PyTorchKR Papers

최신 논문

총 95편
태그·기간 등 상세 필터

카테고리 탐색

arXiv:2611.06055v1cs.DS

Softmax Attention with Any Temperature in Fixed Dimension: Matching the Best Known Maximum-Inner-Product Bounds

Attention is central to modern AI models for language, vision, and scientific modeling. However, direct evaluation compares every query with every key, requiring quadratic time in the context length, which limits its scalability to long contexts. Consequently, characterizing the complexity bounds of attention is of fundamental importance. For every fixed dimension $d$, we give an algorithm for $m$ queries, $n$ key-value pairs, and additive error $\varepsilon$, with expected time $ \widetilde{\mathrm{O}}\bigl((mn)^{k/(k+1)}+m+n\bigr),\qquad k=\lfloor d/2\rfloor. $ For fixed dimension and a fixed number of value channels, soft-O hides factors of the form $\operatorname{poly}_d\bigl(b,\log(m+n),\log B,\log(1/\varepsilon)\bigr),$ where $b$ bounds input encoding lengths and $B$ bounds coordinate and value magnitudes. In particular, the dependence on $B$ and $1/\varepsilon$ is polylogarithmic, which allows arbitrary positive temperature. When $m=n$ and $d\ge2$, the bound is $\widetilde{\mathrm{O}}(n^{2-2/(\lfloor d/2\rfloor+1)})$: near-linear in dimensions two and three, with exponent $4/3$ in dimensions four and five and $3/2$ in dimensions six and seven. It improves the previous $\widetilde{\mathrm{O}}(n^{2-1/d})$ upper bound and matches the best known batch maximum-inner-product bound in each fixed dimension. Additionally, we reduce integer batch maximum inner product computation to one attention call with the same dimension and input sizes $m,n$.

arXiv:2611.05813v1cs.CC

Towards the Query Complexity of Sampling under Strong Log-Concavity and Log-Sobolev Inequalities

We study the query complexity of sampling from smooth distributions on $\mathbb R^d$ using first-order oracles. For $d$-dimensional strongly log-concave targets with condition number $\kappa=2$, we prove that any randomized sampler requires $\exp({\Omega(\log^{1/3} d)})$ queries even from a constant warm start. This rules out polylogarithmic dimension dependence and strictly separates strongly log-concave sampling from Gaussian sampling. For targets with $L$-smooth potentials satisfying an $\alpha$-log-Sobolev inequality with $\kappa= L/\alpha$, we prove an $\Omega(\kappa)$ query lower bound for $d\gtrsim\log^2\kappa$. This shows that log-Sobolev sampling is strictly harder than strongly log-concave sampling. Our proofs construct instance-dependent tests that target samples typically pass but outputs of algorithms making few queries typically fail. For strongly log-concave targets, we embed a sequence of hidden Gaussian into high-degree tensor polynomial with smooth gates, forcing any algorithm to uncover the layers sequentially. The log-Sobolev construction hides a constant fraction of the target mass in angular caps while controlling the log-Sobolev constant via delicate estimates.

arXiv:2611.04074v2cs.CC

The Complexity of ReLU Network Synthesis from a Parameter Multiset

We ask how hard it is to build a ReLU network when its parameter values are given in advance. An instance supplies an appropriate encoding of a target behavior, a multiset of rational numbers, and a bound on the number of neurons (output included). The question is whether one can choose the graph of a ReLU network within the bound, and place the numbers on its edges and biases so that it meets the target. This separates the numbers from the architecture. Matching two labeled examples exactly is already $\mathsf{NP}$-complete, so we ask what changes when the target must hold on every input or on average. Agreeing with a Boolean circuit on every Boolean input is $\Sigma_2^{\mathsf P}$-complete, already for four neurons. Agreeing with a given ReLU network on every real input lies in $\mathsf{DP}$ for every fixed neuron bound and is $\mathsf{DP}$-complete for three neurons. Approximating a ReLU network on the unit cube within absolute error one everywhere is $\Sigma_2^{\mathsf P}$-complete, even with no hidden neurons and all numbers equal and positive. Deciding whether some network has expected squared loss at most a threshold is $\mathsf{NP}^{\mathsf{PP}}$-complete for uniform Boolean inputs. For uniform inputs on the unit cube and for standard Gaussian inputs with a positive threshold, the same holds for two neurons when a positive risk gap, given in binary and possibly exponentially small, separates yes from no instances. Without a gap, the cube version is complete for guessing a polynomial-length string, then comparing a sum of exponentially many rationals with a threshold, all computable in polynomial time from the input and the guess.

