Selected work

Representative Papers

Denoising Models Develop Human-Like Perceptual Illusion Representations Across Architectures

Jul 19, 2026

This study investigates whether denoising models genuinely encode human visual illusions and their underlying mechanisms within internal representations. By analyzing internal activations across multiple architectures—combined with feature visualization, channel ablation, psychophysical modeling (FLODOG), and parametric illusion-strength experiments—the work uncovers, for the first time, a perception-like “phantom” representation that is decoupled from model output. Specifically, certain channels in intermediate layers exhibit high sensitivity to brightness illusions: their activation magnitudes correlate strongly with human perceptual judgments (Spearman ρ ≥ 0.70) and vary monotonically with illusion strength, yet they exert no influence on the final reconstructed pixels. These findings provide causal evidence for human-like perceptual mechanisms embedded within deep neural networks.

0 citationsRead paper

Frontier Coding Agents Use Metaprogramming to Adapt to Unfamiliar Programming Languages

Jun 09, 2026

This study investigates the capability of state-of-the-art large language model agents to handle esoteric programming languages—such as Brainfuck and Befunge-98—where their performance remains unclear despite strong results in mainstream languages. The authors introduce a systematic evaluation pipeline encompassing file editing, local execution, and hidden testing to assess multiple leading agents. Findings reveal that top-performing models, including Claude Opus 4.6 and GPT-5.4 xhigh, predominantly rely on metaprogramming strategies—specifically, generating target-language code via intermediate Python scripts—rather than directly writing in unfamiliar languages; disabling this approach leads to a marked performance drop. Moreover, distilling these auxiliary programs into weaker models (e.g., Sonnet 4.6 and GPT-5.4 mini) substantially enhances their effectiveness, highlighting resource orchestration as a critical factor underlying performance disparities among agents.

0 citationsRead paper

Discovering Reinforcement Learning Interfaces with Large Language Models

May 05, 2026

This work addresses the heavy reliance on manual design in defining environment interfaces—specifically observation mappings and reward functions—in reinforcement learning, for which automated solutions are largely absent. The authors propose LIMEN, a framework that achieves, for the first time, the joint automatic discovery of both observation and reward functions. LIMEN leverages a large language model to guide an evolutionary algorithm that generates executable programs from raw simulation states, iteratively refining the entire interface using policy training feedback. Experiments demonstrate that, given only trajectory-level success signals, LIMEN successfully discovers effective interfaces in both discrete grid-world and continuous control tasks. In contrast, optimizing either component in isolation fails completely in at least one domain, underscoring the necessity and superiority of co-design and substantially reducing the engineering cost of interface specification.

0 citationsRead paper

See, Symbolize, Act: Grounding VLMs with Spatial Representations for Better Gameplay

Mar 12, 2026

This work addresses the challenge faced by vision-language models (VLMs) in translating perceptual inputs into executable actions within interactive environments. The authors propose a method that integrates raw visual frames with symbolic scene representations and present the first systematic evaluation of how symbolic information influences VLM-based action generation. Multimodal policy experiments are conducted across Atari, VizDoom, and AI2-THOR platforms. Results demonstrate that high-quality symbolic representations substantially enhance VLMs’ decision-making performance in gameplay. However, symbols extracted autonomously by the model are often compromised by its inherent limitations and environmental complexity, and noisy or inaccurate symbols can severely degrade action efficacy. The study identifies the reliability of symbol extraction as a critical bottleneck for achieving effective symbol grounding in embodied interactive tasks.

0 citationsRead paper

EsoLang-Bench: Evaluating Genuine Reasoning in Large Language Models via Esoteric Programming Languages

Mar 10, 2026

This work addresses the limitation of existing code generation benchmarks, which often conflate genuine reasoning capabilities of large language models with memorization of training data. To this end, the authors propose the first evaluation benchmark based on five esoteric programming languages—such as Brainfuck and Whitespace—that are unlikely to appear in standard training corpora. By incorporating documentation-based learning, interpreter feedback, and iterative experimentation, the benchmark emulates human-like learning processes and effectively mitigates data contamination, thereby isolating transferable reasoning abilities. Experimental results reveal a stark performance gap: while state-of-the-art models achieve high scores (85–95%) on conventional benchmarks, they attain only 0–11% on this new benchmark and consistently fail to solve tasks of moderate or higher difficulty. This highlights a critical deficiency in their generalization and reasoning capacity, offering a novel paradigm for evaluating true model intelligence.

