Sección

Investigación

Papers y preprints de los feeds de investigación que seguimos.

6265documentos

FPTR: Feasibility-Preserving Transactional Refinement for Deadline-Constrained Joint Beam and Resource Allocation

Short-interval wireless schedulers must return a complete, feasible beam-and-resource allocation before each execution deadline, and a late or invalid output gives no utility in the current control cycle. Searching longer can increase transmitted traffic, but it may also miss the deadline or leave only an infeasible intermediate solution. We propose Feasibility-Preserving Transactional Refinement (FPTR), which always keeps a validated feasible allocation, the incumbent. Each refinement stage builds a separate candidate and leaves the incumbent unchanged. A candidate replaces the incumbent only if it is complete, passes the feasibility validator before its stage cutoff, and strictly improves the objective; otherwise it is discarded. Four refinement stages use this rule: buffer-aware resource selection, compatibility-group (CG) user sharing, beam-mask rebuilding from remaining demand, and joint reassignment of two resources. On 150 fixed synthetic instances, the full pipeline (Full) raises traffic by 3.83% on average over a lightweight beam-first reference and by 1.79% over its Base constructor. In a separate suite with matched inputs, Full exceeds six independent search baselines (ALNS, tabu search, GA, simulated annealing, iterated local search, and GRASP) by 6.68-8.35%, where outputs later than 100 ms receive zero utility. All 750 Full runs in the internal suite returned feasible allocations, and none exceeded 100 ms. Separating mandatory output construction from optional refinement lets spare computation time improve the allocation without endangering the incumbent. The open-source implementation supports independent reproduction.

arXiv AI Papershace 9 días

Risk-Sensitive Non-Parametric and Kernel Bandits

We study non-parametric bandits in which arms are evaluated using Cumulative Prospect Theory (CPT) rather than expected reward. We consider continuous action spaces with arm-dependent noise and propose Non-Parametric CPT Bandits ({NPCB}), an optimistic algorithm that decouples mean estimation from CPT-based decision-making. We prove CPT regret of order ${O}({T\,_E(F,T^{-2})\,N(F,T^{-2},\|\|_)}), where _E is the Eluder dimension of the function class F, T is the time horizon, and N$ is its covering number in the sup norm. This matches the order-wise regret of classical non-parametric bandits, showing that CPT risk sensitivity incurs no additional order-wise regret cost. We extend {NPCB} to unknown noise CDFs, retaining the known-CDF mean-learning term and adding a CDF-learning cost governed by a distributional Eluder dimension. We also study Kernel-CPT bandits and obtain information-gain-based regret bounds for a broad range of kernels, matching the corresponding risk-neutral kernel rates up to logarithmic factors.

arXiv AI Papershace 9 días

Model Merging Does Not Require Uniform LoRA Ranks

Model merging combines separately fine-tuned experts into a single model without retraining. Standard practice trains every expert at the same LoRA rank, which couples the rank choice across experts even though each domain has its own best rank. We show this coupling is a property of the merge method, not of merging. We train 36 experts on three domains (mathematics, instruction following, multilingual) at four ranks and three model scales, then merge with six methods at uniform ranks and at each domain's own best rank. Giving experts different ranks costs averaging and utility-based selection nothing, and gains them up to 1.8 pp of multi-domain accuracy, while the conflict-resolution methods are inconsistent and can lose up to 4 pp. The split has a simple cause. Merging keeps only a fixed share of the pooled spectral components, and methods that choose them by size discard many that would have reduced loss. They also tend to discard the same expert: the instruction adapter has the smallest singular values, so it is dropped almost entirely and IFEval falls by 16 to 25 pp. Choosing components by their effect on loss instead keeps all three experts, and leaves accuracy nearly flat across an 8rank range (0.7 pp, against 6.7 pp for TIES at 4B scale). Rank can therefore be chosen per expert whenever the merge operator does not select by magnitude.

arXiv AI Papershace 9 días

What Does Multi-Turn Reinforcement Learning Really Improve in LLM Agents?

