Task-Conditioned Measurements

• Task-conditioned measurements are procedures where the measured property is defined relative to task-specific variables, altering the observable's meaning based on context.
• They serve as compact intermediate representations that enhance signal extraction in domains such as quantum metrology, sparse transformer routing, and tool-output pruning.
• Their design directly influences performance in applications like thermodynamic metrology, motion planning, and robot manipulation by emphasizing task relevance and contextual fidelity.

Task-conditioned measurements are measurement procedures in which the measured quantity, its representation, or the preferred measurement property is specified relative to a task variable rather than fixed independently of context. The cited literature uses the term in several related senses: prompt-conditioned summaries of expert activation in sparse Mixture-of-Experts transformers; metrological measurement types whose utility depends on the sensing protocol; query-conditioned extraction of minimal evidence from tool outputs; real-time selection of multi-qubit Pauli observables; sparse task-relevant trajectory facts treated as noisy observations in diffusion planning; and scene-level affordance masks grounded from manipulation instructions (Avinash, 11 Mar 2026, Vetrivelan et al., 2023, Kovács, 4 Apr 2026, Blumoff et al., 2016, Kim et al., 30 Sep 2025, Liu et al., 2 Jun 2026).

1. Conceptual scope and formal conditioning

A common formal theme is that conditioning is not merely a label on data; it changes the meaning of the measurement itself. In operational quantum theory, an effect $a$ has probability $P_\rho(a)=\operatorname{tr}(\rho a)$, but if an operation $\mathcal I$ measures $a$, then the conditional probability of an effect $b$ is instrument-dependent: $$P_\rho(b\mid a)=\frac{\operatorname{tr}[\mathcal I(\rho)b]}{P_\rho(a)}.$$ For observables, if an instrument $\mathcal I$ measures $A$, then the observable $B$ conditioned on $A$ relative to $P_\rho(a)=\operatorname{tr}(\rho a)$0 is

$P_\rho(a)=\operatorname{tr}(\rho a)$1

This makes the conditioning context part of the observable definition rather than an external annotation (Gudder, 2023).

The same structural idea appears outside quantum foundations. In sparse transformers, routing signatures summarize expert usage induced by a prompt; in coding agents, the system is asked to measure relevance under a focused query and select the smallest verbatim evidence block; in motion planning, key states and timings are treated as uncertain observations of a trajectory; and in manipulation, a task instruction determines which functional part of a scene should be localized (Avinash, 11 Mar 2026, Kovács, 4 Apr 2026, Kim et al., 30 Sep 2025, Liu et al., 2 Jun 2026).

Domain	Task condition	Measurement output
Sparse MoE transformers	Prompt/task category	Routing signature
Thermodynamic metrology	WVA or repeated metrology without resetting	UB or NI measurement
Coding agents	Focused query and one tool output	Smallest verbatim evidence block
cQED and quantum theory	Chosen Pauli operator or prior instrument	Subset parity or conditioned observable
Diffusion planning	Key states and timings	Soft observations $P_\rho(a)=\operatorname{tr}(\rho a)$2
Manipulation	Task instruction $P_\rho(a)=\operatorname{tr}(\rho a)$3	Functional-part mask $P_\rho(a)=\operatorname{tr}(\rho a)$4

2. Statistical task-conditioning in machine learning systems

In sparse Mixture-of-Experts transformers, a routing signature is a layer-wise normalized summary of which experts are activated for a prompt. If $P_\rho(a)=\operatorname{tr}(\rho a)$5 is the number of times expert $P_\rho(a)=\operatorname{tr}(\rho a)$6 is activated at layer $P_\rho(a)=\operatorname{tr}(\rho a)$7 for prompt $P_\rho(a)=\operatorname{tr}(\rho a)$8, the layer-wise signature is

$P_\rho(a)=\operatorname{tr}(\rho a)$9

and concatenating all layer-wise distributions gives $\mathcal I$0. For OLMoE-1B-7B-0125-Instruct, $\mathcal I$1 and $\mathcal I$2, so the routing signature is 1024-dimensional. Similarity is computed by mean layer-wise cosine similarity,

$\mathcal I$3

Prompts from the same task category induce highly similar signatures: within-category similarity is $\mathcal I$4, across-category similarity is $\mathcal I$5, and Cohen’s $\mathcal I$6. A logistic regression classifier trained solely on routing signatures achieves $\mathcal I$7 cross-validated accuracy on four-way task classification with macro F1 of $\mathcal I$8. Permutation and load-balancing baselines show that the separation is not explained by sparsity or balancing constraints alone, and layer-wise effect size grows in deeper layers, peaking around layer 13. The paper releases MOE-XRAY as a lightweight toolkit for routing telemetry and analysis (Avinash, 11 Mar 2026).