arXiv:2009.03996v5math.LO

Combining Determinism and Indeterminism

Our goal is to construct mathematical operations that combine indeterminism measured from quantum randomness with computational determinism so that non-mechanistic behavior is preserved in the computation. Formally, some results about operations applied to computably enumerable (c.e.) and bi-immune sets are proven here, where the objective is for the operations to preserve bi-immunity. While developing rearrangement operations on the natural numbers, we discovered that the bi-immune rearrangements generate an uncountable subgroup of the infinite symmetric group (Sym$(\mathbb{N})$) on the natural numbers $\mathbb{N}$. This new uncountable subgroup is called the bi-immune symmetric group. We show that the bi-immune symmetric group contains the finitary symmetric group on the natural numbers, and consequently is highly transitive. Furthermore, the bi-immune symmetric group is dense in Sym$(\mathbb{N})$ with respect to the pointwise convergence topology. The complete structure of the bi-immune symmetric group and its subgroups generated by one or more bi-immune rearrangements is unknown.

arXiv:2611.05346v1cs.LG

Agnostic Learning by Imitating

Given a class of machine learning models and sample access to a data distribution with binary labels, which computational task is harder: training a discriminative model that distinguishes positive and negative labels as well as the best member of the class, or training a generative model whose outputs are indistinguishable from real data by all members of the class? We prove that these two fundamental tasks in computational learning theory are not just closely related, but actually equivalent to each other in complexity. One direction of this equivalence was previously known. Indeed, the powerful regularity lemma in complexity theory gives an efficient oracle reduction from the indistinguishability task (a.k.a. multiaccuracy) to the distinguishing task (a.k.a. agnostic learning) in the style of boosting algorithms. Our main contribution is an efficient oracle reduction in the other direction. Taking this idea further, we show that a multiaccuracy oracle is a remarkably powerful primitive, achieving not just agnostic learning but also a wide range of seemingly more demanding learning objectives via similarly efficient and direct constructions. Our results crucially use the power of making adaptive queries to the oracle, and stand in stark contrast to prior work, which showed that post-processing a single multiaccurate predictor is insufficient for even weak forms of learning. On a technical level, we draw a connection between our (offline) problem and a version of no-regret learning in the online setting, which allows us to use follow-the-leader type algorithms to modify the label distribution and aggregate the predictions of the multiaccuracy oracle.

arXiv:2611.05296v1cs.LG

Universal Hardness From Masking Channels

We prove a broad learning-theoretic generalization of the celebrated hardcore lemma concerning the average-case complexity of computing a Boolean function (Impagliazzo, 1995). Our starting point is the observation that the classic notion of a hardcore measure, which reweights a function's input distribution, is equivalent to an input-dependent masking channel that injects randomness into the function's output distribution until it appears fully pseudorandom. The analogue of density of the measure is the fidelity of the channel. From this viewpoint, we develop a general framework for certifying hardness that allows for non-Boolean functions $f$ and a wide range of notions of hardness besides the classic 0-1 loss, by choosing the right notion of fidelity via convex duality. We show the difficulty of learning $f$ using a class $\mathcal{C}$ to minimize a convex loss is tightly characterized by the maximum fidelity possible for a masking channel, which measures the least amount of noise the channel must add to hide the structure of $f$ and make it $\mathcal{C}$-indistinguishable from random noise. Using our framework, we derive a number of multiclass generalizations of the hardcore lemma, and reveal some unexpected subtleties in going to larger alphabets. We construct a universal masking channel that applies to many natural notions of hardness simultaneously. Such channels are dual to omnipredictors, which optimize multiple losses at once (Gopalan, Kalai, Reingold, Sharan, and Wieder, 2022). We observe a symmetry between masking channels and generalizations of the hardcore lemma using pseudoentropy. We present a tight characterization of pseudoentropy using the MaxEnt algorithm from machine learning. A long line of research has shown that learning and hardness are two sides of the same coin; our framework gives a crisp and general formulation of this principle via convex duality.

