces heuristic update rules with uncertainty-aware, Kalman-style gain schedules to maintain scene coherence over long sequences.
• In data systems, FILT3R unifies the evaluation of learned, stacked, and adaptive filters by comparing false positive rates, latency, and robustness under dynamic workloads.

FILT3R is a stylized name used in two distinct 2026 arXiv contributions. In computer vision, it denotes a training-free adaptive filtering layer for streaming 3D reconstruction that casts recurrent latent-state updates as stochastic state estimation in token space (Jin et al., 19 Mar 2026). In data systems, the same name is used for a systematic comparison of learned, stacked, and adaptive approximate membership filters, with emphasis on false positive rate, latency, construction cost, workload sensitivity, and robustness (Sabale et al., 13 Feb 2026). The shared spelling does not indicate a common method; rather, the term refers to two unrelated technical objects whose commonality is an explicit focus on filtering under uncertainty.

1. Nomenclature and domain-specific meanings

The two principal meanings of FILT3R differ in both mathematical object and application regime. One concerns token-wise latent memory in a recurrent reconstruction pipeline. The other concerns approximate set-membership data structures that exploit side information beyond the stored key set.

Usage of FILT3R	Domain	Core mechanism
Latent State Adaptive Kalman Filter	Streaming 3D reconstruction	Per-token variance, adaptive Kalman-style gain, drift-based process noise
Comparative filter study	Approximate membership filtering	Unified evaluation of learned, stacked, and adaptive filters

In the 3D reconstruction setting, FILT3R addresses a specific long-horizon failure mode: recurrent memory updates either overwrite useful history or become too inert to track genuine scene change. The method therefore replaces heuristic update coefficients with an uncertainty-aware, token-wise gain schedule (Jin et al., 19 Mar 2026). In the approximate membership setting, FILT3R addresses a different problem: modern filter families had been studied largely in isolation, often against weak baselines like Bloom filters, leaving their assumptions and trade-offs poorly understood in practice. The study therefore compares the three paradigms under multiple datasets, workloads, and metrics, rather than false positive rate alone (Sabale et al., 13 Feb 2026).

2. FILT3R as latent-state filtering in streaming 3D reconstruction

Streaming 3D reconstruction systems maintain a persistent latent state $\mathbf{s}_t$ across frames. The generic update rule highlighted in FILT3R is

$$\mathbf{s}_t=(1-\boldsymbol\beta_t)\odot\mathbf{s}_{t-1}+\boldsymbol\beta_t\odot\tilde{\mathbf{s}}_t,$$

where $\tilde{\mathbf{s}}_t$ is the decoder-produced candidate state and $\boldsymbol\beta_t\in[0,1]^N$ is a per-token update coefficient (Jin et al., 19 Mar 2026). Within this formulation, full overwrite is $\boldsymbol\beta_t=\mathbf{1}$, fixed EMA-style smoothing is constant $\boldsymbol\beta_t=\beta$, and heuristic gates are cue-driven coefficients that do not propagate confidence over time.

The central diagnosis of the paper is that the update rule, rather than merely the backbone architecture, becomes the bottleneck in long-horizon performance. Aggressive overwriting causes catastrophic forgetting and drift accumulation. Overly conservative updates fail to track genuine scene changes. Heuristic gates improve over pure overwrite but remain ad hoc, because they do not explicitly represent uncertainty. A further consequence is length generalization failure: update biases that are benign over training-length streams accumulate and destabilize the model once rollout length exceeds the training horizon.

FILT3R addresses this by recasting the recurrent latent state as a stochastic state-space model in token space:

$$\mathbf{s}_t = \mathbf{s}_{t-1} + \mathbf{w}_t,\qquad \mathbf{w}_t \sim \mathcal{N}(\mathbf{0}, \mathbf{Q}_t),$$

$$\tilde{\mathbf{s}}_t = \mathbf{s}_t + \mathbf{v}_t,\qquad \mathbf{v}_t \sim \mathcal{N}(\mathbf{0}, \mathbf{R}),$$

with fixed measurement noise $\mathbf{R}=r\mathbf{I}$ and adaptive process noise $\mathbf{Q}_t$. The persistent latent memory is treated as a belief about the scene, while the decoder candidate is treated as a noisy measurement of that belief. This converts memory update into a token-wise stochastic estimation problem, rather than a fixed overwrite-or-gate heuristic.

