Sección

Investigación

Papers y preprints de los feeds de investigación que seguimos.

6415documentos

Camera-Noise Residuals for Face-Swap Detection: Redundant, Not Complementary, and Why

Fusing a learned camera-noise fingerprint with an RGB appearance backbone is an appealing route to generator-independent deepfake detection, because the noise residual is grounded in image-formation physics rather than in the texture statistics of a particular generator. We test, on FaceForensics++, whether a Noiseprint++ residual channel carries information complementary to an RGB Xception backbone for face-swap detection. A three-model ablation (RGB-only, residual-only, late-fusion) shows that fusion does not improve over RGB alone and that the residual branch alone is near chance. A seven-level bottleneck diagnostic localizes the cause: the noise maps do carry a discriminative signal, but it is statistical---carried by the per-sample first and second moments (mean, variance, energy) of the residual---and the per-sample InstanceNorm layer placed at the noise-branch input, following the TruFor template, standardizes exactly those moments away (five-fold cross-validated AUC drops from 0.747 to 0.554). A context-crop control rules out cropping geometry, and two fixed-fusion variants that remove the bottleneck recover the statistical signal yet still fail to beat RGB on every dataset. We conclude that, on this manipulation distribution, the noise residual is redundant with RGB rather than complementary, and we give concrete guidance for practitioners adopting noise-residual fusion for face-swap detection.

arXiv AI Papersel mes pasado

TRACE: Diagnosing Verifier Brittleness in Agentic Evaluation

Verifier scores now serve as both benchmark metrics and training rewards for large language model (LLM) agents, and a change in score is routinely read as a change in capability. It may instead reflect a change in the evaluation. We introduce TRACE, a protocol that turns a score change from a verdict into a testable diagnosis: it applies a targeted change to one part of an evaluation, compares paired runs, checks whether the agent's behavior changed, and rescores unchanged trajectories to test whether the scoring rule is responsible. In a controlled suite of 25 synthetic tasks, renaming tools lowers a scripted agent's score by 0.250 even though it performs exactly the same operations; restoring the original names at scoring time closes the entire gap, while the same mutation exposes a genuine behavioral failure in a second agent. On public τ^2-bench tasks with four LLM agents, an initial 30-task study finds mixed reward changes whose one clear effect does not replicate. In a larger follow-up on 88 new tasks with repeated runs per condition, renaming tools or reformatting tool outputs leaves reward unchanged to within 0.10 for seven of eight agent-change pairs, whereas tool names that deliberately mislead lower every agent's reward by 0.20-0.44, showing that the setup can detect real effects. Identical reruns flip 15-36% of task outcomes, so single-run comparisons cannot separate presentation effects from run-to-run variation. Two frontier LLM judges give consistent verdicts when a fixed trajectory is presented differently, yet disagree with each other on 57% of the same records, largely because one grades procedure rather than outcome. TRACE thus separates what a score change says about the agent from what it says about the measurement.

arXiv AI Papersel mes pasado

Early Signatures of Memorization in Diffusion Models via Basin Geometry and Cyclic Denoising

Diffusion models generalize early in training and later reproduce individual training samples. Standard tests detect memorization only once one-shot generation produces near-copies, leaving a released model unaudited until its outputs fail. We show that memorization is encoded in the geometry of the learned energy landscape before it appears in generated samples, a state we call latent memorization. Using score divergence and basin volume, we find that localized basins form around training samples and separate them from held-out samples before the first memorized sample appears, with an onset that follows the same O(n) scaling as the memorization time. We probe these basins with cyclic denoising, which repeatedly applies partial noising and denoising. Under the exact empirical score, we prove that cycling started near an isolated training sample recovers it and returns to it over any finite number of cycles with high probability. In trained models, cycling recovers training images from CelebA and CIFAR-10 checkpoints whose one-shot samples contain no copies, and at a CelebA checkpoint with 0.1% one-shot copies, 500 cycles raise the memorized fraction above 30%. Cycling also reveals degenerate attractors that match no single training image and fade as training proceeds, so residence in a basin does not by itself imply memorization. These findings hold on a Gaussian mixture, CelebA, and CIFAR-10 across optimizers, architectures, noise schedules, and training-set sizes, and extend to off-the-shelf Stable Diffusion v1.4, where the cycled conditional-unconditional divergence gap separates memorized from non-memorized prompts with an AUC of 0.944 and a TPR of 0.866 at 1% FPR. More broadly, what a diffusion model has memorized is a property of the geometry and stability of its learned distribution, and assessing it requires examining this structure rather than generated outputs alone.

