cs.LGAug 24, 2026

Revelation Control

Authors: Qinyou Wang

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Revelation Control is the problem of choosing priced interventions that reveal hidden state only insofar as the revealed distinctions can change a consequential decision, while accounting separately for any useful progress created by the intervention itself. We develop this theory for learning systems, where states equivalent under declared current information can respond differently to future training and favor different actions. The framework defines decision-sufficient revelation and revelation depth, separates pure information value from productive reuse, embeds static Bayes refinement into state-dependent continuation value, and gives an exact cost-adjusted factorization criterion: an additional shallow coordinate is decision-nonredundant only when states sharing a scalar summary lie on opposite sides of the priced Stop/Continue boundary. We also give a target-independent protocol for model-specific instantiation and prove that bounded stop-flip risk alone cannot certify positive expected utility under unrestricted severity. Across Qwen2.5-7B and Mistral-7B-v0.3, deeper future-learning probes have positive decision value and productive reuse yields strict equal-compute utility advantages. Qwen additionally provides evidence for a decision-nonredundant shallow revealability regime; in Mistral, a scalar continuation architecture fit only on an independent development panel retains positive familywise-adjusted lower bounds on a disjoint target panel, consistent with scalar decision sufficiency within the tested architecture family and resolution. The evidence supports structural rather than numerical transfer: the decision theory, cost accounting, continuation logic, and evaluation protocol transport, while empirical proxies, coefficients, thresholds, and even the required shallow state dimension may be system-specific.

