Trust Filtering in Computational Systems

• Trust filtering is a family of approaches that compute trust scores to selectively admit, weight, or reject signals, balancing purity and coverage in various applications.
• It employs both hard and soft filtering methods—using explicit trust, inferred reliability, and uncertainty measures—to enhance accuracy and robustness.
• Key implementations span recommender systems, weak supervision, influence networks, and distributed security, where design trade-offs affect propagation depth and adversarial resilience.

Trust filtering denotes a family of mechanisms that decide which signals should be admitted to computation, how strongly they should be weighted, or when they should be rejected, using explicit trust relations, inferred trust, reputation indicators, learned reliability scores, or joint trust–uncertainty gates. In the literature, the filtered object varies by domain: recommender systems filter neighborhoods or rating contributions; weak-to-strong learning filters weak labels; opinion dynamics filters influence edges; IoT and distributed security filter recommendations, reports, or even the execution path of a filter itself (Ghenai et al., 2016, Uzunoglu et al., 31 May 2026, Masoumi et al., 23 Jun 2026, Ziauddin et al., 24 Aug 2025, 0908.3930).

1. Core abstraction and formal structure

A recurring structure is to compute a trust-relevant score, apply a gating or ranking rule, and then aggregate only the retained information. The score may come from social links, path-based inference, reputation, hidden-state features, or joint trust/distrust assessments. The gating may be a hard threshold, a top-$n$ selection, or a continuous weight in the downstream objective.

Setting	Trust signal	Filtered object
Recommender systems (Ghenai et al., 2016, He et al., 2017)	explicit trust, inferred implicit trust, reputation	neighbors and rating contributions
Weak supervision and distillation (Uzunoglu et al., 31 May 2026, Yoo et al., 6 Oct 2025)	scalar trust score on weak labels or rationales	training examples and loss terms
Opinion dynamics (Masoumi et al., 23 Jun 2026)	net trust $\tau_{ij}$ and uncertainty $\sigma_{ij}$	interpersonal influence edges
IoT and distributed security (Ziauddin et al., 24 Aug 2025, 0908.3930)	direct trust, precision trust, report confidence	recommendations, reports, accepted evidence

In trust-based recommendation, the filtering step appears directly in the neighborhood definition. The rating predictor is

$$P_{a,i} = \bar r_a + \frac{ \sum_{u\in N(a,i)} (r_{u,i}-\bar r_u)\,w_{a,u} }{ \sum_{u\in N(a,i)} |w_{a,u}| },$$

so the primary design choice is how $N(a,i)$ is formed (Ghenai et al., 2016). In weak-to-strong learning, the trust object is the weak label itself: a trust function $\tau:\mathcal G\to[0,1]$ produces $t_i=\tau(g_{\pi_{\mathcal W}}(x_i,\hat y_i))$, with $t_i\approx \Pr(\hat y_i=y_i)$ under the source distribution (Uzunoglu et al., 31 May 2026). In ambivalent social networks, the filter is edgewise: $$[A_{\rm eff}(\theta,\phi)]_{ij} = A_{ij}\,\mathbf 1\{\tau_{ij}\ge \theta \wedge \sigma_{ij}\le \phi\},$$ so trust filtering becomes the construction of an effective influence graph (Masoumi et al., 23 Jun 2026).

Two implementation styles recur. One is hard filtering, in which low-trust inputs are excluded altogether. The other is soft trust weighting, in which all or most inputs remain but high-trust examples receive larger influence. Self-Filtered Distillation combines both: examples with $CTS(x_i)<T_0$ are omitted, while retained samples contribute a loss scaled by trust (Yoo et al., 6 Oct 2025). This suggests that trust filtering is not a single algorithmic primitive but a design pattern that can be instantiated at neighbor selection time, at data curation time, or at graph-construction time.

2. Trust filtering in recommender systems

The earliest and most developed use of the term is in trust-aware collaborative filtering. In “Exploring Trust-Aware Neighbourhood in Trust-based Recommendation,” the “Trust-Aware Neighbourhood” algorithm restricts the neighborhood to direct trustees,

$\tau_{ij}$0

thereby avoiding propagated trust beyond distance $\tau_{ij}$1 and the “very expensive all-pairs trust-distance computations of Massa & Avesani (2004)” (Ghenai et al., 2016). On Epinions, RMSUE improves from $\tau_{ij}$2 to $\tau_{ij}$3 with coverage rising from $\tau_{ij}$4 to $\tau_{ij}$5; on Flixster, coverage rises from $\tau_{ij}$6 to $\tau_{ij}$7, with both improvements significant at $\tau_{ij}$8 (Ghenai et al., 2016). The hybrid variant

$\tau_{ij}$9

targets cold-start users and yields coverage increases from $\sigma_{ij}$0 to $\sigma_{ij}$1 on Epinions and from $\sigma_{ij}$2 to $\sigma_{ij}$3 on Flixster (Ghenai et al., 2016).

