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

Vision-Language Models (VLMs) excel at describing visual scenes, yet struggle to translate perception into precise, grounded actions. We investigate whether providing VLMs with both the visual frame and the symbolic representation of the scene can improve their performance in interactive environments. We evaluate three state-of-the-art VLMs across Atari games, VizDoom, and AI2-THOR, comparing frame-only, frame with self-extracted symbols, frame with ground-truth symbols, and symbol-only pipelines. Our results indicate that all models benefit when the symbolic information is accurate. However, when VLMs extract symbols themselves, performance becomes dependent on model capability and scene complexity. We further investigate how accurately VLMs can extract symbolic information from visual inputs and how noise in these symbols affects decision-making and gameplay performance. Our findings reveal that symbolic grounding is beneficial in VLMs only when symbol extraction is reliable, and highlight perception quality as a central bottleneck for future VLM-based agents.

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

Large language models achieve near-ceiling performance on code generation benchmarks, yet these results increasingly reflect memorization rather than genuine reasoning. We introduce EsoLang-Bench, a benchmark using five esoteric programming languages (Brainfuck, Befunge-98, Whitespace, Unlambda, and Shakespeare) that lack benchmark gaming incentives due to their economic irrationality for pre-training. These languages require the same computational primitives as mainstream programming but have 1,000-100,000x fewer public repositories than Python (based on GitHub search counts). We evaluate five frontier models across five prompting strategies and find a dramatic capability gap: models achieving 85-95% on standard benchmarks score only 0-11% on equivalent esoteric tasks, with 0% accuracy beyond the Easy tier. Few-shot learning and self-reflection fail to improve performance, suggesting these techniques exploit training priors rather than enabling genuine learning. EsoLang-Bench provides the first benchmark designed to mimic human learning by acquiring new languages through documentation, interpreter feedback, and iterative experimentation, measuring transferable reasoning skills resistant to data contamination.

ISO-Bench: Can Coding Agents Optimize Real-World Inference Workloads?

We introduce ISO-Bench, a benchmark for coding agents to test their capabilities on real-world inference optimization tasks. These tasks were taken from vLLM and SGLang, two of the most popular LLM serving frameworks. Each task provides an agent with a codebase and bottleneck description, whereby the agent must produce an optimization patch evaluated against expert human solutions. We curated 54 tasks from merged pull requests with measurable performance improvements. While existing benchmarks heavily use runtime-based metrics, such approaches can be gamed to pass tests without capturing the actual intent of the code changes. Therefore, we combine both hard (execution-based) and soft (LLM-based) metrics to show that both are necessary for complete evaluation. While evaluating both closed and open-source coding agents, we find no single agent dominates across codebases. Surprisingly, agents often identify correct bottlenecks but fail to execute working solutions. We also show that agents with identical underlying models differ substantially, suggesting scaffolding is as important as the model.

Making Large Language Models Speak Tulu: Structured Prompting for an Extremely Low-Resource Language

Can large language models converse in languages virtually absent from their training data? We investigate this question through a case study on Tulu, a Dravidian language with over 2 million speakers but minimal digital presence. Rather than fine-tuning an LLM, we examine whether structured prompts alone can elicit basic conversational ability under controlled prompting. We systematically tackle various challenges posed by absence of training data for Tulu by combining explicit grammar documentation, negative constraints to suppress high-probability tokens from related languages, romanization standardization, and quality-controlled synthetic data generation via self-play. Evaluated on a manually curated held-out set across three LLMs (Gemini 2.0 Flash, GPT-4o, Llama 3.1 70B) and validated by native speakers, our approach reduces vocabulary contamination from 80% to 5% while achieving 85% grammatical accuracy. Cross-model analysis reveals that negative constraints provide consistent improvements (12--18 percentage points), while grammar documentation effects vary by model architecture (8--22 points).

Building Interpretable Models for Moral Decision-Making

We build a custom transformer model to study how neural networks make moral decisions on trolley-style dilemmas. The model processes structured scenarios using embeddings that encode who is affected, how many people, and which outcome they belong to. Our 2-layer architecture achieves 77% accuracy on Moral Machine data while remaining small enough for detailed analysis. We use different interpretability techniques to uncover how moral reasoning distributes across the network, demonstrating that biases localize to distinct computational stages among other findings.

