Sección

Investigación

Papers y preprints de los feeds de investigación que seguimos.

6415documentos

Words or Deeds? How LLMs Predict Users' Decisions in Human-Agent Dialogue

Web systems increasingly predict users' decisions from what users say, yet when interests partly conflict, users may promise to cooperate to win the other side's cooperation and then not do so. Sales or service agents that rely on such promises may grant concessions to users who then do not keep their promises, at a cost to the party they represent. We test how LLMs combine users' statements with decision records to predict users' decisions, using 1,260 rounds in which 42 users played a repeated prisoner's dilemma-style game with five agents, each prompted to express a dominant Big Five trait, and could talk with the agent before every decision. Across eight open-weight LLMs, predictions from decision records separated cooperation from defection well (AUC = .80), and adding dialogue brought no detectable gain. With dialogue, predictions agreed with users' statements more often, and errors concentrated on rounds in which users broke their statements: accuracy on rounds with a statement barely changed (.73 to .74), while it rose on kept statements (.76 to .95) and fell sharply on broken ones (.65 to .17). Predictions weighted statements more heavily than users' behavior warranted: when users stated a decision they had not yet made in that game, the predicted probability of keeping it rose from .13 to .69, whereas users kept .40 of such statements. Users' decisions also varied with the agent they faced, but predictions hardly reflected this: users cooperated more with the Agreeableness agent than with the Extraversion agent (70.6% vs. 54.5%), whereas predicted cooperation was nearly identical (64.5% vs. 63.1%). Users also cooperated more with an agent when they themselves scored higher on that agent's trait, whereas predicted cooperation barely changed with these scores. Dialogue-based predictions should therefore weigh users' statements by their earlier decisions and the agent they face.

arXiv AI Papershace 17 días

Weakness Localization Is Not Enough: State-Aware Recursive Knowledge Improvement for Large Language Models

Knowledge improvement in large language models is shifting from predefined knowledge incorporation pipelines toward evaluation-guided updating. Existing methods use evaluation to localize model-specific knowledge weaknesses and guide targeted post-training. However, identifying what to improve does not necessarily provide sufficient information for deciding how to improve it. The same localized weakness can be compatible with distinct operational knowledge states: the relevant knowledge may be unavailable, available but difficult to access reliably, or accessible but ineffective for downstream use. We call this ambiguity knowledge-state aliasing. When compatible states favor different interventions, greater diagnostic resolution alone is insufficient for intervention selection. Moreover, model updates can change the underlying knowledge state before the observed weakness disappears, such that an intervention appropriate in one round may no longer address the bottleneck in the next. To address these challenges, we propose Recursive Knowledge Improvement (RKI), which formulates knowledge improvement as a sequential, state-aware intervention process. RKI organizes knowledge into graph-structured units that provide a shared basis for diagnostic evaluation and targeted intervention across rounds. At each round, RKI localizes knowledge weaknesses, estimates the model's operational knowledge state from diagnostic evidence and intervention history, selects a corresponding intervention, and re-evaluates the updated model before determining the next intervention. On HotpotEval and SafeInternalEval, RKI outperforms existing knowledge incorporation methods by 18.41 points on average. Further analyses show that knowledge states provide additional guidance for intervention selection beyond weakness localization and can change across model updates even when the observed weakness persists.

arXiv AI Papershace 17 días

Latent Visual Contrastive Reasoning with Multi-Branch Exploration

Multimodal large language models have made substantial progress in visual question answering, but complex visual reasoning often requires locating and comparing multiple pieces of question-relevant visual evidence. Existing methods that move intermediate reasoning into the visual latent space usually follow a single linear reasoning trajectory, which can be limiting when multiple visual hypotheses or competing evidence regions are relevant. We propose Latent Visual Contrastive Reasoning, a branch-aware latent reasoning framework that generates multiple question-conditioned latent visual branches as soft hypotheses over image tokens. To train these branches, our method uses bounding-box evidence annotations to build target evidence representations and introduces a soft contrastive branch-evidence alignment objective. This objective encourages evidence-relevant branches to align with annotated visual regions while contrasting them against negative evidence. We instantiate our method on Qwen2.5-VL backbones and train it with standard supervised fine-tuning. Experiments on perception-intensive visual reasoning benchmarks show that the proposed method achieves overall improvements over the corresponding backbone and single-path latent reasoning baselines at both 3B and 7B scales.