While multi-turn agentic reinforcement learning (RL) excels on long-horizon tasks, its performance gains remain highly sensitive to interaction constraints. Evaluating across sampling breadth K and interaction depth T via Pass@(K,T), we show that trajectory-level RL advantages depend on sufficient turn budgets, diminishing significantly when budgets tighten. Besides, agentic RL does not uniformly elevate single-step decision quality; rather, its strength concentrates in error recovery and long-horizon resilience. To mitigate the deployment risks caused by this budget dependence, we provide a preliminary discussion on the benefits and limitations of process supervision signals and introduce BudgetMix RL as an alternative that varies interaction budgets during training. Experiments demonstrate that BudgetMix RL achieves robust performance across diverse execution settings. Project Page: https://rethinking-agentic-rl.github.io/

arXiv AI Papershace 9 días

ArcRAG: Agentic Compositional Retrieval over Knowledge Graphs

Graph-based retrieval-augmented generation (GraphRAG) answers questions by retrieving over a knowledge graph built from a corpus. Yet retrieval strategies are designed around the structure their own construction pipeline produces, and whether they survive a change of construction is rarely tested. We study this cross-construction portability: whether one retrieval procedure, with no construction-specific schema, relation names, or retuning, stays effective when the graph is built differently. We introduce ArcRAG, which induces a query-local graph interface at inference time (the node labels, relation descriptors, and property keys reachable from the entities a question names), plans multi-hop traversals against it, and executes and repairs them as Cypher. We build each of the MuSiQue, HotpotQA, and TriviaQA corpora as a schema-guided graph (typed nodes, a closed relation set, source provenance) and as an open-information-extraction graph (untyped nodes, free-text relations), with the retriever's configuration fixed across them. We compare against a construction-agnostic Direct-Cypher agent, ablate each stage across retry and pass budgets on both graph families, and compare with results reported for prior GraphRAG systems, most of which design retrieval together with their own graph construction. Across the two constructions, we find that retrieval need not be co-designed with the graph: query-time interface induction keeps a construction-agnostic retriever effective when the construction changes, although construction still shapes absolute accuracy.

arXiv AI Papershace 9 días

The Alignment Tax Is Graded: Task-Selective Capability Loss Under Length-Normalized DPO with SFT Retention

Setting. This is an empirical characterization study; we propose no new method and run no comparisons against alternative objectives. We fine-tune Qwen3.5-4B with sparse LoRA and length-normalized DPO plus an SFT retention term (the RPO objective) on structured financial-filing tasks, and track downstream capability as the implicit reward margin grows to 7 β-scaled units (nf4 training; 3 under fp16 measurement on the unquantized base). Question. Does margin growth degrade capability uniformly across tasks, or selectively? Findings. In a pre-registered study (8 seeds, 6 unseen at freeze) with artifact-aware scoring and TOST equivalence testing, the alignment tax is graded and task-selective: extraction is preserved (Δ= -0.55 pts, equivalent at 5 pts), json loses little (Δ= -3.52 pts, equivalence fails on a clustered item-level interval), and router degrades severely (Δ= -30 pts). Router's loss concentrates in a margin 1-3 damage window containing a transient format-break spike. The margin is accounted for almost entirely (96\%) by suppression of rejected completions, with chosen reward approximately flat, consistent with the retention term anchoring it; we do not attribute the behavioral damage to any single gradient channel. Gold-logp deflation is task-selective (extraction -0.15, json -1.04, router -1.47 nats/token under fp16), with gold selectively protected relative to incorrect controls on extraction (8/8 seeds). QLoRA reward margins differ up to 3.4between nf4 training and unquantized fp16 measurement, with an axis-dependent gap. Scope. Length-normalized rewards with SFT retention, sparse LoRA on 8/32 layers, one 4B model, easily-separable synthetic preferences, three structured tasks. A no-retention control and a measured format-alignment metric are left to future work.

arXiv AI Papershace 9 días

How many sorted projections determine n points in k dimensions?