A closely related but operationally narrower formulation appears in coding agents as task-conditioned tool-output pruning. The task is defined as: given a focused query $\mathcal I$9 and one tool output $a$0, return the smallest verbatim evidence block the agent should inspect next. The output is a set of contiguous line spans,

$a$1

over the original tool output. The released benchmark contains 11,477 examples: 9,205 SWE-derived examples, 1,697 synthetic positive examples, and 575 synthetic negative examples, spanning 27 tool types. On the manually curated 618-example test set, a Qwen 3.5 2B model fine-tuned with LoRA reaches precision $a$2, recall $a$3, strict F1 $a$4, exact $a$5, F1 $a$6, and compression $a$7. This removes 92% of input tokens while preserving most relevant evidence. The same paper reports that on 59 negative test examples the fine-tuned model returns empty output 80% of the time, whereas zero-shot Qwen 3.5 35B A3B does so only 7% of the time, indicating that task-specific supervision improves calibration for pruning (Kovács, 4 Apr 2026).

Taken together, these two lines of work suggest that task-conditioned measurement in machine learning often appears as a compact intermediate representation: not a solution to the end task, but a measurable, decodable, or executable summary of task-relevant structure.

3. Measurement design under thermodynamic constraints

In thermodynamically consistent metrology, task-conditioning appears as a preference between incompatible measurement desiderata. The relevant distinction is between unbiased measurements, which preserve the pre-measurement statistics in the pointer,

$a$8

and non-invasive measurements, which preserve the system’s reduced-state statistics,

$a$9

The same work defines faithfulness by

$b$0

An ideal measurement requires a non-full-rank pointer satisfying $b$1, but finite-temperature pointers are full rank, so ideal measurements are thermodynamically forbidden at finite resources. At nonzero temperature one may make the measurement as faithful as possible, but not simultaneously unbiased and non-invasive (Vetrivelan et al., 2023).

The metrological advantage then becomes explicitly task-conditioned. In weak value amplification, with interaction Hamiltonian $b$2, the task depends on faithfully transmitting statistics into the post-selected pointer readout. The paper therefore finds that unbiased measurement outperforms non-invasive measurement. It defines a thermally degraded weak value

$b$3

with $b$4 at any nonzero temperature, so the true amplification is always smaller than the ideal one. Finite-temperature transmon simulations show that for an IBM transmon example, pointer temperatures above about $b$5 mK become detrimental under NI with $b$6, whereas under UB the amplification depends only on system temperature; system temperatures below about $b$7 mK still allow $b$8, and below about $b$9 mK one gets near-target amplification (Vetrivelan et al., 2023).

For repeated metrology without resetting, the preferred property reverses. The probe evolves under $$P_\rho(b\mid a)=\frac{\operatorname{tr}[\mathcal I(\rho)b]}{P_\rho(a)}.$$0, is measured $$P_\rho(b\mid a)=\frac{\operatorname{tr}[\mathcal I(\rho)b]}{P_\rho(a)}.$$1 times per preparation, and $$P_\rho(b\mid a)=\frac{\operatorname{tr}[\mathcal I(\rho)b]}{P_\rho(a)}.$$2 is estimated by maximum-likelihood estimation using

$$P_\rho(b\mid a)=\frac{\operatorname{tr}[\mathcal I(\rho)b]}{P_\rho(a)}.$$3

Here the information-bearing quantity is the post-measurement system state, not just the readout statistics. NI converges to the true value with reasonably small variance, whereas UB converges to the wrong value with large standard deviation. After $$P_\rho(b\mid a)=\frac{\operatorname{tr}[\mathcal I(\rho)b]}{P_\rho(a)}.$$4 measurements, the reported purity is about $$P_\rho(b\mid a)=\frac{\operatorname{tr}[\mathcal I(\rho)b]}{P_\rho(a)}.$$5 for NI and $$P_\rho(b\mid a)=\frac{\operatorname{tr}[\mathcal I(\rho)b]}{P_\rho(a)}.$$6 for UB. The paper’s broader conclusion is that the optimal realistic measurement is not universal but depends on whether the task is readout-centric or state-preservation-centric (Vetrivelan et al., 2023).

