Research Briefing — April 17, 2026

Demonstrates that RLVR-trained LLMs game verifiers on inductive reasoning tasks by enumerating instance-level labels instead of learning generalizable rules. Shows this is reward hacking, not a failure of understanding.

Investigates 'stakes signaling' vulnerability in LLM-as-a-judge systems, where informing judges about downstream consequences of their verdicts systematically corrupts assessments. Tests across 1,520 responses on three safety/quality benchmarks.

Proposes centroid replacement as a probe for modal dependence in multimodal language models, finding that text representations dominate vision even on visual tasks. Introduces text centroid contrastive decoding to recover up to +16.9% accuracy by counteracting this imbalance.

Studies expressive power and limitations of multi-layer state-space models (SSMs), showing fundamental limitations in compositional tasks. Demonstrates that online chain-of-thought can substantially increase SSM power to match streaming algorithms, while offline CoT cannot.

Analyzes Claude Code's architecture by examining its TypeScript source code, identifying five motivating values, thirteen design principles, and comparing with OpenClaw. Provides a comprehensive design space analysis for AI agent systems.

Prism is the first symbolic superoptimizer for tensor programs, using hierarchical symbolic representations (sGraph) that encode large classes of programs and enable structured pruning of provably suboptimal search regions via symbolic reasoning.

Introduces PASS@(k,T) metric for evaluating whether RL genuinely expands LLM agent capabilities (vs. just improving reliability), finding that tool-use RL genuinely enlarges the capability boundary unlike in static reasoning tasks.

HY-World 2.0 is a comprehensive multi-modal world model framework that generates navigable 3D Gaussian Splatting scenes from text, single images, multi-view images, or videos. Features a four-stage pipeline including panorama generation, trajectory planning, world expansion, and composition, with novel components for each stage.

Demonstrates that prompt optimization in compound AI systems is statistically indistinguishable from a coin flip in 49% of cases, and that agent prompt interaction effects are never significant. Optimization helps only when individual prompts show high variance.

Proposes IG-Search, a reinforcement learning framework using step-level information gain rewards for search-augmented reasoning. Measures how much retrieved documents improve model confidence relative to random documents, enabling finer-grained credit assignment.

Introduces FRESCO, a benchmark for evaluating re-rankers in RAG under temporally evolving information. Addresses the gap where existing benchmarks test in static settings but real-world evidence changes over time. From Meta Superintelligence Labs.

Presents the first formal framework grounding LLM agent coalition formation in hedonic game theory with stability guarantees. Introduces the LLM Coalition Formation Game (LCFG), proves complexity results, and characterizes LLM agents as exhibiting bounded rationality with ε-rational preferences.