Let d_1,,d_m be directions in {R}^k, where k2. A multiset X of n points of {R}^k projects to the m multisets of numbers \{\{xd_j : xX\}\}, or equivalently to the m sorted lists of these numbers. How many directions are needed for these projections to determine every multiset of n points? Two simple bounds are known in every dimension: more than (k-1)_2 n directions are always needed, and (k-1)n+1 directions in general position always suffice. In the plane much more is known: Matousek, Privetivy and Skovron showed that almost every choice of O(n n) directions suffices. Whether anything similar holds in higher dimensions was left open by Dym, Wellershoff, Tsoukanis, Levy and Balan. We observe that two lemmas of Hair and Sahai, proved for list recovery of random linear codes, answer that question, and by tightening their argument we show that in every dimension almost every choice of (4k-2)(2n)(k_2 2n)+k+1 directions suffices. For k10 this beats the linear bound once n is larger than about 100: for example, 144 directions suffice for multisets of 1000 points in the plane, and an exact computation lowers this to 133. Letting some of the random blocks in their argument fail removes a factor k: almost every choice of O(knn) directions suffices, uniformly in k and n. We also show that 2k-1 directions never determine three points, so that the logarithmic lower bound is not tight; that every m directions in the plane fail on some pair of multisets of size O(1.78^m), improving the bound O(1.80^m) of Matousek, Privetivy and Skovron; and that for each k, m and n, either almost every choice of m directions works or almost none does.

arXiv AI Papershace 9 días

Microdroplet Generator Surrogate Enhancement and Inverse Design Through an Uncertainty-Guided Framework

Microfluidic droplet generation is critical for many applications, yet efficient device design in high-dimensional parameter spaces remains computationally prohibitive due to complex multiphase flow dynamics. This work presents a unified methodology that couples adaptive surrogate enhancement with simulation-driven inverse design. High-fidelity Volume-of-Fluid simulations provide ground-truth data, while neural network ensembles predict flow regimes and droplet diameters. Ensemble-based uncertainty quantification identifies epistemic uncertainty using consensus levels for regime classification and the coefficient of variation for diameter prediction, revealing that uncertainty concentrates near regime transition boundaries where physics is highly sensitive to parameter perturbations. A physics-informed enrichment strategy targets these critical zones and addresses class imbalances by deploying additional targeted simulations and retraining the surrogates forming the ensembles. A framework that integrates Particle Swarm Optimization for constrained inverse design, balancing accuracy with predictive confidence and physical consistency is therefore proposed. Validation on independent test datasets demonstrates that the enriched ensembles achieve higher accuracy, with the fraction of predictions within the 0-3% error range on droplet diameter nearly doubling, and high-error cases (above 10%) dropping substantially. The inverse design procedure reliably identifies physically realizable configurations meeting target droplet sizes. This approach significantly reduces computational costs while ensuring reliable predictions, offering a framework with potential applicability to the optimization of a broad range of microfluidic systems.

arXiv AI Papershace 9 días

Within-Family EfficientNet Scaling under Fixed-Resolution Transfer Learning for Spectral-Domain Optical Coherence Tomography Retinal Disease Classification

