arXiv subjects

Gabriel Garcia

Publications and source records attributed to Gabriel Garcia.

9 recordsLinked to original sources

arXiv2026

Protection Is (Nearly) All You Need: Structural Protection Dominates Scoring in Globally Capped KV Eviction

We study KV cache eviction under a shared globally capped decode-time harness. Seven policies (LRU, H2O, SnapKV, StreamingLLM, Ada-KV, QUEST, Random) share a prompt-boundary vulnerability: without structural protection, they collapse to near-zero quality on six pure-transformer models (F1$\leq$0.064). Reserving 10\% of cache at each boundary recovers 69--90\% of the $C{=}2{,}048$ reference-ceiling quality on seven LongBench models at $C{=}256$ (13\% retention); a ten-model panel spans 68--98\%. An attention-mass pilot (Qwen2.5-3B, $N{=}30$) suggests why: the position-0 sink holds ${\sim}75\%$ of prefix mass, while other boundary tokens sit near ${\sim}0.41{\times}$ uniform expectation, so attention scorers retain the sink but still drop structurally critical tokens. With protection, simplified score-isolation variants are TOST-equivalent to LRU at $K{=}32$ ($\Delta{=}0.02$); at $K{=}8$, attention policies pairwise converge yet beat LRU by 0.011--0.021 F1 across $C{=}256$ and $C{=}512$. Faithful Ada-KV/QUEST add ${\sim}0.03$--$0.04$ F1 on Mistral-7B and Phi-3.5 beyond simplified variants. A NIAH-32K regime-transfer pilot on Qwen3-4B (decode vs.\ prefill, $C{\in}\{512,2048\}$) shows near-identical protection lifts (ratio 0.99--1.00). At 64K, protection helps but recovery is modest; faithful per-head scoring matches full-cache ceiling on Gemma-3-4B at 6.3\% retention only when the model already supports strong 64K retrieval without eviction. Overall: protection dominates; scoring differences are secondary once boundaries are guarded; per-head allocation gives a further modest gain.

cs.LG↗

arXiv2026

No Free Swap: Protocol-Dependent Layer Redundancy in Transformers

When researchers ask whether two transformer layers are "equivalent" for compression, they often conflate distinct tests. Replacement asks whether one layer's map can substitute for another's in place; interchange asks whether two layers approximately commute when their positions are swapped. Both are output-grounded swap-KL probes, but they need not agree: on pretrained transformers the protocol gap can change which layers look safe to prune by several-fold under the same evaluator, especially when replacement distances are high. We measure both protocols across checkpoints and architectures. On a Pythia training trajectory (410M and 1.4B), the replacement-interchange gap grows from initialization to convergence. Under one matched WikiText-2 contract at 8B scale, Qwen3-8B enters a divergent regime: interchange-guided removal is several-fold safer than replacement-guided at the same layer budgets, while Llama-3.1-8B ties the two protocols for pruning cost even though interchange KL is lower, showing metric gaps need not map one-to-one to removal. Before layer removal or merging, score both swap-KLs on the target checkpoint; the diagnostic requires only unlabeled forward passes.

cs.LG↗

arXiv2026

The Last Word Often Wins: A Format Confound in Chain-of-Thought Corruption Studies

Corruption studies, the standard tool for evaluating chain-of-thought (CoT) faithfulness, infer which steps are ``computationally important'' from accuracy loss when steps are corrupted. We show that when benchmark chains end with an explicit terminal answer line, as in GSM8K and MATH, these tests largely measure \emph{answer placement} rather than where intermediate computation is carried out. Using matched GSM8K examples, removing only the final answer statement while preserving all reasoning collapses suffix sensitivity by about $19\times$ for Qwen~2.5-3B ($N{=}300$, $p{=}0.022$). Conflicting-answer prompts, which contain correct reasoning but a wrong explicit final answer, drive accuracy to zero or near-zero at 7B across five open-weight model families; wrong-answer following is strong at 3B--7B and attenuates sharply at larger scales. Replications on MATH, within-stable comparisons at 7B, and suffix-free chains show the same pattern in different guises: corruption sensitivity tracks the location of explicit answer text, not a fixed computational depth in the reasoning. Generation-time probes indicate that final answers are rarely early-determined during generation (${<}5\%$ early commitment), yet consumption-time behavior systematically follows explicit answer text. The confound is therefore largely a readout effect when the chain is consumed. We propose a three-prerequisite protocol (question-only control, format characterization, and an all-position sweep) as a practical minimum for future corruption-based faithfulness studies.

