Multi-Slot Probing: Concepts and Applications

• Multi-Slot Probing is defined as systematically interrogating or allocating across several candidate compartments rather than a single option, with distinct implementations in areas like hashing, sponsored search, LLM interpretability, audio processing, and biomedical engineering.
• In distributed systems and sponsored search, the technique leverages multiple independent probes to optimize load balance and attention quantification while meeting domain-specific performance constraints.
• In LLM interpretability, audio decomposition, and hyperthermia applications, multi-slot probing enables the disentanglement of overlapping signals and control of spatial-temporal distributions, leading to measurable performance improvements.

Multi-slot probing is a polysemous technical term used across several research literatures to denote procedures that interrogate, allocate, or decode across multiple slots rather than committing to a single candidate. In distributed systems, it refers to hashing each key multiple times and choosing the successor with minimum probe-to-node distance; in sponsored search, it denotes either slot-conditioned attention prediction over SERP modules or exploration across multiple ad positions; in mechanistic interpretability, it denotes recovery of separable current-entity and prior-entity representations from a single token’s residual stream; in audio and dialogue, it refers to probing unordered latent slots or induced span segments; and in biomedical engineering it denotes physical air slots in a coaxial antenna that shape electromagnetic deposition (Appleton et al., 2015, Villaizán-Vallelado et al., 30 Apr 2025, Sarma et al., 2010, Bogdan et al., 22 Apr 2026, Taenzer, 31 May 2026, Nguyen et al., 2023, Gas, 2020).

1. Terminological scope and shared abstraction

Across the cited literature, the term is not tied to a single formalism. The object called a “slot” may be a ring successor region, a sponsored-search position, a SERP module, an entity-specific subspace in a residual vector, a source-like latent channel, an induced phrase segment, or an antenna aperture.

Domain	Meaning of “slot”	Primary objective
Consistent hashing	Candidate successor nodes on a ring	Load balance with low memory
Sponsored search SERPs	Page modules such as direct-top or organic-bottom	Predict TFT, TFC, and “noticed” attention
Sponsored search auctions	Ad positions $j \in M$	Learn $\mu_{ij}$ while preserving truthfulness
LLM interpretability	Current-entity and prior-entity subspaces	Decode multiplexed entity bindings
MI-MPE and dialogue	Unordered source-like outputs or induced spans	Recover source/stem activity or slot boundaries
Hyperthermia	Air slots in a coaxial antenna	Control axial heating and SAR

The shared pattern is repeated querying or allocation over multiple candidate compartments, but the optimization target changes by domain: peak-to-average load ratio in hashing, attention quantification and rank quality in SERPs, truthful welfare maximization in auctions, linear decodability and causal usage in LLMs, permutation-invariant source decomposition in audio, boundary induction in dialogue, and longitudinal thermal coverage in hyperthermia (Appleton et al., 2015, Villaizán-Vallelado et al., 30 Apr 2025, Sarma et al., 2010, Bogdan et al., 22 Apr 2026, Taenzer, 31 May 2026, Nguyen et al., 2023, Gas, 2020). This suggests a family resemblance rather than a single standardized definition.

2. Multi-probe consistent hashing

In "Multi-probe consistent hashing" (Appleton et al., 2015), multi-slot probing is instantiated as multiple independent key probes on a consistent-hashing ring. Each live node is hashed exactly once to a position $H_{\text{node}}(\text{node}) \in [0,1)$. For a key, one computes $m=K$ independent hashes $H_{\text{key},j}(\text{key})$, finds the clockwise successor of each probe, measures the probe-to-successor distance, and returns the candidate with minimum distance. The operational data structure is a ring over the unit interval with nodes stored once; practically, the implementation uses an array of bucketed, sorted vectors of 64-bit hashes with inlined storage of about 6 nodes per bucket to keep successor lookups cache-friendly. The method requires no additional storage beyond the hash table because it stores only node positions and IDs, without per-node counters, load state, or balancing metadata (Appleton et al., 2015).