arXiv:2604.07233v2cs.LG

How Does Machine Learning Manage Complexity?

We provide a computational complexity lens to understand the power of machine learning models, particularly their ability to model complex systems. Machine learning models are often trained on data drawn from sampleable or more complex distributions, a far wider range of distributions than just computable ones. By focusing on computable distributions, machine learning models can better manage complexity via probability. We abstract away from specific learning mechanisms, modeling machine learning as producing P/poly-computable distributions with polynomially-bounded max-entropy. We illustrate how learning computable distributions models complexity by showing that if a machine learning model produces a distribution $\mu$ that models the distribution generated by a cryptographic pseudorandom generator at least as well as the uniform distribution does, then $\mu$ must be close to uniform. We view training as compression that removes the randomness from the data, keeping only its structure, and inference as adding randomness back to generate new outputs. A computationally bounded model cannot distinguish pseudorandomness from true randomness, so training removes both, treating anything it cannot efficiently predict as noise. The same holds at inference, where the randomness added back can be, and in practice is, pseudorandom. In the appendix, we make the connection to compression precise and, using a characterization of pseudoentropy by Vadhan and Zheng, show that the best loss achievable by an efficient model is the next-bit pseudoentropy of the data.

arXiv:2611.03472v1cs.LG

Tilted Convexity and an Exact Offline-Online Separation

Convexity has three equivalent descriptions: support by affine functions, monotonicity of the subdifferential, and increasing secant slopes. We weaken all three by one distortion with parameter $\gamma\in(0,1]$, which multiplies positive changes by $\gamma$ and divides negative ones by $\gamma$. The resulting classes are the convex functions at $\gamma=1$ and, for $\gamma<1$, are distinct and nested, with the parameter dropping by at most a square along the chain. For proper lower semicontinuous functions in these classes, the effective domain is convex, local minimizers are global, and the function is locally Lipschitz on the relative interior of its domain. The classes separate offline from online optimization over compact convex sets with efficient projection, with access to function values and bounded support vectors. A smooth member of any class can be minimized to additive accuracy $\varepsilon$ with $\mathrm{poly}(1/\varepsilon)$ first-order queries and projections. Online, with nonpositive losses and $\alpha$-regret measured against $\alpha$ times the best fixed point, a polynomial-time algorithm achieves $\gamma$-regret $O(\sqrt T)$ on the largest class. For every fixed, effectively specified $\gamma\in(0,1)$ and every $\alpha\in(\gamma,1]$, no randomized polynomial-time algorithm has sublinear expected $\alpha$-regret unless $\mathbf{NP}=\mathbf{RP}$, already for smooth concave polynomial losses in the smallest class, each minimized by one linear optimization. Thus, unless $\mathbf{NP}=\mathbf{RP}$, $\gamma$ is the exact online coefficient of each class, while the offline coefficient is one.

