arXiv subjects

Hua-Dong Xiong

Publications and source records attributed to Hua-Dong Xiong.

8 recordsLinked to original sources

arXiv2026

The Convention Gap: Beyond Literal Content in Human-AI Cooperation

Cooperative AI agents are often evaluated against other AIs, yet human cooperation is thought to rely on implicit conventions: shared rules for reading meaning beyond the literal message. We propose the convention gap as a diagnostic of such implicit communication. In the card game Hanabi, where hints are the only messages about one's own hand, the convention gap is the failure probability predicted exactly from the literal content of the hints and the visible cards, minus the observed failure rate of the cards played. We prove that the population gap (the value the gap approaches over many games) is zero when the partner's choices convey nothing beyond their literal content, given a fair shuffle and a player who uses only what was said and seen. A nonzero population gap thus indicates that partner choices conveyed more; it does not measure how much. The gap was +26.2 percentage points (pp) in public human-human games (+46 pp on plays of cards that no hint had pointed to), +16.4 pp for human plays with AI partners, and -0.6 pp in AI-AI games we simulated among rule-based agents. With three AI agents of a published human-AI study, the literal failure probability of the cards humans played was similar across partners (38-41%), yet human failure ranged from 14.4% to 34.4% and the gap from +24.1 to +6.2 pp, and the contrast between partners persisted within participants. Across five levels of Off-Belief Learning agents, trained so that the first level reads past actions only for their literal content and each higher level interprets more, the gap rose monotonically from +1.6 to +21.7 pp. These results argue for reporting the convention gap on human-partnered play, not only AI-AI performance, when evaluating agents intended for human partners.

cs.AI↗

arXiv2026

In-context superposition: human-like working memory interference in large language models

Intelligent systems must maintain and manipulate task-relevant information online to adapt to dynamic environments. This capacity, known as working memory, is fundamental to human reasoning. Yet, human working memory is strikingly limited, maintaining only three to four items in a brain with billions of neurons. Surprisingly, large language models (LLMs), despite different substrates and direct access to prior context through attention, exhibit similar working memory limitations. Why should such different systems face analogous constraints? We propose that working memory limitations reflect a general trade-off of shared representations: representational compression and reuse support efficient learning and generalization, but also cause simultaneously active representations to interfere. We show a two-layer transformer trained on a working memory task can solve it perfectly, but diverse trained LLMs exhibit human-like limitations: performance declines with memory load, while retrieval is biased by recency and stimulus statistics. Mirroring humans, working memory performance in LLMs is also associated with broader model capability. Mechanistically, we show that LLMs encode multiple memories in entangled representations --- a condition we call \emph{in-context superposition} --- and progressively suppress competing content while aligning the target with the readout. Moreover, a causal intervention that suppresses interfering information improves performance. Together, these findings suggest that working memory capacity reflects the ability to select task-relevant information under interference, a computational challenge shared by biological and artificial systems.

cs.LG↗

arXiv2026

Large language models reorganize representational geometry during in-context learning

Large language models (LLMs) show remarkable flexibility in adapting to novel tasks without parameter updates, a capacity known as in-context learning (ICL). Prior work has sought to understand ICL by studying the circuits, algorithms, and representations that support it. Yet why some ICL tasks are easy to solve while others are difficult remains unresolved. In this paper, we ask whether LLMs can adapt their representations arbitrarily to solve a simple linear classification task. Specifically, we construct a family of binary classification tasks in which labels are defined by projecting LLMs' own representations onto different axes. Surprisingly, although all tasks are linearly separable by construction, their in-context learnability varies systematically across axes. We find that successful ICL is accompanied by a geometric reorganization of internal representations that increases task-relevant separability. Causal interventions that amplify neural activity along the axis defining the task are insufficient to improve behavioral performance or induce this representational reorganization. We also show that LLM behavior is best described by a prototype-like algorithm operating on representations that are themselves reorganized in context to adapt to the task. Together, these findings offer a geometric account of ICL in LLMs, showing that representations acquired through training constrain what can be exploited through in-context learning.

cs.CL↗

arXiv2026

Thinking Under Uncertainty: Evidence Use and Information-Seeking in Language Models

Inference-time thinking improves the performance of large language models, but aggregate outcomes do not reveal whether models use available evidence more effectively or seek information that could improve future decisions. We distinguish these responses by measuring action preference, thinking length, and reported confidence under matched uncertainty. Ten open-weight models completed matched horizon-style two-armed bandit trials in thinking and non-thinking modes. A cognitive model separated value-guided action and uncertainty-independent choice noise from two behavioral signatures of exploration: a UCB-like preference for the less-known arm and Thompson-like choice variability that increases with total uncertainty. On average, thinking strengthened value-guided action and reduced uncertainty-independent choice noise, without producing UCB-like exploration or strengthening Thompson-like exploration. Outside action, the information-imbalanced history condition, which also displayed more observations than the matched balanced condition, was associated with greater thinking length. Reported confidence became more sensitive to decision difficulty and more strongly associated with chosen task evidence. We interpret these thinking-length and reported-confidence patterns as consistent with metacognitive control and metacognitive monitoring, respectively, without establishing either process. Decoder sweeps, especially temperature, altered choice noise and thinking length but did not reproduce the joint cross-output pattern. In this controlled decision setting, thinking improved how models acted on current evidence, while neither measured signature supported a shift toward a more information-seeking policy.