4. Programmable quantum measurements and instrument-conditioned observables

Task-conditioned measurement also appears as programmable operator selection in superconducting hardware. A circuit QED implementation uses four highly coherent 3D transmon qubits collectively coupled to a single high-$$P_\rho(b\mid a)=\frac{\operatorname{tr}[\mathcal I(\rho)b]}{P_\rho(a)}.$$7 superconducting microwave cavity, with one ancilla qubit and three register qubits. By programming only the timing of echo pulses, the experimenters can choose in real time which register-wide Pauli operator or subset parity to measure, without changing the underlying hardware. The seven demonstrated nontrivial subset-parity observables on the three-qubit register are

$$P_\rho(b\mid a)=\frac{\operatorname{tr}[\mathcal I(\rho)b]}{P_\rho(a)}.$$8

Echo pulses determine whether each register qubit contributes an $$P_\rho(b\mid a)=\frac{\operatorname{tr}[\mathcal I(\rho)b]}{P_\rho(a)}.$$9 factor, refocusing its phase, or a $\mathcal I$0 factor, retaining its phase contribution (Blumoff et al., 2016).

The measurement is characterized at two levels. Detector behavior is reconstructed by quantum detector tomography using a POVM $\mathcal I$1, including a reduced form $\mathcal I$2, from which the specificity angle

$\mathcal I$3

is defined. Specificity is within roughly $\mathcal I$4. Detector J-fidelities are about $\mathcal I$5 unheralded and $\mathcal I$6 with heralding; detector S-fidelities are about $\mathcal I$7 unheralded and $\mathcal I$8 heralded. Conditioned process tomography then reconstructs the quantum instrument, whose J-fidelity is about $\mathcal I$9 unheralded and $A$0 with heralding, and whose S-fidelity is about $A$1 unheralded and $A$2 heralded. An additional success-heralding measurement verifies that the cavity has returned to vacuum and substantially improves both detector and instrument fidelities (Blumoff et al., 2016).

The foundational counterpart is the theory of conditioned observables and instruments. Bayes’ quantum second rule does not hold in general; it holds only under compatibility conditions such as $A$3, equivalently $A$4. For Lüders operations $A$5, Bayes’ second rule holds exactly when the effects commute. The same framework proves that two observables $A$6 and $A$7 are jointly commuting if and only if there exists an atomic observable $A$8 such that $A$9 and $B$0, and it derives a general uncertainty principle for conditioned observables. This formalism makes explicit that quantum conditioning depends on the instrument used to perform the prior measurement (Gudder, 2023).

5. Soft observations in planning and scene-grounded manipulation

In hierarchical diffusion motion planning, task-conditioned measurements are sparse trajectory facts treated as noisy observations. The measurements are key states $B$1 and timings encoded by a binary selector matrix $B$2, with observation covariance $B$3. They satisfy

$B$4

where $B$5 is the full trajectory. Conditioning a GPMP prior on $B$6 produces a task-conditioned posterior mean $B$7 and covariance $B$8, which are then used directly in a biased, non-isotropic diffusion process: $B$9 The standard DDPM is recovered when $A$0 and $A$1. The hierarchy separates prior instantiation from trajectory denoising: the upper level produces key states and timings, and the lower level denoises the full trajectory under the fixed structured prior. On Maze2D, success rates are $A$2 for isotropic without key states, $A$3 for isotropic with key states as conditioning, $A$4 for GPMP prior without key states, and $A$5 for GPMP prior with key states. On KUKA block stacking, the corresponding success rates are $A$6, $A$7, $A$8, and $A$9. The reported ablation conclusion is that structuring the corruption process offers benefits beyond merely conditioning the network (Kim et al., 30 Sep 2025).