3. Token-wise variance propagation and adaptive Kalman gain

FILT3R adopts a diagonal covariance approximation, storing one scalar variance per token. The variance prediction step is

$$\mathbf{s}_t=(1-\boldsymbol\beta_t)\odot\mathbf{s}_{t-1}+\boldsymbol\beta_t\odot\tilde{\mathbf{s}}_t,$$0

where $$\mathbf{s}_t=(1-\boldsymbol\beta_t)\odot\mathbf{s}_{t-1}+\boldsymbol\beta_t\odot\tilde{\mathbf{s}}_t,$$1 is posterior variance and $$\mathbf{s}_t=(1-\boldsymbol\beta_t)\odot\mathbf{s}_{t-1}+\boldsymbol\beta_t\odot\tilde{\mathbf{s}}_t,$$2 is process noise (Jin et al., 19 Mar 2026). The corresponding Kalman-style gain is

$$\mathbf{s}_t=(1-\boldsymbol\beta_t)\odot\mathbf{s}_{t-1}+\boldsymbol\beta_t\odot\tilde{\mathbf{s}}_t,$$3

The state update becomes

$$\mathbf{s}_t=(1-\boldsymbol\beta_t)\odot\mathbf{s}_{t-1}+\boldsymbol\beta_t\odot\tilde{\mathbf{s}}_t,$$4

or equivalently

$$\mathbf{s}_t=(1-\boldsymbol\beta_t)\odot\mathbf{s}_{t-1}+\boldsymbol\beta_t\odot\tilde{\mathbf{s}}_t,$$5

Posterior variance is updated by the scalar Joseph-form recursion,

$$\mathbf{s}_t=(1-\boldsymbol\beta_t)\odot\mathbf{s}_{t-1}+\boldsymbol\beta_t\odot\tilde{\mathbf{s}}_t,$$6

This yields the paper’s core dynamic. Stable evidence contracts variance, so gains shrink and memory retention increases automatically. Genuine scene change raises uncertainty, so gains increase and the model becomes more responsive. In the paper’s interpretation, FILT3R therefore generalizes overwrite and gating rather than simply replacing them: overwrite corresponds to complete trust in the measurement, fixed-gain EMA is a constant point on the gain curve, and heuristic gates resemble policies that do not accumulate confidence recursively. The idealized zero-process-noise analysis further shows that gains decay roughly like $$\mathbf{s}_t=(1-\boldsymbol\beta_t)\odot\mathbf{s}_{t-1}+\boldsymbol\beta_t\odot\tilde{\mathbf{s}}_t,$$7, implying progressively more conservative updates in stable regimes.

The adaptive heart of the method is online process-noise estimation from EMA-normalized temporal drift. Per-token drift is defined as

$$\mathbf{s}_t=(1-\boldsymbol\beta_t)\odot\mathbf{s}_{t-1}+\boldsymbol\beta_t\odot\tilde{\mathbf{s}}_t,$$8

with stream-level mean drift

$$\mathbf{s}_t=(1-\boldsymbol\beta_t)\odot\mathbf{s}_{t-1}+\boldsymbol\beta_t\odot\tilde{\mathbf{s}}_t,$$9

and EMA baseline

$\tilde{\mathbf{s}}_t$0

Normalized drift is

$\tilde{\mathbf{s}}_t$1

which is mapped to process noise through

$\tilde{\mathbf{s}}_t$2

The practical stabilizers include adding $\tilde{\mathbf{s}}_t$3 to denominators, clamping the gain to $\tilde{\mathbf{s}}_t$4, and clamping the EMA baseline with a floor $\tilde{\mathbf{s}}_t$5. The default hyperparameters are fixed across tasks: $\tilde{\mathbf{s}}_t$6, $\tilde{\mathbf{s}}_t$7, $\tilde{\mathbf{s}}_t$8, $\tilde{\mathbf{s}}_t$9, $\boldsymbol\beta_t\in[0,1]^N$0, $\boldsymbol\beta_t\in[0,1]^N$1, and $\boldsymbol\beta_t\in[0,1]^N$2.

4. Empirical behavior of FILT3R in long-horizon 3D tasks

The empirical profile of the latent-state FILT3R is defined by long-horizon stability rather than only short-horizon accuracy (Jin et al., 19 Mar 2026). On TUM-RGBD, the clearest reported gain is in origin-aligned drift suppression. At 800 frames, $\boldsymbol\beta_t\in[0,1]^N$3 is $\boldsymbol\beta_t\in[0,1]^N$4 for CUT3R, $\boldsymbol\beta_t\in[0,1]^N$5 for TTT3R, and $\boldsymbol\beta_t\in[0,1]^N$6 for FILT3R. Rotation error also improves relative to TTT3R. On Bonn, FILT3R gives the best metric-scale depth across lengths 300, 400, and 500; at 500 frames, Abs Rel is $\boldsymbol\beta_t\in[0,1]^N$7 for CUT3R, $\boldsymbol\beta_t\in[0,1]^N$8 for TTT3R, and $\boldsymbol\beta_t\in[0,1]^N$9 for FILT3R. The paper also reports improvements in $\boldsymbol\beta_t=\mathbf{1}$0 and log RMSE, indicating fewer severe outliers.