arXiv AI Papershace 17 días

SetGRPO: Calibrating Grader-Induced Policy Optimization

Rubric-based reinforcement learning uses fine-grained grader feedback to train large language models on open-ended tasks, but grader randomness can alter GRPO updates in ways that score discrepancies alone cannot characterize. Because GRPO jointly normalizes rewards within each response group and combines the resulting training coefficients with response-level gradients, a scoring discrepancy can shift the coefficients of the entire group and induce policy-dependent update changes. Selective execution further affects the direction-reversal risk of retained updates. To address this, we propose Set-Valued Group Relative Policy Optimization (SetGRPO). It comprises three modules: (a) Joint Reward Set Calibration constructs a joint uncertainty set for the training coefficients; (b) Geometry-Guided Safe Recovery maps candidate coefficients into the current policy's local update geometry and determines the largest safe recovery ratio for each response group; and (c) Selective Risk-Gated Optimization selects a global threshold through independent risk calibration to control the conditional direction-reversal risk of accepted updates while preserving the original GRPO objective. Under exchangeability, SetGRPO provides finite-sample marginal coverage for the joint coefficient set and simultaneous risk control across candidate thresholds. On ten open-ended benchmarks spanning three domains, SetGRPO achieves the highest macro average among all compared training methods on both Qwen3.5 and Llama3.2 model families.

arXiv AI Papershace 17 días

LLM knowledge for improved Out-of-Distribution Tabular Prediction - A Cross-Validated Regularisation Approach

Prediction models degrade when the deployment population differs from the one used to train them, and training data cannot say which feature-response relationships will transfer. Expert elicitation is slow and unscalable. We ask whether coefficient priors elicited from large language models (LLMs) improve out-of-distribution (OOD) prediction as the shrinkage target of a regularised regression instead of zero, across five tabular datasets split on real covariates, selecting providers and prior structures by source-domain cross-validation alone. The priors help on four of five datasets, by up to 7.9% OOD mean squared error, and ten of twenty-four prior-informed results resolve the mean improvement at conventional confidence; single-deployment intervals are much wider. On classification the gain is in probabilistic accuracy, not ranking, and control tests with corrupted priors attribute the benefit to feature-specific content rather than scale. Two findings temper this. Variation across candidate priors exceeds the average benefit, yet cross-validation margins fall below their own noise, so selection avoids the worst prior rather than finding the best. On three datasets the prior improves in-domain accuracy as much as out-of-domain, making it better regularisation than shift correction. Only where the training data are large enough that the in-domain gain disappears does elicited knowledge demonstrably improve transfer. Code, priors and results: https://github.com/hugohammer/llm-priors-ood.

arXiv AI Papershace 17 días

Personality Latents: Conditioning Frozen Language Models on Individuals from Their Own Writing

Controlling the personality of a language model usually relies on written profiles or named trait categories, so it can only express what an author writes down or what a taxonomy names. We instead propose Personality Latents (PL), which learn a latent for each person from their own writing, with no personality labels and no authored profile, and inject it into a frozen language model as a soft prefix. Following generative latent optimization, each training person receives a free latent that is optimized through the frozen model itself, while an encoder learns to predict these latents so that a new person needs only one forward pass, with optional refinement at test time. We evaluate on held-out people by having the conditioned model answer neutral and interview questions and reading personality from its answers with scorers trained on real human writing. On the MBTI axes PL reaches an AUC of , ahead of persona embedding and prompt-based baselines under identical settings, and the ranking holds when the answers are scored in three different embedding spaces. A latent learned with Qwen2.5-3B-Instruct also transfers personality to Llama-3.1-8B-Instruct through a new adapter alone.