Background/Objectives: CNNs perform well in retinal disease classification from spectral-domain optical coherence tomography (SD-OCT), but whether EfficientNet's natural-image scaling hierarchy transfers to medical-image transfer learning at fixed resolution remains unclear. This study evaluates EfficientNet-B0-B7 for four-class SD-OCT classification and accuracy-efficiency trade-offs. Methods: Eight ImageNet-pretrained variants (4.05-64.10M instantiated backbone parameters) were evaluated at 224x224 on the Kermany dataset. Grayscale images were replicated to RGB and processed using model-embedded preprocessing and augmentation. Training comprised frozen-backbone head training followed by partial fine-tuning. Metrics covered classification, discrimination, efficiency, and visual analyses. Results: Test accuracy varied non-monotonically from 95.93% (B0) to 96.95% (B1); B7 achieved 96.73%. B1 had the highest accuracy and Drusen F1 (0.8780), while B7 had Drusen F1 of 0.8720 and the highest Drusen recall (0.9177). Drusen had the lowest F1 across all variants and contributed 82.2% of the summed squared deviation in class-specific F1; class ranking remained Normal > CNV > DME > Drusen. The OCT accuracy ranking differed from the reported ImageNet hierarchy, where B7 ranked first. The accuracy-to-parameter ratio fell from 19.82 to 1.48 percentage points per million total parameters. The final train-test gap had a weak, non-significant association with test accuracy (r = 0.26, p = 0.53). Conclusions: Larger EfficientNet variants did not consistently improve SD-OCT classification. B1 achieved the highest accuracy and Drusen F1, whereas B7 showed a secondary accuracy recovery at greater model size and CPU latency, indicating task and protocol dependent scaling effects. As cross-partition duplicates have not been independently audited, results are conditional on the supplied split.

arXiv AI Papershace 9 días

AmbiHTMLBench: A Benchmark for HTML-based Visually Rich Document parsing

Traditional document parsing systems use Markdown as their output representation, which is effective for linear text but does not fully preserve fine-grained spatial structure, table geometry, visual hierarchy, or rendering semantics. We show that HTML is advantageous for heterogeneous and layout-sensitive documents whose representations must remain editable, inspectable, and reusable, and introduce AmbiHTMLBench to systematically study this setting. In this benchmark, given a real-world document, a model must recover executable HTML that preserves its content, organization, and visual appearance. The benchmark contains 1,115 diverse pages with structured annotations and evaluates predictions in two different ways: The objective protocol aligns ground-truth and predicted rendered blocks and measures table fidelity, page-level text error, and a two-dimensional layout consistency; a complementary reference-anchored protocol measures perceived fidelity by aligning LLM judges with human preferences. Across ten primary multimodal systems, the best objective score is 0.8092, while the subjective winner is different, demonstrating that automatic decomposition and human-perceived fidelity capture complementary properties. We further fine-tune a small-parameter expert model and find that task-specific adaptation yields good performance on document-to-HTML reconstruction task. These results position AmbiHTMLBench as a diagnostic testbed for building smarter and more logical document parsing systems.

arXiv AI Papershace 9 días

SUMONE: A Simple and Efficient Summarizer for Plug-and-Play Agent Memory

Long-horizon language-model agents accumulate increasingly large interaction histories, making context compaction essential for efficient execution. Existing learned compaction methods often rely on downstream task outcomes or solver behavior, coupling training to specific tasks or agents. We introduce SumOne, a standalone summarizer trained to preserve information in agent histories through masked reconstruction. Given a summary of the expired history, a frozen reconstructor recovers randomly masked portions of the original trajectory; reconstruction quality, combined with a length penalty, provides a task-agnostic reward for optimizing the summarizer with GRPO. This yields self-supervision directly from raw trajectories, without reference summaries or downstream task execution. At inference time, only the lightweight summarizer is retained, enabling plug-and-play context compression for otherwise unchanged agents. We conduct a systematic evaluation of SumOne in long-horizon search and software-engineering tasks. SumOne reduces estimated inference cost by approximately 78\% on DeepSearchQA and 75\% on SWE-bench Verified relative to operating with full context, with absolute decreases of 0.04. The same checkpoint also transfers across downstream models and agent harnesses without additional fine-tuning.