Figures & tables

Figure 1: Decision-relative revelation. A coarse information cell can contain learning states that require different terminal actions. Deeper future-learning responses are useful only when they separate that decision quotient; full hidden-state reconstruction is unnecessary.
Figure 2: Equal-budget compute geometry for the two-action H12 comparison. Restart probing discards both H4 trials and executes the selected action freshly; productive revelation probes both actions to H8 and retains the selected tested prefix. Both consume 20 policy-visible updates, but promotion buys twice the revelation depth in the two-action case.
Figure 3: Schematic margin–revealability geometry implied by Theorem 5.5 . At fixed compute price, larger current decision margins require greater state-dependent revealability to justify deeper future learning. The paired points illustrate the nonredundant case in which two states with the same current margin can fall on opposite sides of the stopping boundary because their revealability differs. The curve σh=g(∣mh∣)\sigma_{h}=g(|m_{h}|) illustrates the stronger value-redundant special case in which continuation value itself reduces to a scalar function. More generally, scalar decision sufficiency requires only that the accessible states at a fixed margin remain on one side of the priced boundary; residual revealability variation is allowed. The boundary is theoretical and schematic, not an empirical fit.
Method / theoryPartial trainingProspective training signalPath reuseHidden learning-state decision aliasingExplicit finite-sample value ledger
Freeze–Thaw [ 37 ]yeslearning curvespause/resumenono
Hyperband [ 25 ]yesintermediate performancecontinuation of survivorsnono
PBT [ 17 ]yespopulation performanceyesnono
DynaMiCS [ 14 ]yeslocal cross-domain slopesnonono
Risk-controlled early exit [ 35 , 19 ]noshallow/deep prediction signalnonorisk control, not utility ledger
Revelation Control (this work)yesfuture-learning responsepromotion separated from information valuecentral targetyes
Table 1: Novelty boundary relative to representative neighboring methods. Entries describe the central mechanism emphasized by each work, not every implementation variant.
QuantityIndependent panel A (336)Independent panel B (336)
Minimum point margin1.2136812 ⁣× ⁣10−41.2136812\!\times\!10^{-4}1.4878650 ⁣× ⁣10−41.4878650\!\times\!10^{-4}
Minimum Student- tt LCB6.9363323 ⁣× ⁣10−56.9363323\!\times\!10^{-5}9.4261533 ⁣× ⁣10−59.4261533\!\times\!10^{-5}
Minimum bootstrap LCB7.0252125 ⁣× ⁣10−57.0252125\!\times\!10^{-5}9.5137645 ⁣× ⁣10−59.5137645\!\times\!10^{-5}
Minimum max- tt LCB4.4947532 ⁣× ⁣10−54.4947532\!\times\!10^{-5}6.9209070 ⁣× ⁣10−56.9209070\!\times\!10^{-5}
Comparators satisfying criterion22/2222/22
Table 2: Independent Qwen2.5-7B validation against the 22-component current-information library. Each entry is the worst component across the declared comparator set; positive lower bounds are required simultaneously.
Figure 4: Two independent Qwen2.5-7B panels across all 22 current-information comparators. Open squares mark simultaneous max- tt 95% lower confidence bounds and filled circles mark realized point margins. Every lower-bound endpoint remains strictly positive; the two panels are not pooled.
BankXi -selected meanLCBKeep -selected meanUCB
T1T_{1}2.9246×10−42.9246\times 10^{-4}6.5097×10−56.5097\times 10^{-5}−4.1166×10−4-4.1166\times 10^{-4}−2.2471×10−4-2.2471\times 10^{-4}
T2T_{2}3.6323×10−43.6323\times 10^{-4}9.5547×10−59.5547\times 10^{-5}−2.8389×10−4-2.8389\times 10^{-4}−8.2269×10−5-8.2269\times 10^{-5}
Table 3: Independent-bank structural sign replication inside the fixed H4 frontier-ambiguity event. Bounds are one-sided 95% Student- tt bounds.
Model / evidence blockProductive pathEnd-to-end pointOne-sided 95% LCB
Qwen, independent panel 18.9913×10−58.9913\times 10^{-5}8.9515×10−58.9515\times 10^{-5}2.9246×10−52.9246\times 10^{-5}
Qwen, independent panel 21.4776×10−41.4776\times 10^{-4}1.4737×10−41.4737\times 10^{-4}8.8394×10−58.8394\times 10^{-5}
Mistral, independent family3.0939×10−43.0939\times 10^{-4}2.9748×10−42.9748\times 10^{-4}2.0016×10−42.0016\times 10^{-4}
Table 4: Equal-compute productive-revelation evidence across the two Transformer families. The two Qwen rows use independent productive-path panels and are not pooled.
Theoretical implicationQwen2.5-7B evidenceMistral-7B-v0.3 evidence
Decision-relevant shallow aliasingFixed H4 ambiguity cell contains H8-refined groups with opposite terminal action gaps in two independent banks.Using the H4 threshold fixed from development data, the bank-A-defined ambiguity cell contains 216/480 anchors, including 88 stable- Xi and 58 stable- Keep anchors whose terminal-gap signs agree across both banks (descriptive structural check).
Positive value of deeper revelationH8 beats the declared current-information library on two independent 336-anchor panels; the separate 480-history H8-over-H4 value is strictly positive.Exact-anchor H8-over-H4 value 2.2509×10−42.2509\times 10^{-4} with positive tt and bootstrap lower bounds.
Productive information-to-utility closureProductive path value is positive in two independent panels and yields strict 20-update end-to-end advantage over the strengthened restart frontier.Same-path reuse 3.0939×10−43.0939\times 10^{-4} and equal-compute end-to-end advantage 2.9748×10−42.9748\times 10^{-4} , both with positive lower bounds.
Adaptive Revelation ControlSupporting evidence for incremental decision value from a nonredundant revealability coordinate; a held-out bounded-risk safe-compute evaluation has positive panel net-value lower bounds.Consistent with scalar decision sufficiency at the tested resolution; a familywise structural validation of the scalar continuation architecture, fit only on the separate development panel, retains positive Bonferroni-adjusted exact-anchor tt and bootstrap lower bounds within the two-element pre-existing architecture family.
Risk–severity boundaryPositive held-out safe-compute behavior is explicitly limited to panel-level evidence without an unrestricted tail guarantee.Sub-2% stop–flip risk coexists with nonpositive risk-controller net value because a few flips have high severity.
Table 5: Theory-level evidence across the two Transformer families. The fixed-depth and equal-compute Mistral results use an independent target panel; the adaptive row evaluates a pre-existing two-architecture family with familywise-adjusted inference.
Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

DirectionStop rateStop–flip rateHarm eventsMax harm severityNet value
A →\to B26.67%0.833%12.083×10−42.083\times 10^{-4}+1.186×10−5+1.186\times 10^{-5}
B →\to A11.67%0.417%12.010×10−32.010\times 10^{-3}−4.331×10−6-4.331\times 10^{-6}
Appendix
Table 6: Held-out risk–severity diagnostic for a derived risk-calibrated stopping rule. Net value is relative to always-H8.
More in this topic

Active Information Gathering

Active information gathering studies how agents choose questions, tests, or observations that reduce uncertainty and improve decisions under interaction costs.