A second line of work infers missing trust values rather than limiting attention to explicit trustees. iTrace constructs a dense estimated trust matrix $\sigma_{ij}$4 from a sparse explicit trust matrix $\sigma_{ij}$5 using direct trust, common explicit trustees, and Dijkstra-based longer-path propagation, then fuses $\sigma_{ij}$6 with Pearson similarity in the prediction rule (He et al., 2017). On FilmTrust, iTrace-I achieves the lowest MAE across all $\sigma_{ij}$7 values, at approximately $\sigma_{ij}$8 versus $\sigma_{ij}$9 for standard CF, and the attenuation term $$P_{a,i} = \bar r_a + \frac{ \sum_{u\in N(a,i)} (r_{u,i}-\bar r_u)\,w_{a,u} }{ \sum_{u\in N(a,i)} |w_{a,u}| },$$0 in the path-based trust formula is reported as crucial (He et al., 2017). Related implicit-trust work dispenses with explicit trust statements entirely by defining user–user and item–item implicit-trust correlations from average ratings, rating ranges, and overlap size; the hybrid user/item method then improves MAE, RMSE, and hit rate on MovieLens, Netflix, and Jester, with particularly strong behavior at small neighborhood sizes $$P_{a,i} = \bar r_a + \frac{ \sum_{u\in N(a,i)} (r_{u,i}-\bar r_u)\,w_{a,u} }{ \sum_{u\in N(a,i)} |w_{a,u}| },$$1 (Xuan et al., 2011).

Later systems broaden the notion of trust beyond direct social links. The adaptive similarity measure of Neshati et al. blends three components—rating similarity, direct-trust-group similarity, and indirect-trust-group similarity—through learned weights $$P_{a,i} = \bar r_a + \frac{ \sum_{u\in N(a,i)} (r_{u,i}-\bar r_u)\,w_{a,u} }{ \sum_{u\in N(a,i)} |w_{a,u}| },$$2, reaching on Epinions a MAE of $$P_{a,i} = \bar r_a + \frac{ \sum_{u\in N(a,i)} (r_{u,i}-\bar r_u)\,w_{a,u} }{ \sum_{u\in N(a,i)} |w_{a,u}| },$$3, RMSE of $$P_{a,i} = \bar r_a + \frac{ \sum_{u\in N(a,i)} (r_{u,i}-\bar r_u)\,w_{a,u} }{ \sum_{u\in N(a,i)} |w_{a,u}| },$$4, and coverage of $$P_{a,i} = \bar r_a + \frac{ \sum_{u\in N(a,i)} (r_{u,i}-\bar r_u)\,w_{a,u} }{ \sum_{u\in N(a,i)} |w_{a,u}| },$$5, compared with $$P_{a,i} = \bar r_a + \frac{ \sum_{u\in N(a,i)} (r_{u,i}-\bar r_u)\,w_{a,u} }{ \sum_{u\in N(a,i)} |w_{a,u}| },$$6, $$P_{a,i} = \bar r_a + \frac{ \sum_{u\in N(a,i)} (r_{u,i}-\bar r_u)\,w_{a,u} }{ \sum_{u\in N(a,i)} |w_{a,u}| },$$7, and $$P_{a,i} = \bar r_a + \frac{ \sum_{u\in N(a,i)} (r_{u,i}-\bar r_u)\,w_{a,u} }{ \sum_{u\in N(a,i)} |w_{a,u}| },$$8 for baseline CF (Zarei et al., 2019). Graph-embedding approaches “filter” the trust network into latent user vectors $$P_{a,i} = \bar r_a + \frac{ \sum_{u\in N(a,i)} (r_{u,i}-\bar r_u)\,w_{a,u} }{ \sum_{u\in N(a,i)} |w_{a,u}| },$$9, then perform cosine-based $N(a,i)$0NN; random-walk methods such as DeepWalk and Node2vec are reported as best across Epinions, Ciao, and FilmTrust, with user coverage of $N(a,i)$1 for all graph embeddings and a particularly large gain over directed trust on Epinions, where directed trust covers only $N(a,i)$2 of cold-start users (Duricic et al., 2020). Regular-equivalence filtering via Katz similarity likewise densifies sparse trust, improving cold-start ranking on Epinions from $N(a,i)$3 to $N(a,i)$4 in nDCG and from $N(a,i)$5 to $N(a,i)$6 in Recall at $N(a,i)$7 (Duricic et al., 2018).