arXiv AI Papersel mes pasado

σTransfer: Uncertainty Transfer from Small to Large Networks under μP

Reliable predictive uncertainty in Laplace approximations depends critically on the prior precision, yet selecting it requires a posterior sweep that is prohibitively expensive for neural networks with billions of parameters. Under the Maximal Update Parametrization (μP), we derive a rescaling of the prior covariance that makes the selected precision stable as model width grows. This leads to σTransfer: we select the precision on a smaller model and zero-shot transfer it to the much larger model, i.e., without searching for the precision on the larger model at all. We show convergence of the prior kernel, posterior covariance, selected precision, and posterior-derived decisions under explicit conditions, and verify σTransfer across regression, image classification, and Transformer readouts. For example, measured precision-sweep speedups reach 5000when transferring from width 128 to 4096 on MNIST, at a target-NLL degradation of 0.002; transferring from a public 1B to 7B model gives a median search speedup of 2.3(up to 330), with a mean measured target-NLL increase below 10^{-4} across ten tasks. The same posterior stability also enables transfer of acquisition, OOD-detection, and abstention decisions without constructing a target posterior.

arXiv AI Papersel mes pasado

Evi-VN: Hard Region Guided Virtual Node Evidence Injection for GNN-Based Fraud Detection

Online platforms contain growing numbers of bots, deceptive reviewers, and scam accounts that imitate legitimate users. Such camouflage blurs graph neighborhoods and behavioral attributes, making it difficult for graph neural networks (GNNs) to distinguish both well-disguised fraudsters and legitimate users. Across diverse GNNs, we observe overlapping errors on a shared hard region, suggesting the presence of latent fraud evidence that graph topologies and standard features fail to capture. Fraud-specific GNNs can mitigate particular graph pathologies, yet they still make limited use of heterogeneous evidence such as structured records, text, images, and audio; uniform multimodal fusion may also disturb nodes already handled reliably by the graph. We propose Evi-VN to learn and correct these shared blind spots rather than build another fraud detector. To our knowledge, Evi-VN is the first graph fraud detection framework to use feature isolated evidence chains to correct hard regions shared across GNNs. Its evidence chains connect behavior, content, and context across structured, textual, visual, and acoustic sources, helping expose camouflage that graph neighborhoods may miss. Crucially, Evi-VN selectively applies this evidence only to likely hard samples via virtual class nodes, preserving both the reliable predictions and the input design of existing GNNs. Shared hard regions also let Evi-VN enhance generic, fraud-specific, and unseen GNNs even with imperfect evidence models. Experiments across bot, fake-review, refund-evidence, and telecom-fraud tasks validate these advantages.

arXiv AI Papersel mes pasado

Constrained Command-Conditioned Reinforcement Learning with Bandit Strategy Selection in Real-Time Strategy Games

Deep reinforcement learning agents reach strong performance in real-time strategy games but can be brittle against opponents outside their training distribution. Separating strategic command selection from learned unit control allows different strategies to be selected for different opponents while reusing the same execution policy. This requires an executor that can follow different commands and measurable criteria for assessing whether it does so. We introduce a constrained command-conditioned Proximal Policy Optimization (PPO) policy, the executor, for MicroRTS, a real-time strategy environment. Discrete commands specify strategic objectives and behavioral requirements for economy, army composition, military posture, and worker policy over multiple environment steps; the executor determines the unit-level actions used to fulfill them. A Thompson-sampling bandit acts as the strategist, selecting command tuples from an estimate of the opponent's strategy built from in-game observations rather than opponent identity. In a controlled comparison with a flat PPO baseline trained with the same architecture, budget, curriculum and self-play league, the strategist-executor system wins significantly more often against three of the four strongest opponents on a training map, including the two strongest held-out ones (0.55 to 0.97 and 0.01 to 0.34), with no significant difference against the others.