arXiv:2611.03442v1stat.ML

Subpolynomial-Query Proximal Sampling Beyond Log-Concavity

We study the gradient query complexity of sampling from a density $\pi(x)\propto\mathrm{e}^{-V(x)}$ on $\mathbb{R}^d$, where $V$ has a $\beta$-Lipschitz gradient but need not be convex. Our approach uses the proximal sampler, which alternates Gaussian perturbation with sampling a position conditional on the noisy observation. We implement its observation chain using the high-order noisy conditional sampler of [OpenAI 2026], obtaining improved dimension dependence under functional inequalities. If $\pi$ satisfies a log-Sobolev inequality (LSI) with constant $\alpha$, then for every fixed $\nu>0$, there is a finite constant $C_\nu$ depending only on $\nu$ such that the algorithm returns a sample within total-variation distance $\varepsilon$ using at most $C_\nu\kappa(d+\kappa+\varepsilon^{-1}+\|\nabla V(x_0)\|^2/\beta)^\nu$ queries, where $\kappa=\beta/\alpha$ and $x_0$ is an arbitrary reference point for initialization. The analysis controls center-dependent local errors along exact proximal histories and removes the final observation noise by conditional decoding. Under a Poincar\'e inequality (PI), we obtain a corresponding guarantee from a warm start with a known bound on its R\'enyi divergence. An oracle lower bound shows that a stationary point alone, which suffices under LSI, is insufficient under PI.

arXiv:2611.03156v1cs.LG

Dissecting Geometric Constraints in Grokking: Interaction-Dependent Acceleration in Modular Arithmetic

Spherical transformer configurations can accelerate grokking through simultaneous modifications of several model properties. The present study investigated the contributions of these modifications in the published Sphere+Fourier configuration using 960 controlled modular-arithmetic runs. A complete factorial analysis varied residual projection, readout weight normalization, logit scale, and weight decay. Across four matched settings with weight decay, the successful-run mean crossing times without projection were 8-37 times those observed with projection. However, residual projection was insufficient to induce grokking without additional interventions. Various other unprojected configurations also generalized. Since projection altered the magnitude entering the readout, these comparisons could not separate geometry from output scale. A separate 100-run fixed-horizon study demonstrated faster first crossing for the complete configuration across five train/test partitions. Faster crossing did not necessarily correspond to greater persistence over extended training. Control experiments demonstrated rapid grokking without Fourier initialization. Task and depth sweeps also identified conditions in which the configuration failed to reach the criterion by 50K epochs. Accordingly, the observed acceleration depended on interactions among the components and did not extend uniformly across the tested tasks or depths.

arXiv:2611.02840v1cs.LG

Iterated Narrow Networks: Turing Completeness at Width 4 on Convex Input Spaces and the Complexity of Verification

We study the iterates of a fixed ReLU network $N$ of width 4, that is, the discrete-time system $x_{k+1} = N(x_k)$ on $[0,1]^4$. Equivalently, $x_k$ is the output of a weight-tied network made of $k$ copies of $N$, or of a recurrent network with no external input, started at $x_0$. We prove that every deterministic Turing machine $M$ admits such a network $N_M$. The network can be constructed in polynomial time from $M$. Each iteration has depth $\mathrm{poly}(|M|)$. The iterates of $N_M$ simulate $M$ exactly on every encoded tape. The number of iterations is proportional to the running time of $M$. Moreover, the fourth coordinate irreversibly flags every point of the convex cube $[0,1]^4$ whose simulation reads an invalid digit. This flag is the certificate that makes hardness reductions sound on convex input regions, the regime of practical verifiers. Convex regions were previously inaccessible because of a continuity obstruction inherent to ReLU networks. The correspondence has three consequences. First, reachability for iterated width-4 ReLU networks spans the full complexity hierarchy along the single axis of the time bound: it is NP-complete for a unary bound, NEXP-complete for a binary bound and RE-complete (undecidable) with no bound. Second, two extra neurons of width suffice to extend the hardness to every common smooth activation: at width 6, the robust verification problem (a promise problem with $\delta/2\delta$ gap) is NP-complete for $\tanh$, sigmoid, GELU, Swish, ELU, and any polynomial-time-computable continuous non-polynomial activation with a polynomial-time-computable Lipschitz constant. Third, the whole spectrum transfers to discretized Neural ODEs in dimension 5.