A further extension is multi-faceted trust, in which social links are combined with anonymous public feedback. LOCABAL$N(a,i)$8 weights each observed rating by a multi-faceted trust term $N(a,i)$9 that blends reputation and contribution quality, while also modulating social regularization by $\tau:\mathcal G\to[0,1]$0 (Ardissono et al., 2019). On Yelp-Hotel and Yelp-Food, LOCABAL$\tau:\mathcal G\to[0,1]$1 outperforms U2UCF, SocialMF, LOCABAL, and SVD++ in MAP@10, F1@10, RMSE, and MAE, and the no-social variant is reported to retain almost the same MAP with slightly better RMSE/MAE (Ardissono et al., 2019). By contrast, the U2UCF-based multi-faceted model of Ardissono et al. shows a domain-dependent result: on Yelp, MTRTrust2 is best with Precision@10 $\tau:\mathcal G\to[0,1]$2, Recall@10 $\tau:\mathcal G\to[0,1]$3, F1@10 $\tau:\mathcal G\to[0,1]$4, RMSE $\tau:\mathcal G\to[0,1]$5, and MAE $\tau:\mathcal G\to[0,1]$6, but on LibraryThing all multi-faceted variants perform worse than simple rating CF or LOCABAL (Mauro et al., 2020). This is a central corrective to the common assumption that adding more trust evidence is uniformly beneficial.

3. Learned trust scores for weak supervision and distillation

In recent weak-to-strong generalization, trust filtering is formulated explicitly as a data-selection problem. Uzunoğlu et al. define a trust function $\tau:\mathcal G\to[0,1]$7 over teacher representations $\tau:\mathcal G\to[0,1]$8, instantiated as a small residual MLP, the “Neural Trust Function (NTF),” trained on binary correctness labels from a labeled source dataset with sigmoid final activation and binary cross-entropy loss (Uzunoglu et al., 31 May 2026). The filtering stage retains either the top-$\tau:\mathcal G\to[0,1]$9 weak labels or those above a threshold $t_i=\tau(g_{\pi_{\mathcal W}}(x_i,\hat y_i))$0, and the selected set $t_i=\tau(g_{\pi_{\mathcal W}}(x_i,\hat y_i))$1 is used to fine-tune the strong student. Ablations show that a single-token hidden representation yields AUCs of $t_i=\tau(g_{\pi_{\mathcal W}}(x_i,\hat y_i))$2 for correctness classification, “far outperforming raw output entropy,” and the retained subset is often reported as having purity $t_i=\tau(g_{\pi_{\mathcal W}}(x_i,\hat y_i))$3 (Uzunoglu et al., 31 May 2026).

The empirical consequence is near-lossless weak-to-strong transfer across several domains. On world knowledge, NTF-filtered students nearly match ground-truth fine-tuning across multiple teacher/student pairs, with no significant difference in $t_i=\tau(g_{\pi_{\mathcal W}}(x_i,\hat y_i))$4 settings and one setting in which NTF significantly exceeds ground truth (Uzunoglu et al., 31 May 2026). On quantitative reasoning, the method recovers $t_i=\tau(g_{\pi_{\mathcal W}}(x_i,\hat y_i))$5 of ground-truth gains on AIME and outperforms naive confidence, likelihood-based confidence, and off-the-shelf reward models even when teacher raw accuracy is $t_i=\tau(g_{\pi_{\mathcal W}}(x_i,\hat y_i))$6 (Uzunoglu et al., 31 May 2026). On chess puzzles, NTF achieves up to $t_i=\tau(g_{\pi_{\mathcal W}}(x_i,\hat y_i))$7 recovery and, in an iterative weak-to-strong chain, yields $t_i=\tau(g_{\pi_{\mathcal W}}(x_i,\hat y_i))$8 accuracy by the Qwen3-14B teacher stage, surpassing budget-matched ground-truth training at $t_i=\tau(g_{\pi_{\mathcal W}}(x_i,\hat y_i))$9 (Uzunoglu et al., 31 May 2026). The paper also gives a risk-controlled thresholding procedure based on a calibration set and a Hoeffding upper bound

$t_i\approx \Pr(\hat y_i=y_i)$0

with $t_i\approx \Pr(\hat y_i=y_i)$1 chosen as the most inclusive threshold satisfying $t_i\approx \Pr(\hat y_i=y_i)$2 (Uzunoglu et al., 31 May 2026).