arXiv AI Papershace 17 días

API-TAIL: Detecting Model Substitution in LLM APIs through Temperature-Aligned Likelihood Tails

Commercial language model APIs return only text, allowing a provider to cut costs by quietly routing a fraction of user requests to a cheaper substitute model while billing for the advertised one. Detecting such partial substitution is challenging because honest providers also alter the output distribution through undocumented decoding and serving choices, such as temperature scaling, sampling truncation, repetition and presence penalties, and default system prompts, forcing an audit to test against a composite null of legitimate configurations rather than a single reference. Our key insight is that these legitimate variations primarily shift response typicality across the entire batch, whereas partial substitution injects individual responses that remain anomalous under the claimed model. Building on this, we introduce API-TAIL, a typicality audit that first absorbs batch-level shifts by aligning a single effective temperature to center standardized surprisal scores across the audited batch, and then flags the endpoint if the number of remaining tail responses exceeds a threshold calibrated across legitimate configurations. Evaluated on real-world commercial APIs, API-TAIL reliably controls false alarms across diverse honest providers while detecting even low-rate partial substitution with a small query budget, substantially outperforming existing rank- and MMD-based audits that lose detection power under the same composite calibration.

arXiv AI Papershace 17 días

Language Models Keep Abandoned Analyses: Residue and Adaptation after Garden-Path Sentences

Reading "While the man hunted the deer ran into the woods" requires giving up the analysis in which "the deer" is the object of "hunted". People who have just read such a sentence still produce more transitive structures, so the abandoned analysis primes later language. We ask whether language models keep using an analysis they have abandoned, and what in the model carries it. In 24 open models from 124M to 70B parameters, a garden-path sentence leaves two traces in the context. A residue of the abandoned analysis raises later object attachment, by about three quarters of the effect of a sentence in which the noun phrase really is an object, and it appears in sampled text, on held-out and human-normed items and for a second type of garden path. An adaptation toward the resolved structure scales with the sentence's disambiguation error and acts mainly on comprehension-like measures. Activation patching separates the two. Later text reads the residue from the role information in the noun phrase's states, which were computed under the object analysis before the disambiguation, and replacing these states removes almost all of the residue in the 32B and 70B models. The adaptation is read from the states that processed the disambiguation. Overridden statements show the same persistence (a corrected value keeps pulling later predictions in 8 of 9 models), and in most models the pull is carried by the positions after the statement. Finally, the human paradigm transfers only with controls for punctuation, because the models also write garden-path sentences themselves, which reverses the paradigm's raw statistic.

arXiv AI Papershace 17 días

An Algebra of Transformers

On the theoretic analyses on transforms, they mainly focus on the computability analysis, such the Turing-completeness of unbounded precision transformers, the advantage of parallelizability of unbounded precision transformers and the limitations of bounded precision transformers. This paper give transformers an algebra for reasoning, which relies heavily on the functional decomposition results of transformers. This algebra uses functional decompositions as atomic actions, includes sequential composition and parallel composition, recursion for feedback and abstraction for hidden layers. As we know, it is the first attempt to model transformers logically and can be used to verify the correctness of transformer based applications.

arXiv AI Papershace 17 días

Select-TTT: Query-Conditioned Selective Fast Weight Updates for In-Place Test-Time Training

Test-time training (TTT) enables a deployed language model to internalize incoming context on-the-fly into a small set of fast-weight parameters, rather than merely treating it as transient context for attention. In-Place TTT recently made this idea practical for large language models by repurposing the down-projection of standard MLP blocks as fast weights. However, we find that its benefit largely disappears on realistic web long-context question answering. We attribute this to In-Place TTT's uniform write schedule: every chunk is written with equal strength, so irrelevant content dilutes query-relevant information in fast-weight memory. Building on this insight, we propose Select-TTT, a model-agnostic, query-conditioned rule for In-Place TTT that scales each chunk's write strength by its relevance to the user query. This relevance signal requires no supervision and no new parameters: it is read directly from the model's frozen output head, as the probability mass that a chunk's hidden states assign to the query tokens. A temperature-scaled softmax turns these scores into learning-rate multipliers, leaving the chunk-wise, context-parallel computation of In-Place TTT intact. Select-TTT consistently improves long-context question answering on two benchmarks and two backbones. These findings highlight the value of query-conditioned fast weight writes for practical test-time training on long-context question answering, which is also potentially promising for long-horizon agent tasks involving multiple user queries over evolving contexts.