For 3D reconstruction on 7-Scenes and NRGBD, the main claim is qualitative and structural: FILT3R remains coherent at very long lengths where overwrite and heuristic-gated baselines degrade, and where Point3R runs out of memory. The reported interpretation is that FILT3R preserves geometry more effectively over long rollouts, avoiding fragmentation and instability. Near the training horizon, gains are smaller; the method does not trade away short-horizon quality, but its main benefit emerges under extended streaming inference.

The method is designed to be lightweight. It requires one $\boldsymbol\beta_t=\mathbf{1}$1-dimensional variance vector, one scalar EMA drift baseline, and one buffer for the previous candidate state. Runtime and memory are described as essentially the same as CUT3R, unlike attention-map-based gates such as TTT3R, which require much more memory. The ablation results are also structurally informative. Fixed $\boldsymbol\beta_t=\mathbf{1}$2 harms long-horizon behavior because the filter cannot reopen properly during scene changes. Removing variance propagation is described as disastrous, because the system then cannot become more conservative over time. Fixed-$\boldsymbol\beta_t=\mathbf{1}$3 EMA can reduce global drift but oversmooths and harms local motion accuracy, especially rotation. Adaptive measurement noise $\boldsymbol\beta_t=\mathbf{1}$4 hurts stability and costs memory, leading to the paper’s specific conclusion that adaptivity should live in the process model rather than simultaneously in process and measurement noise.

5. FILT3R as a comparative framework for approximate membership filters

In the approximate membership literature, FILT3R names a unified evaluation of learned, stacked, and adaptive filters (Sabale et al., 13 Feb 2026). The common setting is the standard filter problem: a filter represents a set $\boldsymbol\beta_t=\mathbf{1}$5 approximately, answers membership queries with no false negatives, and controls the false positive rate. The paper situates newer paradigms against traditional Bloom, quotient, cuckoo, XOR, and ribbon filters, arguing that the new approaches exploit extra information and can yield orders-of-magnitude lower false positive rates, but only under paradigm-specific assumptions.

Learned filters use key features and a trained model. A score $\boldsymbol\beta_t=\mathbf{1}$6 is thresholded, and queries below threshold are passed to a backup filter. The overall false positive rate is

$\boldsymbol\beta_t=\mathbf{1}$7

where $\boldsymbol\beta_t=\mathbf{1}$8 is the model false positive rate and $\boldsymbol\beta_t=\mathbf{1}$9 is the backup filter false positive rate. The paper discusses the original sandwiched learned filter and focuses on Ada-BF and PLBF, both of which partition score space and allocate backup protection unevenly across groups. The theoretical discussion includes the condition

$\boldsymbol\beta_t=\beta$0

and the distribution-matching bound

$\boldsymbol\beta_t=\beta$1

These expressions formalize the study’s central point that learned filters are effective only when model quality is high and future queries resemble training and test distributions.

Stacked filters use workload knowledge rather than key features. They construct a cascade of alternating positive and negative layers, explicitly memorizing frequently queried negatives that would otherwise recur as false positives. Their false positive rate is expressed as

$\boldsymbol\beta_t=\beta$2

where $\boldsymbol\beta_t=\beta$3 is the probability that a negative query lies in the set of frequently queried negatives, $\boldsymbol\beta_t=\beta$4 is the false positive rate of layer $\boldsymbol\beta_t=\beta$5, and $\boldsymbol\beta_t=\beta$6 is the number of layers. This makes the workload-sensitive nature of stacked filters explicit.

Adaptive filters use false positive feedback rather than features or workload samples. The representative implementation in the paper is AdaptiveQF, based on quotient filters. Each key hash is split into a quotient $\boldsymbol\beta_t=\beta$7 and remainder $\boldsymbol\beta_t=\beta$8; on a false positive, the remainder is extended using additional hash bits, and a reverse map is consulted to recover the original key that caused the collision. The quoted guarantees are space

$\boldsymbol\beta_t=\beta$9

for target false positive rate $$\mathbf{s}_t = \mathbf{s}_{t-1} + \mathbf{w}_t,\qquad \mathbf{w}_t \sim \mathcal{N}(\mathbf{0}, \mathbf{Q}_t),$$0, and insertion cost

$$\mathbf{s}_t = \mathbf{s}_{t-1} + \mathbf{w}_t,\qquad \mathbf{w}_t \sim \mathcal{N}(\mathbf{0}, \mathbf{Q}_t),$$1

with high probability when the slot fill fraction is $$\mathbf{s}_t = \mathbf{s}_{t-1} + \mathbf{w}_t,\qquad \mathbf{w}_t \sim \mathcal{N}(\mathbf{0}, \mathbf{Q}_t),$$2. The reverse map is characterized as essential for adaptation and can be stored off-heap or on disk because it is accessed only when a false positive occurs.