A related but distinct instantiation appears in patent classification. Self-Filtered Distillation treats LLM-generated rationales as trust indicators rather than ground-truth supervision and computes three unsupervised trust metrics: Self-Consistency,

$t_i\approx \Pr(\hat y_i=y_i)$3

Class Entailment Alignment, and LLM Agreement Scoring; these are averaged into a Combined Trust Score,

$t_i\approx \Pr(\hat y_i=y_i)$4

Samples with $t_i\approx \Pr(\hat y_i=y_i)$5 are filtered, and retained samples are weighted by $t_i\approx \Pr(\hat y_i=y_i)$6 in the student loss (Yoo et al., 6 Oct 2025). On USPTO-2M, the reported performance is Micro-F1 $t_i\approx \Pr(\hat y_i=y_i)$7 and Macro-F1 $t_i\approx \Pr(\hat y_i=y_i)$8 for SFD, compared with $t_i\approx \Pr(\hat y_i=y_i)$9 for label-only distillation, $$[A_{\rm eff}(\theta,\phi)]_{ij} = A_{ij}\,\mathbf 1\{\tau_{ij}\ge \theta \wedge \sigma_{ij}\le \phi\},$$0 for Naïve CoT, and $$[A_{\rm eff}(\theta,\phi)]_{ij} = A_{ij}\,\mathbf 1\{\tau_{ij}\ge \theta \wedge \sigma_{ij}\le \phi\},$$1 for rationale-augmented distillation (Yoo et al., 6 Oct 2025). A threshold sweep identifies $$[A_{\rm eff}(\theta,\phi)]_{ij} = A_{ij}\,\mathbf 1\{\tau_{ij}\ge \theta \wedge \sigma_{ij}\le \phi\},$$2 as best for Micro-F1 and Subset Accuracy, with Subset Accuracy $$[A_{\rm eff}(\theta,\phi)]_{ij} = A_{ij}\,\mathbf 1\{\tau_{ij}\ge \theta \wedge \sigma_{ij}\le \phi\},$$3, while $$[A_{\rm eff}(\theta,\phi)]_{ij} = A_{ij}\,\mathbf 1\{\tau_{ij}\ge \theta \wedge \sigma_{ij}\le \phi\},$$4 correlates with averaged human Likert scores at Pearson $$[A_{\rm eff}(\theta,\phi)]_{ij} = A_{ij}\,\mathbf 1\{\tau_{ij}\ge \theta \wedge \sigma_{ij}\le \phi\},$$5, compared with approximately $$[A_{\rm eff}(\theta,\phi)]_{ij} = A_{ij}\,\mathbf 1\{\tau_{ij}\ge \theta \wedge \sigma_{ij}\le \phi\},$$6 for any single metric (Yoo et al., 6 Oct 2025).

These studies shift trust filtering away from interpersonal trust and toward reliability estimation over model-produced artifacts. A plausible implication is that trust filtering has become a generic mechanism for converting uncertain supervision into a curated training distribution, rather than merely a social-network primitive.

4. Trust and uncertainty gating in influence networks

Trust filtering also appears as edge gating in dynamical systems. In Gated Network Credence, each directed relation $$[A_{\rm eff}(\theta,\phi)]_{ij} = A_{ij}\,\mathbf 1\{\tau_{ij}\ge \theta \wedge \sigma_{ij}\le \phi\},$$7 carries independent trust $$[A_{\rm eff}(\theta,\phi)]_{ij} = A_{ij}\,\mathbf 1\{\tau_{ij}\ge \theta \wedge \sigma_{ij}\le \phi\},$$8 and distrust $$[A_{\rm eff}(\theta,\phi)]_{ij} = A_{ij}\,\mathbf 1\{\tau_{ij}\ge \theta \wedge \sigma_{ij}\le \phi\},$$9, from which one derives net trust

$CTS(x_i)<T_0$0

and uncertainty

$CTS(x_i)<T_0$1

Influence is admitted only when $CTS(x_i)<T_0$2 and $CTS(x_i)<T_0$3, generating an effective adjacency $CTS(x_i)<T_0$4 that drives continuous-time Laplacian averaging $CTS(x_i)<T_0$5 (Masoumi et al., 23 Jun 2026).