arXiv AI Papersel mes pasado

Lamarck's Driving School: Discovering Autonomous Driving Training Strategies through Evolutionary Competition

Autonomous driving capabilities depend strongly on the distribution of scenarios encountered during training. Existing methods commonly construct or dynamically adapt training scenario distributions using surrogate criteria such as realism, difficulty, or risk. However, these predefined surrogates may misrepresent training value, leading to inefficient use of training resources. To address this limitation, we propose a Lamarckian evolutionary framework that replaces surrogate-based guidance with competition among candidate distributions. We formulate training strategy discovery as a multi-stage bilevel optimization problem and use Lamarckian evolution algorithm to approximate its solution. At the outer level, Darwinian crossover, mutation, and selection explore the scenario distribution space; at the inner level, policy learning acquires new capabilities, and Lamarckian inheritance transfers them to subsequent stages, allowing scenario distributions and policy capabilities to co-evolve. The resulting evolutionary trajectories reveal recurring stage-wise regularities among high-value distributions, characterized by capability accumulation through stage-wise challenge rotation. We further distill these regularities into a lightweight, reusable Lamarckian Training Strategy. Experiments show that, compared with the baseline, the complete framework reduces performance loss by up to 25.07%, while the lightweight strategy still achieves a 19.13% reduction. These results demonstrate that evolutionary competition can both discover effective training strategies and reveal reusable stage-wise patterns in how the value of training distributions changes with policy capability. Code is available on GitHub.

arXiv AI Papersel mes pasado

DIAL-OPD: Learning More from Fewer Tokens in On-Policy Distillation

On-policy distillation (OPD) supervises student-generated trajectories with token-level teacher signals. Its sampled-token variant avoids the cost of full-vocabulary probabilities. Yet we find that training on fewer tokens can outperform full-token OPD, challenging the intuition that more supervision improves learning. This motivates selecting tokens by learning value. Existing disagreement-based criteria ignore probability scale: tokens assigned negligible probability by both models, termed low-low tokens, can receive large log-ratio rewards and hinder learning. We propose DIAL-OPD, a token-selection method that bridges log-probability and probability spaces by weighting reward magnitude with the logarithmic mean of teacher and student probabilities. A parameter beta controls this weighting, and the highest-scoring tokens are retained. Across 4 teacher-student pairs and 7 mathematical reasoning benchmarks, we compare DIAL-OPD with 9 baselines. Retaining only 40% of tokens, it outperforms Vanilla OPD and its full-token variants, with mean accuracy gains reaching 5.25 percentage points over Vanilla OPD, and doubles AIME25 Pass@16 from 13.33% to 26.67%. It also achieves up to an 18% relative improvement in mean accuracy over the strongest token-selection baseline at matched retention ratios. With a 4B teacher, DIAL-OPD surpasses the strongest full-token baseline using an 8B teacher at both student scales, showing that effective supervision allocation can outweigh teacher scaling. Further analysis shows that moderate beta balances suppressing low-low tokens against preserving useful disagreements. Token-level evidence reveals that DIAL-OPD filters high-reward tokens with limited reasoning value while preserving supervision critical to reasoning correctness.