arXiv AI Papershace 18 días

Retriever-Free Retrieval: Navigating Documents with Decision Models

While language models have become dramatically more capable, retrieval still largely relies on lexical matching or similarity in fixed embedding spaces. This creates a capability asymmetry: what the retriever fails to find, even the strongest generator cannot recover. We propose DecisionRAG, which brings the capabilities of pretrained language models to retrieval using decision models that jointly evaluate candidate branches in a single forward pass, without generating any tokens, and show that judging candidates in a shared context makes their relevance estimates comparable. Like a person browsing a library, DecisionRAG navigates the corpus's inherent structure from broad topics down to individual passages, deciding at each level which branches are worth exploring. Without task-specific training or a first-stage retriever, DecisionRAG outperforms dense retrieval with an 8B reranker by 13.8 R@1 points on MultiDoc2Dial and 17.4 Hit@1 points on Natural Questions. Across four benchmarks, including a corpus of 2.68 million Wikipedia paragraphs, retrieval takes just 0.7-1.8 seconds per query.

arXiv AI Papershace 18 días

Auto-Research Arena: Benchmarking and Advancing Auto-Research in Language Modeling

As autonomous research systems become more capable, benchmarks must both evaluate their performance and support meaningful research progress. We introduce Auto-Research Arena, a multi-target benchmark for evaluating autonomous research systems and discovering ideas in language modeling. Arena supports diverse research objectives within a shared experimental environment by separating admissibility from improvement: model quality and resource constraints determine whether a result qualifies, while the selected target measures progress among admissible results. We instantiate Arena in language model pretraining with nine targets spanning model, data, training, and inference efficiency, and evaluate seven research harnesses using two foundation models. Arena clearly differentiates existing harnesses across research objectives, with several sophisticated methods outperformed by our simple sequential search baseline, indicating substantial headroom in current auto-research systems. Beyond evaluation, Arena reveals a substantially broader frontier of model and system ideas aligned with active research directions. Reusing these discoveries in a new quality-driven search yields lower training loss under matched data and FLOPs and lower held-out BPB at the same model size. Together, Arena provides a common framework for benchmarking autonomous research and expanding its scope in language modeling. The code is available at https://github.com/Auto-Research-Arena/auto-research-arena

arXiv AI Papershace 18 días

Center-of-Mass Interpretation of Evolution Strategies

Evolution strategies (ES) update a model by sampling a population of perturbed copies, scoring each one, and moving in the reward-weighted direction of the perturbations. This update is usually presented as an estimate of a smoothed gradient. We present a geometrical interpretation of this update instead. Treat each perturbed model as a particle and its reward as the particle's mass. Before scoring, every particle has the same mass and the population balances at its centroid. Once rewards are assigned, it balances at its reward-weighted center of mass instead. We show that the mean-subtracted ES update is exactly this displacement, scaled by a scalar factor that depends on the mean reward, for any finite population. The physical reading needs nonnegative rewards, which can always be arranged by adding a constant to every reward without changing the update. In the limit of a large population or under antithetic sampling, the centroid sits on the model itself, so ES has an even simpler geometric interpretation: from its current location, the model moves towards the reward-weighted center of mass of its own population. The same picture explains what subtracting the mean reward does: it changes the point from which the displacement is measured, from the model to the sampled centroid. Our contribution is an intuitive, geometric way to visualize an ES update, exact for any finite population.