Sweeping the threshold plane yields four regimes. The Pluralistic regime ($CTS(x_i)<T_0$6) preserves almost all original edges. The Selective regime ($CTS(x_i)<T_0$7) is dominated by net-trust filtering and keeps only high-$CTS(x_i)<T_0$8 edges. The Concordant regime ($CTS(x_i)<T_0$9) is dominated by low-uncertainty filtering and prunes edges with $\tau_{ij}$00. The Fortified regime ($\tau_{ij}$01) requires both high net trust and low uncertainty and therefore produces a very sparse graph, often fragmented into many components (Masoumi et al., 23 Jun 2026).

The paper’s most distinctive claim is the hub–periphery reversal. In the Selective regime, simulation evidence shows that post-filter in-degree increases monotonically with pre-filter degree, and the normalized left null vector $\tau_{ij}$02 of $\tau_{ij}$03 grows with degree, so hubs dominate long-run influence (Masoumi et al., 23 Jun 2026). In the Concordant regime, high-degree hubs accumulate both trust and distrust and therefore tend to exceed the uncertainty threshold; post-filter in-degree rises at small pre-filter degree but falls sharply for large pre-filter degree, $\tau_{ij}$04 becomes flat or decreasing in degree, and lower-degree agents gain leverage over the collective equilibrium (Masoumi et al., 23 Jun 2026). This is an important correction to purely structural views of influence: network topology alone does not determine influence concentration once ambivalence is explicitly filtered.

The practical guidance is correspondingly threshold-centric. One estimates empirical distributions of $\tau_{ij}$05 and $\tau_{ij}$06, chooses $\tau_{ij}$07 and $\tau_{ij}$08 as percentiles to control tolerance for low net trust and ambivalence, and inspects connectivity diagnostics such as the largest strongly connected component to avoid unintended fragmentation (Masoumi et al., 23 Jun 2026). The framework generalizes one-dimensional trust filtering by making uncertainty a coequal gate.

5. Security, IoT, and distributed trust infrastructures

In IoT trust computation, trust filtering is used to exclude malicious recommendations and stabilize temporal trust estimation. The model of Alreshidi et al. first determines a dynamic sliding-window length such that the window always contains between $\tau_{ij}$09 and $\tau_{ij}$10 ratings, chosen off-line by a bi-objective NSGA-II fitness $\tau_{ij}$11 (Ziauddin et al., 24 Aug 2025). It then fuses average trust $\tau_{ij}$12 and freshness $\tau_{ij}$13 through a weighted harmonic mean

$\tau_{ij}$14

followed by reward and penalty terms $\tau_{ij}$15 and $\tau_{ij}$16 to produce direct trust $\tau_{ij}$17 (Ziauddin et al., 24 Aug 2025). Recommendation filtering at the Community Server is performed by an $\tau_{ij}$18 one-dimensional subspace clustering algorithm over three direct-trust bins $\tau_{ij}$19, $\tau_{ij}$20, and $\tau_{ij}$21, using historical precision trust $\tau_{ij}$22 to decide which neighbor and “wrong” bins are admissible (Ziauddin et al., 24 Aug 2025). The paper reports approximately $\tau_{ij}$23 improvement in accuracy for detecting on-off attacks and a $\tau_{ij}$24 reduction in recommendation filtering time (Ziauddin et al., 24 Aug 2025).

Distributed security systems use analogous logic at the level of peer reports. SocialFilter assigns initial social trust $\tau_{ij}$25, updates direct trust through an exponential moving average

$\tau_{ij}$26

computes reporter trust as a maximum-product path in the trust graph,

$\tau_{ij}$27

and combines $\tau_{ij}$28, identity uniqueness $\tau_{ij}$29, and report confidence $\tau_{ij}$30 in a weighted average $\tau_{ij}$31 (0908.3930). In simulation on a $\tau_{ij}$32-node Facebook sample, SocialFilter characterizes $\tau_{ij}$33 of spam bot connections with confidence greater than $\tau_{ij}$34 while yielding no false positives (0908.3930). The trust filter here is neither a recommender neighborhood nor a training subset; it is a report-selection and report-weighting layer over distributed evidence.