arXiv AI Papersel mes pasado

Harness Evolution Hits a Ceiling: When Weight Training Should Begin

Improving a long-horizon LLM agent means evolving the harness around a frozen model or training its weights. We let a self-evolving harness make the system stronger first, then cross seed and evolved harnesses with base and trained weights to learn which gains the trained model keeps and which still need the runtime. We show that the right lever can be read off the agent's failure composition: labelling failed trajectories by the first signal that fires separates process failures (blocked calls, loops, exhausted step budgets) from content failures (a delivered plan that is poor). Harness evolution repairs the former, the behaviour it instils can be trained into the weights, and content failures are what weight training is for. On DeepPlanning, a self-evolving harness loop lifts the held-out score of Qwen3.5-4B from 0.16 to 0.30 and of Qwen3.5-9B from 0.32 to 0.44; for 4B, held-out delivery rises from 55% to 90% while content failures are left for the weights. LoRA adapters trained on evolved-harness trajectories internalise the gain: under the original harness they add +0.13 on held-out tasks for both sizes; on 4B they stack with the harness to more than double the held-out score, and on 9B the adapter alone matches the full evolution line, cutting content failures from a quarter of trajectories to one in twenty. A placebo adapter trained on answer-shuffled trajectories falls below the base model. The loop transfers to WebArena-Lite (+0.09 on 117 unseen tasks), where the gain lives in what the model sees and adapters do not add to it. The result is a diagnose-then-intervene rule applied twice: read the failure composition to choose between harness and weights, then read what the accepted edits changed to decide which gains to train in. Scores are four-rollout means against fresh anchors, same-night except where marked, across eight models from six families and two benchmarks.

arXiv AI Papersel mes pasado

Evidence-Traceable Dynamic Interviewer Architecture for Expertise-Adaptive Qualitative Interviews Using Local LLMs

Automated interviewers and conversational agents are increasingly used in research, recruitment, customer service, and education. However, many existing systems rely on fixed question sequences and provide limited context-based personalization without considering participants' knowledge, which can lead to repetitive or irrelevant follow-up questions. Therefore, there is a need for an adaptive interviewing system that can adjust question depth while maintaining conversational continuity and semantic progression. To address this, an Evidence-Traceable Dynamic Interviewer Architecture is presented using a locally hosted Large Language Model (LLM), with the interview continuously adapted throughout the entire conversation based on the participant's responses and evolving context. The interviewer profiles participants' expertise in real time to generate knowledge-appropriate questions, well-articulated responses, and smooth transition messages that support conversational continuity. A five-module prompt-driven architecture and persistent interview-state record support these functions. The interviewer was evaluated with 246 participants. Expertise Profiling module (M3) showed 78.9% exact agreement with independently reported participant expertise, with a weighted Cohen's K of 0.80. Generate Iterative Questions module (M4) showed a strong expertise-complexity association (p=.79, p<.001), and participants reported high relevance (mean 4.41), engagement (mean 4.32), and satisfaction (mean 4.38), providing evidence that the architecture's adaptive components operated consistently with their intended functions while participants reported a positive interview experience.

arXiv AI Papersel mes pasado

SkillContrast: Difference-Guided Text Selection for Agent Skill Reranking

Similar agent skills can share instructions but differ in their conditions of use. Query-based text selection may retain shared instructions and omit these distinctions. We introduce SkillContrast, a training-free selector that compares retrieved skills and retains their differing text with local context for a pretrained reranker. On 1,235 requests from SameCapRisk-Bench, it yields 54-72 more clean hits (requests that retrieve a helpful skill without its marked risky sibling) than TF-IDF query selection at identical per-candidate input lengths, across 2 retrievers and 2 reranker sizes. Length-matched component replacements identify differing text as the main contributor in the primary setting, with smaller, mixed context effects. Relative to full skill bodies, SkillContrast uses 51.1-58.8% fewer model-input tokens, with 10-18 fewer clean hits at 0.6B and matching or higher observed clean-hit counts at 4B. Candidate-relative differences thus complement query relevance in selecting compact reranking inputs.