In robot manipulation, scene-level task-conditioned affordance grounding treats the instruction as the measurement condition for a functional-part mask. The formal goal is

$P_\rho(a)=\operatorname{tr}(\rho a)$00

where $P_\rho(a)=\operatorname{tr}(\rho a)$01 is the scene image and $P_\rho(a)=\operatorname{tr}(\rho a)$02 is the task instruction. A2A-Bench explicitly includes both single-region and multi-region instruction correspondences, reflecting the one-to-many nature of affordance grounding in cluttered scenes. Its composition includes a point-prompting policy training corpus of about 40K single-object part masks, a human-verified core of 5,000 multi-object, multi-part scenes with about 18K single-part masks, and an initial batch of 5,000 generated multi-instance scenes after quality verification. The associated A2A-AffordGen pipeline combines language-model filtering, interactive part segmentation, instance-level mask-out refinement, task-reasoning instruction generation, and human verification. A2A-GroundingModel uses ORPS-to-TRPS adaptation and text-conditioned visual prompt injection, and is reported to run at about 9.7 FPS at $P_\rho(a)=\operatorname{tr}(\rho a)$03. The policy experiments indicate that A2A-Explicit achieves the best average success rate on LIBERO-object and performs best overall on real-world Piper tasks such as opening the microwave, stacking cubes, and placing objects (Liu et al., 2 Jun 2026).

A plausible implication is that, in both planning and manipulation, task-conditioned measurements act as structured spatial or temporal priors rather than hard constraints: uncertain key states anchor diffusion trajectories, and affordance masks provide robot-friendly spatial support for downstream action.

6. Evaluation regimes, misconceptions, and broader significance

A recurring point in this literature is that no single quantity completely encapsulates the performance of a measurement (Blumoff et al., 2016). Accordingly, different subfields use different evaluation regimes. Sparse transformer routing uses within-category and across-category similarity, effect sizes, permutation and load-balancing baselines, and linear decodability from routing signatures (Avinash, 11 Mar 2026). Tool-output pruning uses line-level precision, recall, strict F1, exact match, tolerant F1, and compression, with negative examples testing whether unsupported queries yield empty output (Kovács, 4 Apr 2026). Thermodynamic metrology uses post-selected Fisher information, maximum-likelihood estimation behavior, and purity after repeated measurements (Vetrivelan et al., 2023). Programmable quantum measurements use detector tomography, specificity, J-fidelity, S-fidelity, and conditioned process tomography (Blumoff et al., 2016). Diffusion planning uses success rates and mean absolute error between velocity and finite-difference position derivatives (Kim et al., 30 Sep 2025). Affordance grounding uses single-instance and set-based multi-instance protocols, real-time latency, and downstream policy success (Liu et al., 2 Jun 2026).

Several recurrent misconceptions are tested directly in the cited work. In sparse MoE models, the observed routing structure is not explained by balancing constraints alone, because within-task similarity lies above the load-balancing baseline while across-task similarity lies below it (Avinash, 11 Mar 2026). In coding agents, the task is not solved by large zero-shot models or lexical heuristics alone, because Qwen 3.5 35B A3B, Kimi K2, BM25, First-N, Last-N, and Random all underperform the fine-tuned small model on the curated test set (Kovács, 4 Apr 2026). In diffusion planning, simply conditioning the denoiser is not equivalent to structuring the noise model, because the full GPMP prior with key states clearly outperforms isotropic baselines (Kim et al., 30 Sep 2025). In affordance grounding, generic segmentation, VLM-based grounding, and mask-decoding MLLMs remain misaligned with scene-level one-to-many affordance localization in cluttered scenes (Liu et al., 2 Jun 2026). In thermodynamic metrology, no measurement type is uniformly best: UB is preferred for weak value amplification, whereas NI is preferred for repeated metrology without resetting (Vetrivelan et al., 2023).

Taken together, these results suggest a broad research program in which measurements are designed, selected, or summarized relative to the informational bottleneck imposed by the task. Under this view, task-conditioned measurements are not merely auxiliary diagnostics. They are task-specific observables, statistics, evidence blocks, priors, or spatial supports whose quality determines whether subsequent inference, control, or estimation remains aligned with the quantity the protocol actually cares about.