Its central guarantee is asymptotic load balance. With $K$ independent key probes per key and one hash per node, the most loaded node satisfies, with high probability,

$$L_{\max} = \frac{K}{K-1}\cdot \frac{1}{n} + o(1/n),$$

so that

$$\frac{L_{\max}}{L_{\text{avg}}} = \frac{K}{K-1} + o(1).$$

Setting

$K = 1 + \frac{1}{\epsilon}$

yields

$$\frac{L_{\max}}{L_{\text{avg}}} \le 1+\epsilon+o(1).$$

The analysis assumes random hashing from a universal ranged hash family, independence between the node hash and the $\mu_{ij}$0 key-hash functions, and $\mu_{ij}$1. The paper models successor distances with a distribution $\mu_{ij}$2 and derives

$\mu_{ij}$3

then uses McDiarmid’s bounded differences inequality to show concentration around the mean (Appleton et al., 2015).

The algorithm trades lookup cost for load balance. Lookup time is $\mu_{ij}$4, with $\mu_{ij}$5 expected for the bucketed ring structure and $\mu_{ij}$6 for a sorted array. Measured performance for $\mu_{ij}$7 is roughly 350–600 ns per assignment on a Xeon W3690, with memory about 22 bytes per node. The paper reports the empirical convergence $\mu_{ij}$8, $\mu_{ij}$9, $H_{\text{node}}(\text{node}) \in [0,1)$0, and $H_{\text{node}}(\text{node}) \in [0,1)$1 for the peak-to-average load ratio, while preserving consistent-hashing monotonicity under node arrival and departure. Adding a node moves approximately $H_{\text{node}}(\text{node}) \in [0,1)$2 of keys, and removing a node remaps approximately $H_{\text{node}}(\text{node}) \in [0,1)$3 of keys (Appleton et al., 2015).

3. Sponsored search: slot-conditioned attention and truthful exploration

In "AdSight: Scalable and Accurate Quantification of User Attention in Multi-Slot Sponsored Search" (Villaizán-Vallelado et al., 30 Apr 2025), multi-slot probing refers to predicting slot-level attention across heterogeneous SERP modules from mouse trajectories. Each page module is treated as a slot, with four primary slot categories in the study: direct-top, direct-right, organic-top, and organic-bottom. The input is an asynchronous multivariate time series of cursor events whose per-timestep features are normalized $H_{\text{node}}(\text{node}) \in [0,1)$4 coordinates, time spent at the coordinate, slot type at position, and normalized sequence index in $H_{\text{node}}(\text{node}) \in [0,1)$5. AdSight uses an encoder–decoder Transformer in which the encoder contextualizes the cursor trajectory and the decoder forms slot-conditioned queries from normalized slot centers $H_{\text{node}}(\text{node}) \in [0,1)$6 and slot-type embeddings. Cross-attention aligns slot queries to the cursor trajectory and produces per-slot predictions for Total Fixation Time $H_{\text{node}}(\text{node}) \in [0,1)$7, Total Fixation Count $H_{\text{node}}(\text{node}) \in [0,1)$8, and category-level “noticed” probabilities. Auxiliary AOIs are inserted to refine outside-slot cursor semantics, with the best configuration reported at $H_{\text{node}}(\text{node}) \in [0,1)$9 and auxiliary-loss weight $m=K$0. On the headline Seq2Seq time-series Transformer configuration, the paper reports TFT $m=K$1 and $m=K$2, TFC $m=K$3 and $m=K$4, and weighted classification $m=K$5, $m=K$6 (Villaizán-Vallelado et al., 30 Apr 2025).

A distinct sponsored-search usage appears in "Multi-Armed Bandit Mechanisms for Multi-Slot Sponsored Search Auctions" (Sarma et al., 2010), where multi-slot probing denotes exploration policies and allocation/payment rules for learning click probabilities $m=K$7 over advertisers $m=K$8 and slots $m=K$9 across $H_{\text{key},j}(\text{key})$0 rounds while maintaining incentive compatibility. In the unconstrained case of arbitrary $H_{\text{key},j}(\text{key})$1, deterministic non-degenerate DSIC mechanisms are characterized by strong pointwise monotonicity and weakly separatedness, with payments