cs.LG↗

arXiv2026

The Right Answer, the Wrong Direction: Why Transformers Fail at Counting and How to Fix It

Large language models often fail at simple counting tasks, even when items to count are in the prompt. We investigate whether this failure occurs because transformers do not represent counts internally, or because they cannot convert representations to the correct output tokens. Across three model families: Pythia, Qwen3, and Mistral, ranging from 0.4B to 14B parameters, we find evidence for the second explanation. Linear probes recover the correct count from intermediate layers with $R^2>0.99$, showing that the information is present. However, the internal directions that encode counts are nearly orthogonal to digit-token output-head rows ($|\cos| \leq 0.032$). In other words, the model stores the count in a form that the digit logits do not naturally read out. We localize this failure with two interventions. Updating only the digit rows of the output head (36,864 parameters) substantially improves constrained digit prediction (60.7--100.0% on four tasks), but it does not fix unconstrained generation (0%); we do not claim that digit-row repair fixes open-ended text. By contrast, small LoRA on attention Q/V (7.67M parameters) improves upstream routing and achieves 83.1%$\pm$7.2% in true greedy autoregressive generation (deployable fix). Logit-lens at layer 35 (entity counting; correct-digit rank): (i) median over 3 seeds drops from order-$10^4$ to 1; (ii) seed 42 shows $54{,}332 \to 838$ (median top-1 while one seed stays far below). Norm, logit-lens, and cross-task analyses generalize the bottleneck to counting, addition, and list length; nulls on MMLU and GSM8K and limited DROP transfer. These results identify counting failure as a geometric readout bottleneck, not an internal-representation failure: the model knows the count but the output pathway is misaligned with tokens needed to express it.

cs.LG↗

arXiv2025

AGNOMIN -- Architecture Agnostic Multi-Label Function Name Prediction

Function name prediction is crucial for understanding stripped binaries in software reverse engineering, a key step for \textbf{enabling subsequent vulnerability analysis and patching}. However, existing approaches often struggle with architecture-specific limitations, data scarcity, and diverse naming conventions. We present AGNOMIN, a novel architecture-agnostic approach for multi-label function name prediction in stripped binaries. AGNOMIN builds Feature-Enriched Hierarchical Graphs (FEHGs), combining Control Flow Graphs, Function Call Graphs, and dynamically learned \texttt{PCode} features. A hierarchical graph neural network processes this enriched structure to generate consistent function representations across architectures, vital for \textbf{scalable security assessments}. For function name prediction, AGNOMIN employs a Ren\'ee-inspired decoder, enhanced with an attention-based head layer and algorithmic improvements. We evaluate AGNOMIN on a comprehensive dataset of 9,000 ELF executable binaries across three architectures, demonstrating its superior performance compared to state-of-the-art approaches, with improvements of up to 27.17\% in precision and 55.86\% in recall across the testing dataset. Moreover, AGNOMIN generalizes well to unseen architectures, achieving 5.89\% higher recall than the closest baseline. AGNOMIN's practical utility has been validated through security hackathons, where it successfully aided reverse engineers in analyzing and patching vulnerable binaries across different architectures.

cs.SE↗

arXiv2024

A Fly on the Wall -- Exploiting Acoustic Side-Channels in Differential Pressure Sensors

Differential Pressure Sensors are widely deployed to monitor critical environments. However, our research unveils a previously overlooked vulnerability: their high sensitivity to pressure variations makes them susceptible to acoustic side-channel attacks. We demonstrate that the pressure-sensing diaphragms in DPS can inadvertently capture subtle air vibrations caused by speech, which propagate through the sensor's components and affect the pressure readings. Exploiting this discovery, we introduce BaroVox, a novel attack that reconstructs speech from DPS readings, effectively turning DPS into a "fly on the wall." We model the effect of sound on DPS, exploring the limits and challenges of acoustic leakage. To overcome these challenges, we propose two solutions: a signal-processing approach using a unique spectral subtraction method and a deep learning-based approach for keyword classification. Evaluations under various conditions demonstrate BaroVox's effectiveness, achieving a word error rate of 0.29 for manual recognition and 90.51% accuracy for automatic recognition. Our findings highlight the significant privacy implications of this vulnerability. We also discuss potential defense strategies to mitigate the risks posed by BaroVox.