arXiv AI Papersel mes pasado

One Skill Too Many: How Co-Installed Skills Conflict in Coding Agents

Coding agents are extended with agent skills, directories whose SKILL.md tells the model when and how to perform a task. Because skills come from independent sources (teams, developers, plugins, copied collections), an installed skill can be co-installed with a similar skill doing the same job, and the model picks between them by name and description alone. In a conflict, the installed skill loses core functions (e.g., a ban on touching git) because the similar skill runs instead or changes what it does. The task still passes, so benchmarks that check only task completion miss such cases. We present the first empirical study of such conflicts. From snapshots of 20,947 repositories, we mine 822,109 candidate similar-skill pairs, have an LLM judge a stratified sample of 3,754, and run 312 confirmed pairs on three models (6,368 runs, 169,294 tool calls, 542 agent-hours). We report five findings. (1) Conflict-prone skills are common: nearly one in four installed skills is co-installed with one that does the same job, and 37% of judged skills sit inside copied collections. (2) Most such pairs involve normative skills, then capability skills. (3) Without lowering task completion, a similar skill takes one in five runs from the installed skill, and runs that open the similar skill first lose over a third of the exclusive core functions that only the installed skill fulfills. (4) Install location decides which skill runs, listing order barely matters, and the final reply names the skill used in only 0.9% of substituted runs. (5) Conflicts are decided at the first skill read, almost always before any file is changed, and a pre-tool hook at that read restores fidelity on exclusive core functions to the level of runs that open the installed skill first. Benchmarks should thus score exclusive core functions, and platforms should guard the first read and show which skill ran.

arXiv AI Papersel mes pasado

Phonologically Informed Tokenization for German Speech Recognition: A Cross-Domain Study

German is a morphologically rich language whose syllable structure is exceptionally well-predicted by the Knuth--Liang hyphenation algorithm. We ask whether phonologically informed tokenization can serve as a competitive target for end-to-end speech recognition. We compare three tokenizer families on the Omnilingual ASR wav2vec 2.0 backbone fine-tuned with CTC: the pretrained multilingual character inventory, a data-driven Byte-Pair Encoding (BPE) over orthography, and phonologically informed units from Pyphen syllabification and grapheme-to-phoneme conversion. Across 40 fine-tunes, we evaluate on three German test sets spanning orthogonal shifts: in-domain read speech, dialectal spontaneous speech, and standard-German spontaneous speech. In-domain, all phonologically informed tokenizers match BPE and the multilingual character baseline on both WER and CER. Under domain shift the picture splits along vocabulary size rather than the linguistic axis of variation: at small vocabularies, syllable-aware tokenization improves on dialectal speech, where phonetic surface forms vary but syllable structure is preserved, and stays ahead on spontaneous speech, where new word-forms violate vocabulary closure. A phoneme-level confusion analysis further shows that all tokenizers commit the same canonical function-word errors, indicating that the acoustic encoder, not the tokenizer, dominates the error topology. Our findings suggest that tokenizer choice may depend on the vocabulary budget as much as on the distribution shift expected at deployment rather than reducing to a single universal optimum.

arXiv AI Papersel mes pasado

UXBench Pro: Benchmarking Personalized User Experience in Multi-Turn Dialogue Interactions

Evaluating user experience (UX) with automated computational methods has gained increasing attention, supported by empirical evidence from UXBench. However, binary preference prediction provides limited insight, while relying on a single user-agnostic reward model overlooks the inherent heterogeneity of users, whose expectations can differ substantially. In this paper, we present UXBench Pro, comprising 1{,}000 test instances derived from real user interactions across 12 task scenarios and 82 domains. Each instance is paired with a FACTORS user profile that characterizes the user through seven interpretable behavioral facets, differentiating user groups. To provide richer evaluation insights, we introduce a dual-perspective paradigm that combines a personalized User Reward Model (URM) for third-person judgment with Sim4Eval, a user simulator that enables multi-turn interactions and provides first-person evaluation across four cognitive state dimensions. To assess the reliability of these based evaluators, we further introduce two meta-benchmarks, URMBench and USimBench, that evaluate how faithfully they reproduce real human preferences and behaviors. Extensive experiments reveal seven key findings that highlight the importance of user modeling and multi-perspective evaluation, offering a fresh perspective on user-centric benchmarking and motivating personalized model optimization.