arXiv:2611.02344v1cs.DS

Efficient Proper Agnostic Learning of Halfspaces

We give a proper membership-query algorithm for agnostically learning halfspaces under the uniform distribution on $\{\pm1\}^n$. Given query access to an arbitrary function $f:\{\pm1\}^n \to \{\pm1\}$, our algorithm runs in time $\mathrm{poly}(n)+2^{\mathrm{poly}(1/\varepsilon)}$ and outputs a halfspace with error at most $\mathrm{OPT}+\varepsilon$, where $\mathrm{OPT}$ is the error of the best-fitting halfspace for $f$. Previously, even with membership queries, the best known running time was $n^{\mathrm{poly}(1/\varepsilon)}$, achieved by the sample-based algorithm of Kalai, Klivans, Mansour, and Servedio (FOCS 2005). Our algorithm is the first agnostic halfspace learner under the uniform distribution on $\{\pm1\}^n$ whose running time is polynomial in $n$ with an exponent independent of $\varepsilon$. Combining our algorithm with a boosting argument yields a proper agnostic learner using only random examples. It achieves the same $\mathrm{OPT}+\varepsilon$ guarantee, with running time matching that of Kalai et al.\ up to an additive term depending only on $\varepsilon$. To our knowledge, this is the first sample-based proper agnostic learner in the Boolean setting with running time $2^{o(n)}$ for every fixed $\varepsilon>0$.

arXiv:2610.40535v2cs.AI

Efficient Scalable Oversight via Rewinding

Scalable oversight asks how a weak verifier can check the outputs of computations performed by more powerful AI systems. One proposed approach uses relativizing interactive proofs, in which a weak verifier checks the output of a complex oracle-aided computation by interacting with one or more powerful provers. A prominent example is debate, which is highly efficient but assumes that one of two powerful AI models, or provers, is truthful. For certain classes of oracle-aided computations, this assumption can be removed, but at substantially greater computational cost. In this work, we introduce \emph{rewindable interactive proofs} (RIPs), a new model of relativizing interactive proofs for scalable oversight. In an RIP, the verifier may rewind the prover to an earlier state of the interaction and continue from there. This is possible when the prover is an AI model, since the verifier can save a snapshot of the model's state and later restore it. We construct a relativizing RIP for computations whose outputs are robust to changing a small fraction of oracle answers. Our construction is more efficient than prior relativizing proofs that do not assume a truthful prover, and it uses no cryptography. In our construction the honest prover runs in time linear in that of the original computation, and the verifier, after a one-time preprocessing step, runs in time nearly linear in the input length (which can be much smaller than the computation itself). We implement our protocol, report a preliminary MNIST experiment, and show that an RIP can be faster than a state-of-the-art deployed proof system. RIPs may be of independent interest. Rewinding is possible whenever the verifier controls the prover's machine, for example when the prover runs on trusted hardware or in a virtual machine whose state can be checkpointed.

arXiv:2610.38281v2cs.CC

A Witnessing Duality for ReLU Network Depth

A depth lower bound for ReLU networks asserts that every sufficiently shallow network fails on some input. What does a proof of such a bound compute? We study this question for exact computation of the maximum of $n$ real inputs, with depth measured by the number of hidden layers. For every separately fixed prime $q$ and depth bound $K$, we give a uniform $\mathsf{TC}^0$ refuter for rational networks with at most $K$ hidden layers and more than $q^K$ inputs, provided that the reduced denominators of their nonbias weights are not divisible by $q$. Biases may be arbitrary rationals. The refuter finds a rational input on which the network fails, and its correctness is provable in $\mathbf{VTC}^0$, the bounded-arithmetic theory of uniform constant-depth threshold computation. The underlying numerical depth lower bound follows from Averkov, Hojny and Merkert (ICLR 2025), extending Haase, Hertrich and Loho (ICLR 2023); our contribution is an internal proof that makes its geometric argument into a feasible refuter. On the construction side, $\mathbf{VTC}^0$ verifies locally checkable certificates for the balanced integer-weight maximum network of depth $\lceil\log_2 n\rceil$. At $q=2$, the construction and refuter meet: for each fixed $K$, the theory verifies that integer-nonbias-weight networks with arbitrary rational biases can compute the maximum with at most $K$ hidden layers exactly when $n\le 2^K$. We also give certified rational constructions of polynomial encoding length and depth $O(\log n/\log\log n)$. These results instantiate a witnessing duality: bounded proofs of certified existence and of refutation correspond, respectively, to threshold-computable constructors and refuters whose guarantees the theory proves.