arXiv AI Papershace 9 días

EvoGuard: Mitigating Degradation in Online Skill Evolution

Skills let language agents reuse experience across tasks as natural-language documents. Recent work evolves skills from interaction, but existing methods have limitations when skills evolve online. Prior methods update a skill offline from a batch of trajectories collected over many tasks. Some then adopt the new skill without any check, which can degrade the skill over time. Others keep it only if it improves performance on a held-out validation set, which is too slow and costly to run after every update. We introduce EvoGuard, a pairwise skill selector that mitigates skill degradation at low latency and cost. We first run the agent iteratively to build pools of skills and score each skill on a private test set to obtain utility labels. We pair the skills and use rejection sampling to keep only pairs with a clear utility gap. The selector is trained on the logits of two label tokens to pick the better skill. At deployment it decides in a single forward pass, which is significantly faster than evaluation on the private test set. Compared with adopting every new skill, EvoGuard improves success by 11.0% on ALFWorld, 5.9% on ScienceWorld, and 2.7% on LiveCodeBench.

arXiv AI Papershace 9 días

Graph Foundation Models Can Become Effective Multivariate Time Series Anomaly Detectors

Multivariate time series anomaly detection (MTSAD) relies on modeling both temporal dynamics and inter-sensor dependencies to identify complex system faults. Although graph-based detectors explicitly model sensor topologies, they train task-specific networks from scratch on limited target-domain data, remaining vulnerable to spurious correlations and unable to reuse transferable priors. While pretrained Graph Foundation Models (GFMs) offer transferable topological priors and anomaly discriminability, directly adapting them to MTSAD is hindered by cross-modal feature-topology mismatch, unreliable pseudo-anomaly supervision, and decision distortion during node-to-window aggregation. To fill these gaps, we propose a novel adaptation framework, Gfm4Mtsad, which unleashes pretrained GFMs for MTSAD. We introduce the Structure-and-Information Contract Alignment (SICA) module to bridge continuous time series and the pretrained GFM space. This module harmonizes principal subspace token projection, normal-prior-guided graph construction, and residual recovery to suppress spurious edges while retaining fine-grained temporal details. We propose a Mechanism-Aligned Pseudo-Anomaly Screening strategy (MAPS) to steer unsupervised learning. It applies localized dependency breaks while preserving empirical residuals, selecting hard negatives near the normal boundary for targeted supervision. We develop a Tail-Gated Evidence Calibration mechanism (TGEC) to eliminate node-to-window aggregation distortion. It calibrates multi-granular graph responses against normal operational references, strictly preserving benign score rankings while selectively refining extreme anomaly candidates in the tail distribution for reliable window detection and interpretable sensor evidence. Extensive evaluations with two GFM backbones demonstrate that Gfm4Mtsad achieves superior detection performance over existing baselines.

arXiv AI Papershace 9 días

Understanding the Out-of-Distribution Generalization of Chain-of-Thought Reasoning in LLMs

Chain-of-Thought (CoT) reasoning substantially improves Large Language Models (LLMs), yet its out-of-distribution (OOD) generalization mechanism remains theoretically underexplored. We study CoT as compositional OOD generalization: the target task lacks complete reasoning trajectories or direct input-output pairs, while training provides only in-distribution (ID) atomic step data. We show that CoT generalizes by repeatedly applying a shared single-step predictor to compose reusable ID transitions. Theoretically, let T denote the number of reasoning steps in the target task, n the number of training samples per sub-task, and k the number of output tokens per step; then the OOD task risk is controlled by the sum of the ID sub-task risks, scaled by a coefficient of order {O}(k) that is independent of T, although the summed risk itself may still grow with T. A Rademacher-complexity analysis further shows that the estimation term is of order {O}({T/n}), rather than linear in T, due to predictor sharing under a uniform capacity control on the shared class. For autoregressive Next-Token Prediction, the final-task risk is bounded by the best-in-class step-level risk scaled by the same coefficient of order {O}(k), plus an estimation term of order {{O}}(k{T/n}). A parity-coupled construction further shows that, for the sub-task empirical risk minimizer, the expected final-task risk is at least the best in-class risk plus Ω({T/n}) when Tn. Experiments on Atom-Task and GSM8K support the theory: CoT enables short-to-long and compositional generalization, while explicit state tracking is crucial. Our code is available at https://github.com/maxliang114514/CoT-OOD.