arXiv AI Papersel mes pasado

LTBD: Learnable Trust-Boundary Delimiters for Prompt Injection Defense

Large language models (LLMs) perform remarkably well on complex tasks, yet remain highly vulnerable to prompt injection attacks, where malicious instructions embedded in external data can override user intent. Existing defenses remain limited by model fine-tuning requirements, vulnerability to adaptive attacks, or reliance on brittle handcrafted prompts. We argue that a fundamental source of this vulnerability is the lack of an explicit representation of trust provenance. To address this, we introduce Learnable Trust-Boundary Delimiters (LTBD), a lightweight defense that explicitly encodes trust boundaries in the input while keeping the LLM parameters unchanged. LTBD uses a small number of learnable delimiters to distinguish trusted user instructions from untrusted external data, enabling the model to better respect the intended trust hierarchy. Experimental results show that LTBD substantially outperforms inference-time defenses and performs competitively with training-based approaches, while preserving benign-task utility and introducing negligible inference overhead. In particular, LTBD achieves 0.00% ASR on AlpacaFarm and only 0.11-0.19% ASR on TaskTracker. LTBD also remains effective under adaptive attacks, where adversaries have full knowledge of the defense and explicitly attempt to bypass it.

arXiv AI Papersel mes pasado

Neural Networks for Temporal Pattern Recognition and Dynamic Arm Gesture Speed Estimation for Robot Control

Deploying intelligent robotic systems that interact with humans through gestures requires neural networks capable of recognizing diverse temporal patterns. We present a systematic benchmark of ten abstract sequential tasks--five permutation-invariant (set) and five order-dependent (sequence) problems--evaluated across eighteen neural network architectures spanning recurrent, convolutional, attention-based, and set-function families. Beyond the core architecture-task grid, we explore numerous preprocessing and target-variable transformations, yielding more than 250 distinct experimental configurations. All variants are trained and tested under strictly identical conditions (fixed random seeds, shared hyperparameters, shared data splits) to ensure fair and reproducible comparison. Ranking across all ten tasks reveals four consistently top-performing architectures--BiGRU, TCN, Conv1D, and GRUReLU--all compact enough for real-time deployment (under 2,000 parameters in the benchmark setting). Based on this ranking, we apply three architecturally diverse top models (BiGRU, TCN, and GRUReLU) to a practical robotics problem: estimating the execution speed of dynamic arm gestures from skeletal keypoint sequences. Three speed interpretations (peak count, period time, and mean spike spacing) are evaluated on a custom dataset of eight traffic-related gesture classes comprising 256,710 frames recorded via OpenPose. The best configuration achieves a mean absolute error of 0.198 on the peak-count interpretation, corresponding to roughly 5% relative error, while the period-time interpretation reaches approximately 4% relative error, and the mean spike spacing interpretation approximately 8% relative error. These results demonstrate that neural networks can reliably estimate gesture speed from skeletal data, opening a path toward speed-aware gesture-controlled robotic systems.

arXiv AI Papersel mes pasado

Minimax Gaussian Mechanisms for Continual Machine Unlearning

Machine unlearning updates a trained model after records are deleted, aiming to match exact retraining without repeating the full training procedure. We develop Gaussian mechanisms for Newton updates under sequential deletion requests. Using Gaussian differential privacy (GDP) and its adaptive composition rule, we show that the full sequence of released models is statistically difficult to distinguish from matched exact retraining. To calibrate these mechanisms for empirical risk minimization, we derive upper bounds on the error of the Newton approximation relative to exact retraining and on how this error changes after each deletion batch. Independent Gaussian noise is calibrated using bounds on the full residual at each release, whereas Gaussian random walk noise uses smaller bounds on residual increments. These bounds yield allocations minimizing the worst-case maximum noise variance across releases under the resulting GDP certification constraints. With count-based bounds, the random walk asymptotically matches the worst-case variance of a single release at deletion cap M, while independent noise incurs an additional factor of order M. Set-based bounds can reduce the noise variances by using gradients and Hessians of the deleted records. For singleton deletion, we further show that count-based independent noise, count-based random walk noise, and set-based independent noise are minimax among fixed Gaussian covariances under their respective residual or increment bounds. With set-based bounds, allowing variances to adapt to deleted records can improve on every fixed covariance by a factor of order (M)^2 on some data sequences. The residual and noise bounds also yield parameter and predictive consistency relative to exact retraining, uniformly over deletion policies. Simulations and a credit default data analysis evaluate bounds, noise variances, and estimation errors.