arXiv:2610.09166v2cs.DS

병렬 확산 샘플링을 위한 하한

Lower Bounds for Parallel Diffusion Sampling

Standard diffusion samplers generate samples through sequential evaluations of a learned score function, with each query depending on previous answers. Parallel samplers seek to accelerate this process by making multiple score queries simultaneously. We ask how many adaptive rounds are necessary when each round allows polynomially many queries. We establish the first polynomial lower bounds on the round complexity of diffusion sampling with approximate scores satisfying a mean-square accuracy guarantee. Specifically, we prove (1) a $\widetilde{\Omega}(d^{1/3})$-round lower bound for sampling smooth, near-isotropic Gaussian mixtures in $\mathbb{R}^d$, and (2) a tight $\Omega(d)$-round lower bound for uniform sampling from anisotropic axis-aligned boxes contained in the unit ball. Both bounds hold for arbitrary randomized algorithms making polynomially many queries per round at arbitrary locations and noise levels, with inverse-polynomial score error and constant total variation accuracy. Our near-isotropic construction follows the blockwise approach of earlier lower bounds for parallel matroid search and discrete sampling \cite{KUW85,AGR24}. We use independent coordinate blocks containing random codebooks with well-separated rates and construct approximate score oracles whose responses in each block depend on consistency tests in earlier blocks. Together, the rate separation and these tests ensure that each round reveals information about at most one additional block with high probability.

arXiv:2606.13799v2cs.CC

The program is still there, and searching for it obeys a conservation law

The length of the shortest program that generates a sequence, its Kolmogorov complexity, is uncomputable, so we ask what finding a short program costs when one exists. Many search algorithms, including Levin search, evolutionary methods, simulated annealing and the cross-entropy method, learn about a candidate program only through its score, how much of the sequence it reproduces. For such algorithms we define the coupling width of a search problem, the largest number of program fields that must change together before the score moves. We prove an unconditional worst-case lower bound on the number of candidates scored, exponential in that width. Under stated conditions, a conservation law follows, in which each bit of structural knowledge injected into a search removes at most one bit of search. The structure plus the logarithm of the remaining search is at least a fixed fraction of the sequence's Kolmogorov complexity. That fraction tends to one as the instruction set grows, so the lower bound approaches Levin's 1973 upper bound, apart from the program's running time. Search by score must pay this cost to be complete. We prove that our engine's structural inference avoids it only by giving up coverage, as does, if one-way functions exist, every method that runs in polynomial time. Our deterministic engine, built on this theory, recovers a generating program, certified by compressing its data and predicting an unseen continuation, for 2,383 of 3,914 candidate sequences in four populations of integer sequences and cellular automata, including the chaotic Rule 30.

arXiv:2501.19214v3math.OC

A SPIDER-type stochastic subgradient method for expectation-constrained nonconvex nonsmooth optimization

Many real-world problems, such as those with fairness constraints, involve complex expectation constraints and large datasets, necessitating the design of efficient stochastic methods to solve them. Most existing research focuses on cases with no {constraint} or easy-to-project constraints or deterministic constraints. In this paper, we consider nonconvex nonsmooth stochastic optimization problems with expectation constraints, for which we build a novel exact penalty model. We first show the relationship between the penalty model and the original problem. Then on solving the penalty problem, we present a single-loop SPIDER-type stochastic subgradient method, which utilizes the subgradients of both the objective and constraint functions, as well as the constraint function value at each iteration. Under certain regularity conditions (weaker than Slater-type constraint qualification or strong feasibility assumed in existing works), we establish an iteration complexity result of $O(\epsilon^{-4})$ to reach a near-$\epsilon$ stationary point of the penalized problem in expectation, matching the lower bound for such tasks. Building on the exact penalization, an $(\epsilon,\epsilon)$-KKT point of the original problem is obtained. For a few scenarios, our complexity of either the {objective} sample subgradient or the constraint sample function values can be lower than the state-of-the-art results by a factor of $\epsilon^{-2}$. Moreover, on solving two fairness-constrained problems and a multi-class Neyman-Pearson classification problem, our method is significantly (up to 466 times) faster than the state-of-the-art algorithms, including switching subgradient method and inexact proximal point methods.