cs.SD↗

arXiv2024

Hardware-efficient quantum error correction via concatenated bosonic qubits

In order to solve problems of practical importance, quantum computers will likely need to incorporate quantum error correction, where a logical qubit is redundantly encoded in many noisy physical qubits. The large physical-qubit overhead typically associated with error correction motivates the search for more hardware-efficient approaches. Here, using a microfabricated superconducting quantum circuit, we realize a logical qubit memory formed from the concatenation of encoded bosonic cat qubits with an outer repetition code of distance $d=5$. The bosonic cat qubits are passively protected against bit flips using a stabilizing circuit. Cat-qubit phase-flip errors are corrected by the repetition code which uses ancilla transmons for syndrome measurement. We realize a noise-biased CX gate which ensures bit-flip error suppression is maintained during error correction. We study the performance and scaling of the logical qubit memory, finding that the phase-flip correcting repetition code operates below threshold, with logical phase-flip error decreasing with code distance from $d=3$ to $d=5$. Concurrently, the logical bit-flip error is suppressed with increasing cat-qubit mean photon number. The minimum measured logical error per cycle is on average $1.75(2)\%$ for the distance-3 code sections, and $1.65(3)\%$ for the longer distance-5 code, demonstrating the effectiveness of bit-flip error suppression throughout the error correction cycle. These results, where the intrinsic error suppression of the bosonic encodings allows us to use a hardware-efficient outer error correcting code, indicate that concatenated bosonic codes are a compelling paradigm for reaching fault-tolerant quantum computation.

quant-ph↗

arXiv2023

Demonstrating a long-coherence dual-rail erasure qubit using tunable transmons

Quantum error correction with erasure qubits promises significant advantages over standard error correction due to favorable thresholds for erasure errors. To realize this advantage in practice requires a qubit for which nearly all errors are such erasure errors, and the ability to check for erasure errors without dephasing the qubit. We demonstrate that a "dual-rail qubit" consisting of a pair of resonantly coupled transmons can form a highly coherent erasure qubit, where transmon $T_1$ errors are converted into erasure errors and residual dephasing is strongly suppressed, leading to millisecond-scale coherence within the qubit subspace. We show that single-qubit gates are limited primarily by erasure errors, with erasure probability $p_\text{erasure} = 2.19(2)\times 10^{-3}$ per gate while the residual errors are $\sim 40$ times lower. We further demonstrate mid-circuit detection of erasure errors while introducing $< 0.1\%$ dephasing error per check. Finally, we show that the suppression of transmon noise allows this dual-rail qubit to preserve high coherence over a broad tunable operating range, offering an improved capacity to avoid frequency collisions. This work establishes transmon-based dual-rail qubits as an attractive building block for hardware-efficient quantum error correction.

quant-ph↗

arXiv2019

0-Step Capturability, Motion Decomposition and Global Feedback Control of the 3D Variable Height-Inverted Pendulum

One common method for stabilizing robots after a push is the Instantaneous Capture Point, however, this has the fundamental limitation of assuming constant height. Although there are several works for balancing bipedal robots including height variations in 2D, the amount of literature on 3D models is limited. There are optimization methods using variable Center of Pressure (CoP) and reaction force to the ground, although they do not provide the physical region where a robot can step and require a precomputation for the analysis. This work provides the necessary and sufficient conditions to maintain balance of the 3D Variable Height Inverted Pendulum (VHIP) with both, fixed and variable CoP. We also prove that the 3D VHIP with Fixed CoP is the same as its 2D version, and we generalize controllers working on the 2D VHIP to the 3D VHIP. We also show the generalization of the Divergent Component of Motion to the 3D VHIP and we provide an alternative motion decomposition for the analysis of height and CoP strategies independently. This allow us to generalize previous global feedback controllers done in the 2D VHIP to the 3D VHIP with a Variable CoP.

cs.RO↗