0 citationsRead paper

Recent publications

Latest Papers

Denoising Models Develop Human-Like Perceptual Illusion Representations Across Architectures

Jul 19, 2026

This study investigates whether denoising models genuinely encode human visual illusions and their underlying mechanisms within internal representations. By analyzing internal activations across multiple architectures—combined with feature visualization, channel ablation, psychophysical modeling (FLODOG), and parametric illusion-strength experiments—the work uncovers, for the first time, a perception-like “phantom” representation that is decoupled from model output. Specifically, certain channels in intermediate layers exhibit high sensitivity to brightness illusions: their activation magnitudes correlate strongly with human perceptual judgments (Spearman ρ ≥ 0.70) and vary monotonically with illusion strength, yet they exert no influence on the final reconstructed pixels. These findings provide causal evidence for human-like perceptual mechanisms embedded within deep neural networks.

0 citationsRead paper

Frontier Coding Agents Use Metaprogramming to Adapt to Unfamiliar Programming Languages

Jun 09, 2026

This study investigates the capability of state-of-the-art large language model agents to handle esoteric programming languages—such as Brainfuck and Befunge-98—where their performance remains unclear despite strong results in mainstream languages. The authors introduce a systematic evaluation pipeline encompassing file editing, local execution, and hidden testing to assess multiple leading agents. Findings reveal that top-performing models, including Claude Opus 4.6 and GPT-5.4 xhigh, predominantly rely on metaprogramming strategies—specifically, generating target-language code via intermediate Python scripts—rather than directly writing in unfamiliar languages; disabling this approach leads to a marked performance drop. Moreover, distilling these auxiliary programs into weaker models (e.g., Sonnet 4.6 and GPT-5.4 mini) substantially enhances their effectiveness, highlighting resource orchestration as a critical factor underlying performance disparities among agents.

0 citationsRead paper

Discovering Reinforcement Learning Interfaces with Large Language Models

May 05, 2026

This work addresses the heavy reliance on manual design in defining environment interfaces—specifically observation mappings and reward functions—in reinforcement learning, for which automated solutions are largely absent. The authors propose LIMEN, a framework that achieves, for the first time, the joint automatic discovery of both observation and reward functions. LIMEN leverages a large language model to guide an evolutionary algorithm that generates executable programs from raw simulation states, iteratively refining the entire interface using policy training feedback. Experiments demonstrate that, given only trajectory-level success signals, LIMEN successfully discovers effective interfaces in both discrete grid-world and continuous control tasks. In contrast, optimizing either component in isolation fails completely in at least one domain, underscoring the necessity and superiority of co-design and substantially reducing the engineering cost of interface specification.

0 citationsRead paper

See, Symbolize, Act: Grounding VLMs with Spatial Representations for Better Gameplay

Mar 12, 2026

This work addresses the challenge faced by vision-language models (VLMs) in translating perceptual inputs into executable actions within interactive environments. The authors propose a method that integrates raw visual frames with symbolic scene representations and present the first systematic evaluation of how symbolic information influences VLM-based action generation. Multimodal policy experiments are conducted across Atari, VizDoom, and AI2-THOR platforms. Results demonstrate that high-quality symbolic representations substantially enhance VLMs’ decision-making performance in gameplay. However, symbols extracted autonomously by the model are often compromised by its inherent limitations and environmental complexity, and noisy or inaccurate symbols can severely degrade action efficacy. The study identifies the reliability of symbol extraction as a critical bottleneck for achieving effective symbol grounding in embodied interactive tasks.

0 citationsRead paper

EsoLang-Bench: Evaluating Genuine Reasoning in Large Language Models via Esoteric Programming Languages

Mar 10, 2026

This work addresses the limitation of existing code generation benchmarks, which often conflate genuine reasoning capabilities of large language models with memorization of training data. To this end, the authors propose the first evaluation benchmark based on five esoteric programming languages—such as Brainfuck and Whitespace—that are unlikely to appear in standard training corpora. By incorporating documentation-based learning, interpreter feedback, and iterative experimentation, the benchmark emulates human-like learning processes and effectively mitigates data contamination, thereby isolating transferable reasoning abilities. Experimental results reveal a stark performance gap: while state-of-the-art models achieve high scores (85–95%) on conventional benchmarks, they attain only 0–11% on this new benchmark and consistently fail to solve tasks of moderate or higher difficulty. This highlights a critical deficiency in their generalization and reasoning capacity, offering a novel paradigm for evaluating true model intelligence.

0 citationsRead paper