$H_{\text{key},j}(\text{key})$2

This characterization implies $H_{\text{key},j}(\text{key})$3 worst-case regret for DSIC mechanisms in the multi-slot setting. Under separable CTRs, $H_{\text{key},j}(\text{key})$4, the paper gives truthful-in-expectation mechanisms based on bid-independent exploration, weak pointwise monotonicity, weakly separatedness, and exploitation by ranking advertisers with $H_{\text{key},j}(\text{key})$5; the experiments reported in the paper show regret scaling like $H_{\text{key},j}(\text{key})$6 (Sarma et al., 2010). In this setting, “probing” is not perceptual prediction but mechanism design under uncertainty.

4. Mechanistic interpretability in LLMs

In "Slot Machines: How LLMs Keep Track of Multiple Entities" (Bogdan et al., 22 Apr 2026), multi-slot probing is a mechanistic probe for multiplexed entity representations in a single token’s residual stream. Let $H_{\text{key},j}(\text{key})$7 denote the residual activation at token position $H_{\text{key},j}(\text{key})$8. The paper studies a decomposition

$H_{\text{key},j}(\text{key})$9

where $K$0 is a current-entity slot and $K$1 is a prior-entity slot. Rather than imposing explicit orthogonality constraints, the work defines slots implicitly through learned readout directions and then assesses orthogonality empirically. The multi-slot probe is a mixture-of-experts system with $K$2 linear slot readouts $K$3 and per-entity routers $K$4, with predictions

$K$5

The training loss sums cross-entropy over entities already introduced by token $K$6. In the main setup, prompts contain $K$7 entities with 4 sentences each, activations are taken at sentence-ending periods, and the primary probing model is Qwen3-32B at layer 45 (Bogdan et al., 22 Apr 2026).

The main empirical result is that single-token activations carry two separable and largely orthogonal representations. Routing heatmaps show a current-entity slot specialized to entity $K$8 on its own tokens and a prior-entity slot specialized to entity $K$9 on entity $$L_{\max} = \frac{K}{K-1}\cdot \frac{1}{n} + o(1/n),$$0’s tokens. Weight correlation between current and prior slots is $$L_{\max} = \frac{K}{K-1}\cdot \frac{1}{n} + o(1/n),$$1, and RSA second-order similarity is $$L_{\max} = \frac{K}{K-1}\cdot \frac{1}{n} + o(1/n),$$2. Probe accuracy rises sharply from one to two slots and then plateaus: for Qwen3-32B the reported accuracies are 29.7% with 1 slot, 47.0% with 2 slots, 54.2% with 3 slots, 56.3% with 4 slots, and 57.9% with 5 slots; independent per-$$L_{\max} = \frac{K}{K-1}\cdot \frac{1}{n} + o(1/n),$$3 probes reach 50.3% overall. Causal patching and steering separate linearly decodable information from functional use: prior-entity signals support relational inference such as sequence retrieval and conflict detection, but explicit factual retrieval relies on current-entity representations even when prior-entity information is linearly decodable. The paper further reports that open-weight models perform near chance on dual subject-verb-object syntax that forces two bindings on a single token, whereas frontier models such as Claude Opus-4.5 and Gemini-3-Pro succeed more consistently (Bogdan et al., 22 Apr 2026).

5. Slot-based decomposition in audio and dialogue

In "A Lightweight Slot-Attention Framework for Multi-Instrument Multi-Pitch Estimation" (Taenzer, 31 May 2026), multi-slot probing concerns whether a compact set of unordered latent slots can stand in for unknown sources in a music mixture. A mixture CQT is mapped to $$L_{\max} = \frac{K}{K-1}\cdot \frac{1}{n} + o(1/n),$$4 source-like pitch maps $$L_{\max} = \frac{K}{K-1}\cdot \frac{1}{n} + o(1/n),$$5, where $$L_{\max} = \frac{K}{K-1}\cdot \frac{1}{n} + o(1/n),$$6 is an upper bound on the number of active sources. Competitive slot attention assigns pitch-time tokens to learned slot queries with softmax normalization across slots, and Hungarian matching provides permutation-invariant supervision. The total objective combines mixture MPE, matched slot pitch loss, noisy-OR union consistency, slot activity and inactivity penalties, optional timbre supervision from a frozen self-supervised teacher, and optional polyphony losses. On URMP family-level decomposition, moving from fixed slot order to Hungarian matching raises family AP from 24.00 to 61.12 and family F1 from 33.91 to 65.91; on stem-level decoding, FiLM conditioning improves URMP stem F1 from 33.78 to 39.08 and timbre cosine from 0.8452 to 0.8902. The framework remains lightweight, with a base slot model of 1.266M parameters and the largest reported slot model at 1.685M parameters (Taenzer, 31 May 2026).