See all 26 papers

Explore similar work

CardsList
Sep 1, 2026

A Score Is Not a Policy: Measuring the Value of Adaptive Revision

As agentic systems become compound systems, increasingly important decisions move above task execution itself: when should a higher-level controller preserve the strategy guiding another process, and when should it revise it? We study this meta-level control problem in a hierarchical latent reasoner whose manager can retain or replace a commitment governing lower-level computation. Across three precommitted training seeds, learned revision timing produces qualitatively different policies, ranging from an almost deterministic early clock to substantially more state conditioned schedule distributions, yet none outperforms the best forced timing policy evaluated on the same frozen checkpoint. This separates state dependence from decision value: a controller can vary its actions with internal state without turning that variation into a reproducible task-performance benefit. A deeper intervention study on the original checkpoint shows that timing itself is consequential and order-sensitive, while exhaustive enumeration reveals that a strong fixed schedule captures most of the measurable value available from timing at this decision budget. Counterfactual PERSIST/REPLAN diagnostics further show why score-level evidence can be misleading when predictability is dominated by decision position rather than within-position discrimination. Together, these results argue that learned meta-level control should be evaluated along three separate axes: whether its score depends on state, whether that dependence changes realized behavior, and whether those changes capture outcome value beyond a strong non-adaptive policy.

Ayushi Chadha

Hierarchical RLAdaptive Computation
AskPDF
Oct 20, 2026

Decision Sufficiency of Learned Representation Interfaces for Safe Generative Control

A controller can guide a language model toward safe and useful responses by choosing when to assist, ask for clarification, or refuse. When the controller uses selected internal features, poor performance alone cannot distinguish missing information from limitations of its decision rules or training. We study decision sufficiency: whether the best controller using a representation can match the minimum average loss attainable with full state information, under the same strategies and evaluation criteria. For interfaces with finitely many outputs, we compute the optimal loss exactly and separate the three sources of excess loss. Pairwise checks can miss information loss: states with the same controller input may share an optimal strategy in every pair but not across the whole group. We bound the number of states needed to reveal a conflict and characterize the fewest outputs an unrestricted interface needs to avoid all conflicts. We audit selected features from sparse autoencoders, which encode model activations using relatively few active features, in Qwen2.5, SmolLM2, and Phi. On SmolLM2 responses scored by two human raters, restricting the controller to two selected on/off features increases its minimum average loss by 0.0345--0.1166 relative to full state information, even with the governing policy supplied. Both independently trained autoencoders show the increase on both evaluation sets. The audit identifies decision failures that require changing the controller's input, not merely its architecture or training.

Jipeng Li, Yanning Shen

University of California, Irvine

Language Model-Based ControlLanguage Model Safety EvaluationSequential Decision Making
AskPDF
Oct 22, 2026

Jev-DLC: Decision-Level Continual Learning from Typed Feedback

Individual preferences often become visible only after deployment, when a user accepts one decision or revokes it in favor of another. Jev-style typed interfaces make this feedback explicit by identifying the selected action and, under revocation, the decision it replaces. We study how these consequences persist across requests as Decision-Level Continual Learning from Typed Feedback (Jev-DLC). Jev-DLC treats each correction as both evidence about a recurring action identity and a directed relation between the preferred and rejected choices. It preserves these signals in complementary bounded channels above a frozen scorer and combines their scores to shape the next decision. The method retains no raw interaction history, adds no scorer call, and performs no runtime language-model backward pass. We also introduce TCPA-Bench, a chronological suite that credits a correction only through its effect on later decisions. On TCPA-Repeat, Jev-DLC reaches 44.83% user-macro accuracy overall, 2.43 points above Recent-4 history replay. After at least two earlier feedback events, it reaches 56.07%, a 7.54-point margin over the strongest reported comparison at this depth. These results show how explicit typed corrections can become compact causal state for continual personalization.

Yangyang Feng, Jianle Gan, Zhuoyan Feng +1

The Hong Kong University of Science and Technology (Guangzhou) · Sun Yat-sen University

User Preference Modeling
AskPDF