arXiv subjects

Sandro Andric

Publications and source records attributed to Sandro Andric.

5 recordsLinked to original sources

arXiv2026

Diversity Without Fidelity: A Solver-Sampler Mismatch in Multi-Agent LLM Negotiation Simulation

Language models are increasingly used to simulate people: survey respondents, negotiators, stakeholders in policy exercises. In that role a model should reproduce how people plausibly behave, hesitating, conceding late, and settling for imperfect deals, rather than playing the best move. We call this the sampler role, in contrast to the solver role of finding the best move, and we test how the reasoning modes providers ship to strengthen models as solvers affect it. Our testbed is multi-party negotiation: five agents bargain over a regulation for fifteen turns, and unresolved issues are decided by an authority. Agents without a structured memory of the negotiation almost never reach agreement, whether reasoning is on or off: 314 of 315 such runs end with the authority deciding. What reasoning changes is how the failure looks. With reasoning enabled, one model family negotiates visibly, with varied moves, concessions in most runs, and a different path every time, yet still ends without agreement in fifteen runs of fifteen. Diversity checks would pass a model whose endings never change. Two further results show the task permits agreement: when agents write their own short running notes on the negotiation, agreement becomes the norm, while the same notes supplied ready-made change nothing; and hand-coded agents following textbook concession strategies agree in most runs under identical rules. Simulation pipelines should therefore vet models as samplers, on the distributions of outcomes they produce. Fidelity as a sampler must be tested on its own: solver strength is no guide to it, and switching on reasoning leaves it where it was.

cs.LG↗

arXiv2025

Brain-Grounded Axes for Reading and Steering LLM States

Interpretability methods for large language models (LLMs) typically derive directions from textual supervision, which can lack external grounding. We propose using human brain activity not as a training signal but as a coordinate system for reading and steering LLM states. Using the SMN4Lang MEG dataset, we construct a word-level brain atlas of phase-locking value (PLV) patterns and extract latent axes via ICA. We validate axes with independent lexica and NER-based labels (POS/log-frequency used as sanity checks), then train lightweight adapters that map LLM hidden states to these brain axes without fine-tuning the LLM. Steering along the resulting brain-derived directions yields a robust lexical (frequency-linked) axis in a mid TinyLlama layer, surviving perplexity-matched controls, and a brain-vs-text probe comparison shows larger log-frequency shifts (relative to the text probe) with lower perplexity for the brain axis. A function/content axis (axis 13) shows consistent steering in TinyLlama, Qwen2-0.5B, and GPT-2, with PPL-matched text-level corroboration. Layer-4 effects in TinyLlama are large but inconsistent, so we treat them as secondary (Appendix). Axis structure is stable when the atlas is rebuilt without GPT embedding-change features or with word2vec embeddings (|r|=0.64-0.95 across matched axes), reducing circularity concerns. Exploratory fMRI anchoring suggests potential alignment for embedding change and log frequency, but effects are sensitive to hemodynamic modeling assumptions and are treated as population-level evidence only. These results support a new interface: neurophysiology-grounded axes provide interpretable and controllable handles for LLM behavior.

cs.LG↗

arXiv2025

Do Large Language Models Walk Their Talk? Measuring the Gap Between Implicit Associations, Self-Report, and Behavioral Altruism

We investigate whether Large Language Models (LLMs) exhibit altruistic tendencies, and critically, whether their implicit associations and self-reports predict actual altruistic behavior. Using a multi-method approach inspired by human social psychology, we tested 24 frontier LLMs across three paradigms: (1) an Implicit Association Test (IAT) measuring implicit altruism bias, (2) a forced binary choice task measuring behavioral altruism, and (3) a self-assessment scale measuring explicit altruism beliefs. Our key findings are: (1) All models show strong implicit pro-altruism bias (mean IAT = 0.87, p < .0001), confirming models "know" altruism is good. (2) Models behave more altruistically than chance (65.6% vs. 50%, p < .0001), but with substantial variation (48-85%). (3) Implicit associations do not predict behavior (r = .22, p = .29). (4) Most critically, models systematically overestimate their own altruism, claiming 77.5% altruism while acting at 65.6% (p < .0001, Cohen's d = 1.08). This "virtue signaling gap" affects 75% of models tested. Based on these findings, we recommend the Calibration Gap (the discrepancy between self-reported and behavioral values) as a standardized alignment metric. Well-calibrated models are more predictable and behaviorally consistent; only 12.5% of models achieve the ideal combination of high prosocial behavior and accurate self-knowledge.

cs.LG↗

arXiv2025

BlockCert: Certified Blockwise Extraction of Transformer Mechanisms

Mechanistic interpretability aspires to reverse-engineer neural networks into explicit algorithms, while model editing seeks to modify specific behaviours without retraining. Both areas are typically evaluated with informal evidence and ad-hoc experiments, with few explicit guarantees about how far an extracted or edited model can drift from the original on relevant inputs. We introduce BlockCert, a framework for certified blockwise extraction of transformer mechanisms, and outline how a lightweight extension can support certified local edits. Given a pre-trained transformer and a prompt distribution, BlockCert extracts structured surrogate implementations for residual blocks together with machine-checkable certificates that bound approximation error, record coverage metrics, and hash the underlying artifacts. We formalize a simple Lipschitz-based composition theorem in Lean 4 that lifts these local guarantees to a global deviation bound. Empirically, we apply the framework to GPT-2 small, TinyLlama-1.1B-Chat, and Llama-3.2-3B. Across these models we obtain high per-block coverage and small residual errors on the evaluated prompts, and in the TinyLlama setting we show that a fully stitched model matches the baseline perplexity within approximately 6e-5 on stress prompts. Our results suggest that blockwise extraction with explicit certificates is feasible for real transformer language models and offers a practical bridge between mechanistic interpretability and formal reasoning about model behaviour.

cs.LG↗

arXiv2025

An Exact Quantile-Energy Equality for Terminal Halfspaces in Linear-Gaussian Control with a Discrete-Time Companion, KL/Schrodinger Links, and High-Precision Validation

We prove an exact equality between the minimal quadratic control energy and the squared normal-quantile gap for terminal halfspaces in linear-Gaussian systems with additive control and quadratic effort $E(u) = \tfrac12\!\int u^\top M u\,dt$ where $M = B^\top\Sigma^{-1}B$. For terminal halfspace events, the minimal energy equals the squared normal-quantile gap divided by twice a controllability-to-noise ratio $R_T^2(w)=(w^\top W_c^M w)/(w^\top V_T w)$ and is attained by a matched-filter control. We provide an exact zero-order-hold discrete-time companion via block exponentials, relate the result to minimum-energy control, Gaussian isoperimetry, risk-sensitive/KL control, and Schrodinger bridges, and validate to high precision with Monte Carlo. We state assumptions, singular-$M$ handling, and edge cases. The statement is a compact synthesis and design-ready translator, not a universal principle. Novelty: while the ingredients (Gramians, Cauchy-Schwarz, Gaussian isoperimetry) are classical, to our knowledge the explicit quantile-energy equality with a constructive matched-filter achiever for terminal halfspaces, and its discrete-time companion, are not recorded together in the cited literature.

eess.SY↗