In "Slot Induction via Pre-trained LLM Probing and Multi-level Contrastive Learning" (Nguyen et al., 2023), the same phrase is used for inducing multiple slot-bearing spans in task-oriented dialogue without token-level slot annotations. The task is framed as Break/Tie prediction at inter-token boundaries, with evaluation by Break-F1, Tie-F1, and their harmonic mean. The method first applies Unsupervised PLM Probing to build a segmentation tree from a perturbed-masking impact matrix and then refines segment representations with segment-level and sentence-level contrastive learning over BERT-base-uncased embeddings. At inference, two adjacent tokens are Tie iff they fall in the same depth-$$L_{\max} = \frac{K}{K-1}\cdot \frac{1}{n} + o(1/n),$$7 segment. The reported full model reaches H-Mean $$L_{\max} = \frac{K}{K-1}\cdot \frac{1}{n} + o(1/n),$$8 on SNIPS and $$L_{\max} = \frac{K}{K-1}\cdot \frac{1}{n} + o(1/n),$$9 on ATIS, improving over UPL-only baselines; the refined encoder also improves slot-filling F1 on emerging intents, from $$\frac{L_{\max}}{L_{\text{avg}}} = \frac{K}{K-1} + o(1).$$0 to $$\frac{L_{\max}}{L_{\text{avg}}} = \frac{K}{K-1} + o(1).$$1 on SNIPS$$\frac{L_{\max}}{L_{\text{avg}}} = \frac{K}{K-1} + o(1).$$2 and from $$\frac{L_{\max}}{L_{\text{avg}}} = \frac{K}{K-1} + o(1).$$3 to $$\frac{L_{\max}}{L_{\text{avg}}} = \frac{K}{K-1} + o(1).$$4 on ATIS$$\frac{L_{\max}}{L_{\text{avg}}} = \frac{K}{K-1} + o(1).$$5 (Nguyen et al., 2023). Here, multi-slot probing is neither set prediction nor mechanistic readout, but sentence-wide induction of multiple semantic spans.

6. Physical probing with multi-slot coaxial antennas

In "Study on interstitial microwave hyperthermia with multi-slot coaxial antenna" (Gas, 2020), multi-slot probing refers to physical slot configurations in an interstitial applicator rather than latent or algorithmic slots. The antenna is a coaxial structure with a central conductor radius $$\frac{L_{\max}}{L_{\text{avg}}} = \frac{K}{K-1} + o(1).$$6 mm, outer conductor inner radius $$\frac{L_{\max}}{L_{\text{avg}}} = \frac{K}{K-1} + o(1).$$7 mm, outer radius $$\frac{L_{\max}}{L_{\text{avg}}} = \frac{K}{K-1} + o(1).$$8 mm, and plastic catheter outer radius $$\frac{L_{\max}}{L_{\text{avg}}} = \frac{K}{K-1} + o(1).$$9 mm. Air slots of height $K = 1 + \frac{1}{\epsilon}$0 mm are placed at axial positions $K = 1 + \frac{1}{\epsilon}$1 mm, $K = 1 + \frac{1}{\epsilon}$2 mm, and $K = 1 + \frac{1}{\epsilon}$3 mm, with spacing $K = 1 + \frac{1}{\epsilon}$4 mm. The study compares single-, double-, and triple-slot configurations using a 2D axisymmetric FEM model in COMSOL, solving a TM-mode electromagnetic problem at $K = 1 + \frac{1}{\epsilon}$5 GHz coupled to the transient Pennes bioheat equation. Power deposition is modeled by $K = 1 + \frac{1}{\epsilon}$6, and thermal response by