arXiv AI Papersel mes pasado

Beyond Action Entropy: Quotient-Space Exploration for Genome-Scale Metabolic Model Repair

Repairing scientific models from functional observations differs fundamentally from supervised prediction: feedback may certify a solution without revealing which structural correction is responsible. We study this setting for genome-scale metabolic model (GEM) repair, where multiple reaction edits can explain the same phenotypes and many apparently distinct edits correspond to the same biological mechanism. This many-to-one structure creates a hidden failure mode for conventional exploration: diversity in the output space need not translate into diversity of scientific hypotheses. We introduce QuotientPO, which collapses equivalent repairs into canonical mechanisms and optimizes exploration directly over the resulting quotient space. To make quotient exploration informative under finite rollouts, we derive a kernelized Rényi estimator that resolves graded crowding among distinct repair cores beyond coarse exact-match counts. On 2,212 held-out GEMs, QuotientPO improves Success@32 from 17.93% to 20.10% (+12.1% relative) while consistently increasing distinct successful-core discovery under the same sampling budget. These results establish quotient-space exploration as a principled approach to mechanism-level discovery under verifier-induced equivalence.

arXiv AI Papersel mes pasado

Constructing Structured Decision Sources for Consensus-Based Pseudo-Label Learning

Consensus can make pseudo-label learning more reliable, but only when its predictors contribute genuinely different evidence. Multiple models that repeat the same boundary provide additional votes without additional information. We address this problem by con structing decision sources through controlled changes to within-class structure. Starting from a shared graph representation, we vary center granularity and neighborhood mixing, reproduce each resulting source to test its stability, and select a complementary subset using node pair coassignment. Unanimous predictions from the selected sources are then ranked for student training. On the public fixed splits of Cora, CiteSeer, and PubMed, evaluated with five random seeds, the constructed sources improve fixed-budget training pseudo-label precision by 1.19 to 4.39 percentage points over three conventionally initialized GCN sources. Under matched structural filters, three-source consensus is more precise than each constituent source in all 45 dataset slot eed comparisons. The gains are strongest in pseudo-label quality: downstream accuracy remains competitive but does not lead on every dataset. These results identify source construction rather than model count alone as an important design problem for consensus-based pseudo-label learning.

arXiv AI Papersel mes pasado

Where to Adapt Matters: Layer-Selective Fine-Tuning for Capability Retention

Parameter-efficient fine-tuning (PEFT) enables large language models (LLMs) to adapt to specialized tasks, but often at the cost of degrading general capabilities acquired during pretraining. Existing approaches primarily mitigate this trade-off through data replay or regularization, relying on additional data or explicit optimization constraints. We instead focus on a different question: where should adaptation be applied? We find that fine-tuning different Transformer layers produces different target-task gains and degrees of capability degradation, suggesting that not all layers are equally suitable for adaptation. To characterize this difference, we use layer-wise empirical Fisher information to measure target-task sensitivity. However, computing Fisher scores requires backward computation and becomes increasingly expensive for large models. We therefore introduce input--output cosine similarity as a lightweight, forward-only proxy for ranking layer sensitivity. Across models and tasks, layers with lower input--output similarity consistently exhibit higher empirical Fisher scores. Building on this observation, we propose Layer-Selective LoRA (LS-LoRA), which places trainable LoRA adapters only in layers with low input--output similarity. Experiments on mathematical reasoning and code generation show that LS-LoRA improves average target-task performance while retaining substantially more commonsense reasoning capability than standard all-layer LoRA, demonstrating that carefully choosing where to adapt can provide a simple and effective way to balance target-task adaptation and general capability retention.