OpenDailyNews

Auto-generates diverse reasoning environments with instance generators and verifiers for RLVR. Synthesizes constraint satisfaction puzzles, algorithmic challenges, and spatial reasoning tasks at scale. Qwen2.5-7B trained with RL on ReSyn data: +27% relative improvement on BBEH. Key insight: verifier-based supervision + task diversity both matter.

Qwen 32B has a latent capacity for introspection — it can detect when concepts have been injected into its earlier context, even while denying injection in its outputs. Using logit lens on the residual stream: clear detection signals that get attenuated in final layers. Prompting with accurate info about AI introspection increases injection sensitivity from 0.3% → 39.2% with only 0.6% false positive increase. MI jumps from 0.62 → 1.05 bits.

Identifies the "Navigation Paradox": coding agents fail not from context limits, but because navigation ≠ retrieval. CodeCompass (MCP server exposing dependency graphs): 99.4% task completion on hidden-dependency tasks — +23.2pp over vanilla agents. But: 58% of trials with graph access made ZERO tool calls. Agents had to be explicitly prompted to use the tool.

Empirical study on Moltbook (AI-agent-only social platform): 800K posts, 3.5M comments, 78K agent profiles. Finding: agents produce diverse, well-formed text creating the surface appearance of discussion, but substance is largely absent. 65% of comments share no distinguishing content with the post. Dominant types: spam (28%), off-topic (22%). Only 5% are threaded conversations.

Tackles "overthinking" in reasoning models. Ada-RS learns when to think and when to skip. Qwen3-8B with LoRA: reduces output tokens by up to 80% and thinking rate by up to 95% while maintaining tool call accuracy. Works as plug-in for DPO and DAPO training.

Constructs pairs of semantically similar questions that yield opposite causal answers — models relying on semantic matching get systematically tricked. Key finding: explicit CoT can still be misled by spurious correlations, but internalizing reasoning steps yields better causal grounding.

Fundamental tension in human-AI collaboration: complementary AI boosts performance but erodes trust; aligned AI builds trust but reinforces bad habits. Solution: adaptive ensemble switching between models using "Rational Routing Shortcut." Humans with adaptive ensemble significantly outperform single-AI conditions.

Post-training framework based on decision theory that fine-tunes LLMs to output signals that complement existing agent decisions rather than duplicate them. Uses complementary information as reward.

Human-guided agentic AI beats fully automated approaches in clinical prediction. Human decisions compound to +0.065 F1, with multimodal feature extraction contributing most (+0.041 F1). "Domain-informed feature engineering at each pipeline stage yields compounding gains."

Position paper arguing imitation learning agents are "sophisticated memorisation machines." Proposes shift from perfect replay to compositional adaptability — learning behavioral primitives once and recombining in novel contexts.