A related systems inversion appears in verifiable in-network filtering for DDoS defense. VIF does not estimate trust in data sources; it makes the filter itself trustworthy by placing packet-filter logic and rule set $\tau_{ij}$35 inside a Trusted Execution Environment, attesting the enclave state, and exporting authenticated count-min sketches for verification (Gong et al., 2019). The prototype is reported to handle $\tau_{ij}$36 Gb/s per enclave and up to $\tau_{ij}$37 filter rules, while a parallel deployment can scale to $\tau_{ij}$38 Gb/s and $\tau_{ij}$39 filter rules (Gong et al., 2019). This broadens the trust-filtering landscape from “trust-based selection” to “trustworthy execution of filtering.”

6. Trade-offs, limits, and divergent meanings

Several design trade-offs recur across the literature. The first is locality versus propagation depth. Trust-Aware Neighbourhood deliberately stops at direct neighbors for efficiency and reports a runtime of approximately $\tau_{ij}$40 minutes on Epinions, compared with approximately $\tau_{ij}$41 hours for the original Massa–Avesani all-pairs trust-distance approach (Ghenai et al., 2016). By contrast, iTrace, Katz-based regular equivalence, and Dijkstra-style implicit trust methods exploit multi-hop propagation and obtain denser trust coverage and better cold-start accuracy (He et al., 2017, Duricic et al., 2018). This suggests that propagation depth is not intrinsically desirable; it is a computational and statistical trade-off.

The second is purity versus coverage. In recommendation, hybrid trust-aware methods expand the candidate set to recover near-complete coverage for cold-start users (Ghenai et al., 2016). In weak supervision, trust filtering does the opposite: it sacrifices coverage to raise label purity, often above $\tau_{ij}$42, or targets a prescribed noise rate through threshold calibration (Uzunoglu et al., 31 May 2026). In patent distillation, the best reported threshold $\tau_{ij}$43 retains only high-trust samples for maximal Micro-F1 and Subset Accuracy (Yoo et al., 6 Oct 2025). The same term therefore covers both expansion and contraction of the retained set, depending on whether the bottleneck is sparsity or noise.

The third is nominal accuracy versus adversarial robustness. In a first-principles study of recommendation on Epinions, pure CF attains lower MAE under normal conditions than pure trust, but under a “bad-celebrity plus bots” attack its MAE rises from $\tau_{ij}$44 to $\tau_{ij}$45 ($\tau_{ij}$46), whereas pure trust rises from $\tau_{ij}$47 to $\tau_{ij}$48 ($\tau_{ij}$49) and trust-only random walks rise only $\tau_{ij}$50 (Stefanopoulos et al., 2024). The combined trust-plus-item scorer achieves the best normal MAE at $\tau_{ij}$51 yet degrades to $\tau_{ij}$52 ($\tau_{ij}$53) under attack (Stefanopoulos et al., 2024). Trust filtering can therefore improve robustness, but fusion with rating-side information can reintroduce vulnerabilities.

A common misconception is that trust filtering is equivalent to explicit social-link filtering. The literature is broader: trust may be implicit and path-inferred (He et al., 2017), derived from rating behavior (Xuan et al., 2011), composed from anonymous reputation and contribution quality (Ardissono et al., 2019), inferred from hidden representations of a weak teacher (Uzunoglu et al., 31 May 2026), or jointly parameterized with distrust and uncertainty (Masoumi et al., 23 Jun 2026). Another misconception is that richer trust evidence always helps. The Yelp/LibraryThing contrast shows that multi-faceted trust can outperform state-of-the-art trust-aware recommenders in one domain and underperform simple baselines in another when global trust signals are sparse or noisy (Mauro et al., 2020).

Finally, the phrase has a distinct numerical-optimization usage in “Trust-Region Eigenvalue Filtering for Projected Newton,” where “filtering” refers to adaptive choice among Newton, clamped, and absolute-value eigenvalue projections based on the trust-region ratio

$\tau_{ij}$54

The method reports $\tau_{ij}$55 wall-clock speedups over always-clamp and always-abs variants (Chen et al., 2024). This is not social or epistemic trust. Its inclusion underscores that “trust filtering” has become terminologically polysemous: in one literature it filters information sources by estimated reliability; in another it filters curvature directions by trust in a local quadratic model. The unifying idea is selective admission under an explicit criterion of confidence.