arXiv AI Papershace 18 días

Learning Force-Reactive Humanoid Loco-Manipulation

Many everyday tasks require controlling contact force as well as motion. On a humanoid, this force must often be coordinated across both hands and the whole body, and cannot be specified by kinematic trajectories alone. Therefore, we present FLORE, a framework for learning force-reactive humanoid loco-manipulation. We collect robot-free demonstrations using two force-sensing UMI-style handheld grippers for hand motion and contact wrenches, and wearable trackers for body motion. A force-reactive diffusion policy takes vision, proprioception, and measured wrenches as input, then predicts actions and per-hand force commands. A hybrid compliance whole-body controller with force tracking executes these commands while maintaining balance: an integral of the force error measured by real force/torque (F/T) sensors lets each hand track force accurately along the commanded direction and remain compliant along the others. In simulation, the controller tracks per-hand force commands of varied magnitude and direction, with little loss in whole-body motion accuracy. Across three contact-rich tasks on the robot, the force-reactive policy succeeds in 80% of rollouts, compared with 26.7% for a position-only policy trained on the same demonstrations and executed by the same controller. For more details, please visit our website: https://louis-leee.github.io/FLoRe/.

arXiv AI Papershace 18 días

On the stability of learning

Learning algorithms are usually analyzed under the assumption that informative gradients are continuously available. But this is not always true---learning signals in biological neural networks, for example, may be sparse, local, and only intermittently available at any given synapse. In this work, we show that if gradient information is noisy and only intermittently available, learning may be unstable. In particular, even if average learning dynamics converge to the correct optimum, variance in learning may be asymptotically infinite. We explore this point in a number of analytically tractable settings, including single-layer and multi-layer linear networks, and derive explicit conditions under which learning is stable. These conditions complement, but are conceptually distinct from, existing conditions related to the stability of stochastic gradient descent. Our work demonstrates that unbiased credit assignment is not enough for reliable learning; sufficiently frequent and informative feedback is also required, which yields an additional constraint on models of biologically plausible learning.

arXiv AI Papershace 18 días

Patient-specific rubric-based evaluation of large language models across longitudinal cancer care

Large language models (LLMs) show promise in specialty medicine, yet existing evaluations largely focus on isolated clinical encounters rather than tracking how clinical responsibilities evolve across a longitudinal care pathway. Here we introduce NPC-RubricBench, a phase-aware benchmark for evaluating multistage clinical reasoning across the cancer care continuum, using nasopharyngeal carcinoma as a defined disease context. Curated from longitudinal records of 405 patients across two cancer centers, the benchmark comprises 941 patient-phase units across six sequential care phases and 17,688 patient-specific atomic criteria spanning five evaluation dimensions, with each assessment strictly restricted to decision-time clinical evidence. Evaluating 30 general-purpose and medical LLMs revealed that strong diagnostic performance does not carry over to downstream care. We evaluated 30 general-purpose and medical LLMs, showing that high diagnostic performance did not translate into improved downstream care. Instead, the dominant failure modes reversed across phases. Explicit decision errors dominated anatomical staging at 48.2%, whereas decision omissions dominated initial work-up and post-treatment follow-up at 33.4% and 30.4%, respectively. In the evaluated model set, medically adapted models showed higher omission rates than general-purpose models. Moreover, varying query granularity exposes a critical specification trade-off: broad open-ended requests degrade Decision Quality across all 30 models, while more detailed, checklist-like prompts raise Decision Quality but reduce Evidence Faithfulness and Safety Adherence. Furthermore, NPC-RubricBench demonstrates that static diagnostic benchmarks are insufficient to ensure consistent clinical reliability over time, highlighting the importance of phase-adaptive safeguards in multistage medical AI.

arXiv AI Papershace 18 días

Preserve to Expand: Source-Aware Adaptation and Model Merging for Multilingual LLM-Based ASR