arXiv AI Papershace 9 días

Hardwired Distributed Processing Units for Inference: Neural Decisions at the Sensor that Free the Host

We show that a hardwired distributed processing unit (HDPU) can take a neural decision at the sensor and so remove the inference from the host. Built for zebra-crossing detection on a low-cost Artix-7 FPGA, the unit decides on every camera frame in a fixed 2,269,477 clock cycles (34.04 ms depends upon Fpga clk ), inside the 37.49-ms frame period, and on the board it returns the same scores and decisions as its bit-exact software model on all 639 test images. The whole detection path, from camera pixels to decision, takes 1,814 LUTs, with no processor, no operating system and no full-resolution frame store. When the unit answers instead of reporting "not sure", 95.7% of its answers on street photographs are correct, and a hardware contrast stage lifts accuracy on camera-like dashcam frames from 47.7% to 78.2%. The host is spared all of the inference: 8.13 ms of a core in every frame period, where a full-precision ResNet-18 would need more than a frame period even on 16 threads. It also receives 16 bytes per frame instead of 614,400, 38,400 times less. The unit is an 80x60-pixel pipeline feeding a four-layer binarized CNN with 93,408 one-bit weights, verified bit-exactly from a Python model through RTL simulation to the board. Earlier HDPUs did pre-processing (edges, corners, steering); this one shows that the decision itself can move to the sensor.

arXiv AI Papershace 9 días

EviSeek: An Evidence-Centric Agentic Framework for Embodied Active Information Seeking

Active embodied question answering requires agents to acquire evidence through viewpoint changes and physical interaction. The central challenge is to avoid undirected interaction: taking plausible actions without specifying what evidence they should reveal for answering the question. Action plans and accumulated histories alone can leave the informational purpose of an interaction implicit. We introduce EviSeek, a training-free framework that makes question-derived evidence demands explicit acquisition objectives, linking observation-conditioned actions and selectively retained evidence to the information the agent seeks. On 3,436 ESI-Bench questions, EviSeek improves accuracy over the Single-model agent by 7.1 percentage points with Qwen3-VL-32B and 3.7 points with Gemini 3.1 Pro, while improving visual-support-weighted accuracy and reducing mean token usage. These findings support making the informational purpose of interaction explicit, with action selection and evidence retention organized around what the agent needs to learn to answer the question.

arXiv AI Papershace 9 días

PARGAR: Pilot-Aware Graph Attention for Sparse-Pilot Neural OFDM Receivers

Reliable channel estimation and symbol detection in 5G New Radio (NR), as well as emerging 6G radio systems, rely on demodulation reference signals (DMRS). The density and placement of these pilots on the orthogonal frequency-division multiplexing (OFDM) time--frequency resource grid affect both receiver reliability and spectral efficiency. Sparse DMRS layouts increase the time--frequency distance between pilot observations and data resource elements (REs), which can make pilot information harder to exploit with purely local processing. This paper proposes PARGAR (Pilot-Aware Graph Attention Receiver), a graph-attention neural receiver that represents the OFDM resource grid as a graph for bit detection. The graph combines local time--frequency edges with pilot-anchor links from data REs to selected pilot REs, and we further study a channel-scaled pilot-anchor rule based on approximate coherence time and coherence bandwidth. Using Sionna-generated OFDM simulations, we evaluate PARGAR against DeepRx, a convolutional neural network (CNN)-based receiver, across dense, two-symbol sparse, and single-symbol sparse DMRS layouts under channel-mismatch scenarios. Under the considered SISO uncoded setup, channel-scaled PARGAR explicitly injects pilot-aware inductive bias and yields lower BER in selected sparse-pilot scenarios while remaining competitive under dense pilots.