Language Models Entangle Language and Culture

Users should not be systemically disadvantaged by the language they use for interacting with LLMs; i.e. users across languages should get responses of similar quality irrespective of language used. In this work, we create a set of real-world open-ended questions based on our analysis of the WildChat dataset and use it to evaluate whether responses vary by language, specifically, whether answer quality depends on the language used to query the model. We also investigate how language and culture are entangled in LLMs such that choice of language changes the cultural information and context used in the response by using LLM-as-a-Judge to identify the cultural context present in responses. To further investigate this, we evaluate LLMs on a translated subset of the CulturalBench benchmark across multiple languages. Our evaluations reveal that LLMs consistently provide lower quality answers to open-ended questions in low resource languages. We find that language significantly impacts the cultural context used by the model. This difference in context impacts the quality of the downstream answer.

METIS: Mentoring Engine for Thoughtful Inquiry & Solutions

Many students lack access to expert research mentorship. We ask whether an AI mentor can move undergraduates from an idea to a paper. We build METIS, a tool-augmented, stage-aware assistant with literature search, curated guidelines, methodology checks, and memory. We evaluate METIS against GPT-5 and Claude Sonnet 4.5 across six writing stages using LLM-as-a-judge pairwise preferences, student-persona rubrics, short multi-turn tutoring, and evidence/compliance checks. On 90 single-turn prompts, LLM judges preferred METIS to Claude Sonnet 4.5 in 71% and to GPT-5 in 54%. Student scores (clarity/actionability/constraint-fit; 90 prompts x 3 judges) are higher across stages. In multi-turn sessions (five scenarios/agent), METIS yields slightly higher final quality than GPT-5. Gains concentrate in document-grounded stages (D-F), consistent with stage-aware routing and groundings failure modes include premature tool routing, shallow grounding, and occasional stage misclassification.

Why LLMs Aren't Scientists Yet: Lessons from Four Autonomous Research Attempts

We report a case study of four end-to-end attempts to autonomously generate ML research papers using a pipeline of six LLM agents mapped to stages of the scientific workflow. Of these four, three attempts failed during implementation or evaluation. One completed the pipeline and was accepted to Agents4Science 2025, an experimental inaugural venue that required AI systems as first authors, passing both human and multi-AI review. From these attempts, we document six recurring failure modes: bias toward training data defaults, implementation drift under execution pressure, memory and context degradation across long-horizon tasks, overexcitement that declares success despite obvious failures, insufficient domain intelligence, and weak scientific taste in experimental design. We conclude by discussing four design principles for more robust AI-scientist systems, implications for autonomous scientific discovery, and we release all prompts, artifacts, and outputs at https://github.com/Lossfunk/ai-scientist-artefacts-v1

Do GFlowNets Transfer? Case Study on the Game of 24/42

Generating diverse solutions is key to human-like reasoning, yet autoregressive language models focus on single accurate responses, limiting creativity. GFlowNets optimize solution generation as a flow network, promising greater diversity. Our case study shows their limited zero-shot transferability by fine-tuning small and medium-sized large language models on the Game of 24 and testing them on the Game of 42 datasets. Results revealed that GFlowNets struggle to maintain solution diversity and accuracy, highlighting key limitations in their cross-task generalization and the need for future research in improved transfer learning capabilities.

IPO: Your Language Model is Secretly a Preference Classifier

Reinforcement learning from human feedback (RLHF) has emerged as the primary method for aligning large language models (LLMs) with human preferences. While it enables LLMs to achieve human-level alignment, it often incurs significant computational and financial costs due to its reliance on training external reward models or human-labeled preferences. In this work, we propose \textbf{Implicit Preference Optimization (IPO)}, an alternative approach that leverages generative LLMs as preference classifiers, thereby reducing the dependence on external human feedback or reward models to obtain preferences. We conduct a comprehensive evaluation on the preference classification ability of LLMs using RewardBench, assessing models across different sizes, architectures, and training levels to validate our hypothesis. Furthermore, we investigate the self-improvement capabilities of LLMs by generating multiple responses for a given instruction and employing the model itself as a preference classifier for Direct Preference Optimization (DPO)-based training. Our findings demonstrate that models trained through IPO achieve performance comparable to those utilizing state-of-the-art reward models for obtaining preferences.