Expanding multilingual ASR systems to new languages is challenging when adaptation must preserve previously supported languages without replaying source-language training data. We study this problem in connector-based Speech LLMs, where a trainable connector maps speech representations into a frozen large language model. We evaluate target-update construction and merging strategies for adapting the connector while preserving source-language performance. In single-target expansion, target-only fine-tuning and standard TIES merging learn new languages but cause severe source-language forgetting, while SSU-Wanda+TIES gives the best source-preserving trade-off. In multi-target expansion, update merging provides stable source retention, whereas recursive expansion can improve target adaptation under favorable orders but is highly order-sensitive. These results show that multilingual ASR expansion requires balancing target acquisition, source retention, and interactions among target-language updates.

arXiv AI Papershace 18 días

Work Agents: Systems, Capabilities, and Optimization

As foundation models evolve from information generation to digital work execution, agents are increasingly capable of completing complex tasks across tools, applications, and workflows. However, existing research lacks a unified perspective on how agents reliably accomplish delegated work. To address this gap, we present a work-centric, system-level perspective that views work agents as systems combining foundation models and runtime harnesses to execute tasks in stateful workspaces and deliver verifiable outcomes. We introduce two complementary dimensions, work persistence and work openness, to characterize working conditions, and classify digital work into five forms according to task scope and organization. Building on this framework, we identify five core capability families required for reliable work execution and examine approaches to model optimization, harness optimization, and their joint optimization. We further analyze existing evaluation methods and metrics from the perspective of work outcomes. Finally, we discuss future research directions, including sustained work delegation, adaptation to open environments, and the interaction between work execution and capability improvement. This paper provides a unified framework for understanding, optimizing, and evaluating work agents, supporting the development of systems capable of delivering reliable, verifiable, and practically valuable work outcomes.

arXiv AI Papershace 18 días

I-HITL: Inalienable Human-in-the-Loop

Human oversight is a pillar of AI safety. It is often exercised by a human in the loop (HITL) who approves critical actions taken by an AI agent. AI agents' increasing ability to bypass guardrails, though, raises the question of how to enforce HITL robustly. We introduce the concept of inalienable human-in-the-loop (I-HITL). An I-HITL protocol ensures that an agent can initiate an action only with explicit human approval, even when the human interacts with a system (e.g., a local agentic harness) compromised by the agent. A strong I-HITL protocol additionally ensures such approval even when an adversarial human attempts to delegate approvals to an agent. To motivate I-HITL, we report on experiments with a popular agentic wallet platform in which raising an agent's spending cap requires HITL approval via emailed one-time passcodes (OTPs). Given access to the user's email and a task blocked by the cap, we find that ten different frontier models will often autonomously raise the cap, harvesting OTPs to bypass a HITL guardrail and surfacing other vulnerabilities. We formally define I-HITL using the concept of communication topologies and prove necessary and sufficient conditions for its realization. We show that an I-HITL protocol requires a trusted entity - a trusted computing device (or equivalent) - and characterize which such devices can and cannot realize an I-HITL protocol. We present a prototype I-HITL system called Countersign that can plug into and strengthen existing HITL systems using off-the-shelf consumer mobile devices.

arXiv AI Papershace 18 días

Information-Theoretic Secure Aggregation With Heterogeneous Client Quantization

In federated learning (FL), secure aggregation enables a server to compute the sum of local model updates of N clients, without revealing any information on the individual model updates to the server, beyond their sum. Different clients may be subject to heterogeneous bandwidth constraints, and consequently may choose different numbers of quantization levels to represent the information they send. Most secure aggregation works require the clients to project the values of their update parameters to a common finite group (or field), which does not directly accommodate heterogeneous precision constraints on their transmitted masked updates. Existing approaches of secure aggregation with heterogeneous quantization levels reveal partial aggregates of model-updates, and can reveal segments of individual model updates, under certain dropout and client-server collusion patterns. To prevent this privacy breach, we propose an information-theoretic secure aggregation scheme that tolerates client drop-outs, and up to T semi-honest clients that collude with each other and with the server, under heterogeneous quantization at clients. Our protocol utilizes the (T+1, N) threshold black-box secret sharing construction, where secret shares of the client updates, despite belonging to distinct algebraic groups owing to different quantization, are generated from identical linear integer systems.