$K = 1 + \frac{1}{\epsilon}$7

The study’s principal finding is a geometry-dependent trade-off between axial coverage and radial penetration. Electric field and SAR peak near each active slot; multiple active slots produce multiple axial lobes with reduced per-slot peak levels compared to a single slot. At the same input power, double- and triple-slot configurations reduce peak temperature and radial reach $K = 1 + \frac{1}{\epsilon}$8 relative to single-slot antennas, while extending heating along the $K = 1 + \frac{1}{\epsilon}$9-axis. For example, in breast tissue at $$\frac{L_{\max}}{L_{\text{avg}}} \le 1+\epsilon+o(1).$$0 W, the single-slot configuration at slot 1 yields $$\frac{L_{\max}}{L_{\text{avg}}} \le 1+\epsilon+o(1).$$1 and $$\frac{L_{\max}}{L_{\text{avg}}} \le 1+\epsilon+o(1).$$2 mm, whereas the double-slot $$\frac{L_{\max}}{L_{\text{avg}}} \le 1+\epsilon+o(1).$$3 configuration gives $$\frac{L_{\max}}{L_{\text{avg}}} \le 1+\epsilon+o(1).$$4 and $$\frac{L_{\max}}{L_{\text{avg}}} \le 1+\epsilon+o(1).$$5 mm. The paper therefore recommends single-slot configurations for concentrated, deeper radial deposition and multi-slot configurations for elongated targets requiring greater longitudinal coverage, subject to careful power management and, in practice, possible cooling (Gas, 2020).

7. Cross-domain patterns, limitations, and recurring misconceptions

Several recurrent design patterns appear across these otherwise disparate uses. Independence or permutation-tolerance is often central: multi-probe consistent hashing relies on independent key probes and independence between node and key hash families; AdSight’s decoder is empirically permutation-robust to slot order; and MI-MPE explicitly treats slots as unordered and interchangeable, using Hungarian matching to avoid fixed output semantics (Appleton et al., 2015, Villaizán-Vallelado et al., 30 Apr 2025, Taenzer, 31 May 2026). A second recurring pattern is the separation between information that is present and information that is operationally used. The LLM study makes this explicit as a decodability–usage gap, showing that prior-entity information is linearly decodable but not used for explicit factual retrieval (Bogdan et al., 22 Apr 2026).

The term also invites several misconceptions. In hashing, the paper explicitly distinguishes multi-probe consistent hashing from open addressing methods such as linear or quadratic probing, because the former is an inter-node placement scheme with consistency and monotonicity guarantees under node churn rather than collision resolution within a single in-memory table (Appleton et al., 2015). In dialogue, slot induction addresses boundary induction only and does not assign slot types (Nguyen et al., 2023). In SERP attention modeling, mouse trajectories are treated as a strong proxy for gaze, but the paper notes a cursor–gaze proxy gap, especially during pure reading phases or on touch devices (Villaizán-Vallelado et al., 30 Apr 2025). In MI-MPE, slot-based architectures improve family-level decomposition decisively, but stem-level prediction remains more challenging and auxiliary timbre or polyphony cues do not consistently resolve source assignment (Taenzer, 31 May 2026). In hyperthermia, the reported findings derive from a 2D axisymmetric model with idealized EM boundaries, passive catheter modeling, and no explicit S-parameter or active-cooling analysis (Gas, 2020).

Taken together, these studies show that “multi-slot probing” names a recurring research strategy rather than a single algorithmic object. It can mean repeated geometric probes on a ring, slot-conditioned cross-attention over human interaction traces, incentive-compatible exploration over ranked positions, subspace-specific readout of multiplexed representations, permutation-invariant decomposition into latent sources, unsupervised induction of multiple semantic spans, or physical slot design in an applicator. The shared theme is systematic reasoning over multiple candidate compartments, but the mathematical structure, guarantees, and failure modes remain domain-specific.