arXiv:2610.41086v1math.OC

Stochastic and Noisy Oracles in Smooth Continuous Optimization

We review models of inexact computational oracles in smooth continuous optimization and their implications for the convergence, computational complexity, and attainable accuracy of optimization algorithms. We adopt a framework in which zeroth-,first-, and second-order oracles are characterized by quantitative error bounds and may accept additional control parameters that influence both their accuracy and computational cost. We introduce deterministic and stochastic oracle models and illustrate their properties through a variety of examples, including finite-difference approximations, stochastic gradients, and randomized derivative estimators. Particular attention is given to the distinction between absolute and relative errors, bias and variance, and stochastic function evaluations with independent or coupled randomness. A central theme of the review is the use of common analytical tools to recover and compare convergence and complexity guarantees under different oracle models. For deterministic and stochastic first-order methods, we show how a general descent inequality leads to complexity bounds under various assumptions on the objective function and all proposed oracle models. In the stochastic setting, we develop bounds on expected stopping times and the accuracy that can be guaranteed under different oracle conditions. We also examine the total computational complexity of algorithms relying on specific oracle constructions and briefly review models of inexact second- order information and their use in optimization algorithms.

arXiv:2610.10611v2cs.AI

에이전트 AI의 검증과 자기 개선: 기초 및 한계

Verification and Self-Improvement in Agentic AI: Foundations and Limits

A self-improving agent can find accepted solutions more often without expanding the tasks its verification interface can accept. We extend task-relative verification with explicit support and transcript interfaces to distinguish search gains from changes in accepted-task coverage. A stage's native reach contains tasks accepted with default support; its closure frontier contains tasks accepted with any support the declared interface permits. A semantic containment criterion identifies the authority or exposure condition that every verified escape must violate, while guard-transfer invariance prevents a change of representation from being counted as a capability expansion. An executable Boolean-program repair agent makes the distinction concrete: over 36,864 paired runs, counterexample feedback reduces verifier calls without changing coverage, whereas an enforced permission change admits a new target with a proof excluding every old-interface program. Exact and uniformly gapped randomized verification share this account; independent majority amplification improves reliability while preserving the accepted language. Separate adaptive-selection contracts state what is additionally required to interpret a proposed successor's score as evidence of deployment improvement. Classical hierarchy results and finite audits provide supporting boundaries. The framework specifies the evidence needed to establish improved search, wider accepted capability, and reliable successor selection.

arXiv:2605.27381v4cs.CC

자기 개선 에이전트의 상대적 계산가능성 이론

A Relative-Computability Theory of Self-Improving Agents

Self-improving agents can acquire a procedure without knowing when its supporting teacher can safely be removed. We extend an oracle-relative model of evolving agents with explicit acquisition and completion contracts, separating learning a solver from certifying its independence. For a known finite-dimensional Boolean feature language, a query history determines exactly the answers in its feature span. A guarded protocol permits query-proposer revision while guaranteeing completion with the optimal number of teacher calls, and a supplied separating specification gives a finite-family extension. The resulting solver answers every input without its teacher and remains in the baseline computability layer. The central obstruction concerns supplying the completion specification: no baseline-effective procedure can extract a valid span or separating specification uniformly from arbitrary total programs, already in one feature dimension or with two candidate descriptions. Classical relative computation and limiting-revision results locate these claims without imposing a ceiling on practical improvement. Controlled experiments distinguish exact teacher removal from low-risk stopping and measure adaptation benefits and costs against a stronger greedy baseline. The framework makes the information required to acquire, certify, and deploy a capability explicit.