cs.LG↗

arXiv2026

Post-training makes large language models less human-like

Large language models (LLMs) are increasingly used as surrogates for human participants, but it remains unclear which models best capture human behavior and why. To address this, we introduce Psych-201, a novel dataset that enables us to measure behavioral alignment at scale. We find that post-training -- the stage that turns base models into useful assistants -- consistently reduces alignment with human behavior across model families, sizes, and objectives. Moreover, this misalignment widens in newer model generations even as base models continue to improve. Finally, we find that persona-induction -- a popular technique for eliciting human-like behavior by conditioning models on participant-specific information -- does not improve predictions at the level of individuals. Taken together, our results suggest that the very processes that are currently employed to turn LLMs into useful assistants also make them less accurate models of human behavior.

cs.CL↗

arXiv2026

The Position Curse: LLMs Struggle to Locate the Last Few Items in a List

Modern large language models (LLMs) can find a needle in a haystack (locating a single relevant fact buried among hundreds of thousands of irrelevant tokens) with near-saturated accuracy, yet fail to retrieve the last few items in a short list. We call this failure the Position Curse. For instance, even in a two-line code snippet, Claude Opus 4.6 misidentifies the second-to-last line most of the time. To characterize this failure, we evaluated two complementary queries: given a position in a sequence (of letters or words), retrieve the corresponding item; and given an item, return its position. Each position is specified as a forward or backward offset from an anchor, either an endpoint of the list (its start or end) or another item in the list. Across both open-source and frontier closed-source models, backward retrieval substantially lags forward retrieval. To test whether this capability can be rescued by post-training, we constructed PosBench, a position-focused training dataset. LoRA fine-tuning improves both forward and backward retrieval and generalizes to a held-out code-understanding benchmark (PyIndex), yet absolute performance remains far from saturated. As LLM coding agents increasingly operate over large codebases where precise indexing becomes essential for code understanding and editing, position-based retrieval emerges as a key capability for future pretraining objectives and model design.

cs.LG↗

arXiv2026

Hypothesis generation and updating in large language models

Large language models (LLMs) increasingly help people solve problems, from debugging code to repairing machinery. This process requires generating plausible hypotheses from partial descriptions, then updating them as more information arrives. Yet how LLMs perform this form of inference, and how close it is to optimal, remains unclear. We study this question in the number game, a controlled setting in which a learner infers the hypothesis supported by a few positive integers, such as $\{16, 8, 2, 64\}$: a rule like powers of 2 or an interval like numbers near 20. We measure the posterior over hypotheses using three complementary probes: posterior prediction, hypothesis evaluation, and hypothesis generation. We then compare LLM behavior with an optimal Bayesian model and human behavior, and test whether the same posterior is expressed across probes. LLMs are often well described by a two-parameter Bayesian fit, but with systematic offsets: by default they show a strong-sampling assumption that creates an implicit Occam's razor, favoring narrower hypotheses, while thinking mode shifts them toward greater prior reliance. We also find a robust evaluation--generation gap: LLMs select more correct hypotheses during hypothesis evaluation but generate simpler, more rule-like hypotheses. Finally, this Bayesian-with-bias pattern does not extrapolate. Models can behave as if they hold rule-like hypotheses over observed examples, yet generalize poorly to parts of the hypothesis domain not covered by those examples. Our results highlight a limitation of LLMs as general problem solvers, especially for scientific inference, where hypotheses must go beyond the data.

cs.LG↗

arXiv2025

Language Models Are Capable of Metacognitive Monitoring and Control of Their Internal Activations

Large language models (LLMs) can sometimes report the strategies they actually use to solve tasks, yet at other times seem unable to recognize those strategies that govern their behavior. This suggests a limited degree of metacognition - the capacity to monitor one's own cognitive processes for subsequent reporting and self-control. Metacognition enhances LLMs' capabilities in solving complex tasks but also raises safety concerns, as models may obfuscate their internal processes to evade neural-activation-based oversight (e.g., safety detector). Given society's increased reliance on these models, it is critical that we understand their metacognitive abilities. To address this, we introduce a neuroscience-inspired neurofeedback paradigm that uses in-context learning to quantify metacognitive abilities of LLMs to report and control their activation patterns. We demonstrate that their abilities depend on several factors: the number of in-context examples provided, the semantic interpretability of the neural activation direction (to be reported/controlled), and the variance explained by that direction. These directions span a "metacognitive space" with dimensionality much lower than the model's neural space, suggesting LLMs can monitor only a small subset of their neural activations. Our paradigm provides empirical evidence to quantify metacognition in LLMs, with significant implications for AI safety (e.g., adversarial attack and defense).

cs.AI↗