6. Comparative findings, recommendations, and technical significance

The experimental design of the approximate-membership FILT3R is unusually broad, using the Malicious URL, Ember, Shalla, and Caida datasets, workload types including one-pass, uniform random, Zipfian with $$\mathbf{s}_t = \mathbf{s}_{t-1} + \mathbf{w}_t,\qquad \mathbf{w}_t \sim \mathcal{N}(\mathbf{0}, \mathbf{Q}_t),$$3, adversarial, and dynamic churned settings, and metrics including empirical false positive rate, construction time, query latency, and robustness (Sabale et al., 13 Feb 2026). False positive rate is computed as

$$\mathbf{s}_t = \mathbf{s}_{t-1} + \mathbf{w}_t,\qquad \mathbf{w}_t \sim \mathcal{N}(\mathbf{0}, \mathbf{Q}_t),$$4

with explicit care taken to isolate filter costs by measuring learned-model inference online and using an in-memory reverse map for adaptive filters.

The reported trade-offs are sharp. Learned filters can achieve up to $$\mathbf{s}_t = \mathbf{s}_{t-1} + \mathbf{w}_t,\qquad \mathbf{w}_t \sim \mathcal{N}(\mathbf{0}, \mathbf{Q}_t),$$5 lower false positive rates when the query distribution matches the training distribution well, but they exhibit high variance, lack robustness under skewed or dynamic workloads, and incur query latencies up to $$\mathbf{s}_t = \mathbf{s}_{t-1} + \mathbf{w}_t,\qquad \mathbf{w}_t \sim \mathcal{N}(\mathbf{0}, \mathbf{Q}_t),$$6 slower than stacked or adaptive filters. Their construction is at least $$\mathbf{s}_t = \mathbf{s}_{t-1} + \mathbf{w}_t,\qquad \mathbf{w}_t \sim \mathcal{N}(\mathbf{0}, \mathbf{Q}_t),$$7 slower than AdaptiveQF and stacked filters, because model training dominates. Stacked filters can achieve up to $$\mathbf{s}_t = \mathbf{s}_{t-1} + \mathbf{w}_t,\qquad \mathbf{w}_t \sim \mathcal{N}(\mathbf{0}, \mathbf{Q}_t),$$8 lower false positive rates on skewed workloads and, in some comparisons, up to $$\mathbf{s}_t = \mathbf{s}_{t-1} + \mathbf{w}_t,\qquad \mathbf{w}_t \sim \mathcal{N}(\mathbf{0}, \mathbf{Q}_t),$$9 better than weaker learned variants, but they require reliable workload knowledge and remain up to $$\tilde{\mathbf{s}}_t = \mathbf{s}_t + \mathbf{v}_t,\qquad \mathbf{v}_t \sim \mathcal{N}(\mathbf{0}, \mathbf{R}),$$0 slower than AdaptiveQF due to multi-layer traversal. Adaptive filters are the most robust overall, achieving up to $$\tilde{\mathbf{s}}_t = \mathbf{s}_t + \mathbf{v}_t,\qquad \mathbf{v}_t \sim \mathcal{N}(\mathbf{0}, \mathbf{R}),$$1 lower false positive rates under adversarial queries, maintaining low false positive rates across one-pass, uniform, Zipfian, and dynamic settings, and providing the fastest query and construction times among the compared modern filters.

The paper’s final recommendation is explicitly conditional. Learned filters suit stable workloads where key features are genuinely predictive and space constraints matter more than query latency. Stacked filters are appropriate when the query distribution is known in advance, relatively static, and especially skewed. Adaptive filters are the broadest and safest option when robustness is primary, particularly under dynamic or adversarial workloads, provided false positive feedback and the additional reverse-map state are available.

Taken together, the two uses of FILT3R illustrate a broader research pattern: in both cases, filtering is framed not as passive smoothing, but as a structured response to uncertainty and side information. In streaming 3D reconstruction, the side information is temporal drift and accumulated latent confidence (Jin et al., 19 Mar 2026). In approximate membership filtering, it is predictive features, workload samples, or false positive feedback (Sabale et al., 13 Feb 2026). The term therefore denotes different mechanisms in different communities, but in each case it marks an effort to replace fixed heuristics with explicit assumptions about evidence, uncertainty, and adaptation.