arXiv AI Papershace 9 días

Partially Interpretable Decision Trees

Partially interpretable models extend simple interpretable models with the option to defer complicated inputs to an accurate but uninterpretable black-box model. By deferring more inputs, the model gets more accurate; by deferring fewer, it gets more interpretable. Existing procedures to train such partially interpretable models build on linear models or decision rules. We propose three algorithms for decision trees: a partially interpretable extension of the existing Branches algorithm, which performs exact search to optimize the tree; a conversion algorithm that we run on top of standard CART, which is computationally more efficient; and a generalized CART whose splitting criterion takes into account the option to defer while growing the tree. Perhaps surprisingly, the performance of the three algorithms is similar across many realworld datasets. We can explain this through a formal bound on the potential gain of smart tree construction, provided the black box accuracy is sufficiently homogeneous across the domain. We also construct a synthetic dataset where the accuracy of the black box is deliberately inhomogeneous. There, our deferral-aware generalization of CART does outperform regular CART and matches the performance of exact search. Finally, we evaluate an application to concept bottleneck models. Partially interpretable decision trees can replace uninterpretable neural networks after the bottleneck. On undeferred inputs, the model is fully interpretable, and deferring preserves accuracy.

arXiv AI Papershace 9 días

D3D: Decoupled Diffusion Drafting for Asynchronous Speculative Decoding

Target-conditioned diffusion drafters predict candidate tokens in parallel, yet each new block must wait for updated target-model hidden states. This forces drafting and verification to run sequentially, limiting speedup even though each block is generated in one pass. We introduce D3D, a standalone diffusion drafter that removes this inter-round dependency. D3D combines blockwise denoising with auxiliary autoregressive supervision to achieve long accepted prefixes without target hidden states. This independence lets D3D draft successor blocks during verification in one packed pass, enabling asynchronous speculative decoding. We progressively prune layers according to their effect on acceptance length to reduce the cost of packed drafting. Removing five of 36 layers retains over 99\% of mean acceptance length across three benchmarks with Qwen3-8B as the target. Across seven benchmarks, D3D averages 5.00and 5.71decoding speedup with Qwen3-8B and Qwen3-32B targets, compared with 4.03and 3.67for DFlash. On GSM8K with Qwen3-32B, speedup reaches 7.48. Code is publicly available at https://github.com/fudan-generative-vision/D3D.

arXiv AI Papershace 9 días

Mind Control of LLMs via Closed-Loop Stimulation: Out-of-Distribution Robustness

Brains and large language models need not be considered equivalent systems to reveal an interpretable commonality: both exhibit high-dimensional, nonlinear dynamics whose behavior can only be modeled imperfectly. This creates a shared challenge whenever we seek to intervene in their dynamics. A controller may either rely on a model identified around a particular operating regime and risk failure as the system moves away from that regime, or explicitly account for uncertainty and a changing internal state. Neurostimulation has confronted an analogous problem for decades. The limitations of fixed, open-loop stimulation have motivated closed-loop and state-dependent approaches that condition intervention on the ongoing state of the neural system. Inspired by this principle, we formulate LLM steering as a feedback-control problem in which the network's internal state, rather than a fixed intervention alone, determines how the system is perturbed. We further account for the mismatch between tractable dynamical models and the nonlinear transformer they approximate through robust control. As one principled realization of this framework, we develop an H_-based steering controller that limits sensitivity to disturbances and linearization mismatches. Across multiple steering benchmarks, the controller improves performance over existing baselines, with particularly pronounced gains under out-of-distribution conditions and varying length complexities. Beyond controller performance, we introduce robust steerability, S_{rob}, a pre-deployment measure of tolerance to dynamics mismatch; across model-task pairs, it exhibits task-specific structure and varies systematically with model scale. Beyond robust LLM steering, our measure asks not merely whether control is possible, but how much uncertainty it can survive; a principle that may span artificial and biological neural systems.