The Reasoning Gymnasium: How Zero-Sum Games Shape Smarter LLMs
SPIRAL shows that competitive self-play can train transferable reasoning in LLMs, but the real lesson is adaptive pressure, not games for their own sake.
Cognaptus Insights
Industry thoughts, trends, and commentary.
Words, Not Just Answers: Using Psycholinguistics to Test LLM Alignment
A new psycholinguistic benchmark shows where LLMs align with human word judgements—and where sensory meaning still leaks through the floorboards.
Good AI Goes Rogue: Why Intelligent Disobedience May Be the Key to Trustworthy Teammates
A mechanism-first reading of why trustworthy AI teammates may need bounded, transparent ways to refuse, interrupt, escalate, or override human instructions.
Inked in the Code: Can Watermarks Save LLMs from Deepfake Dystopia?
A mechanism-first reading of BiMark, a proposed LLM watermarking framework that tries to carry provenance data without degrading generation quality.
When Text Doesn’t Help: Rethinking Multimodality in Forecasting
A practical reading of when text improves time-series forecasting, when it adds cost without accuracy, and how operators should test multimodal systems before deploying them.
Catalysts of Thought: How LLM Agents are Reinventing Chemical Process Optimization
A mechanism-first reading of how multi-agent LLM systems can infer missing process constraints, guide simulations, and reduce the setup bottleneck in chemical optimisation.
Playing with Strangers: A New Benchmark for Ad-Hoc Human-AI Teamwork
A practical reading of AH2AC2, a Hanabi benchmark that tests whether AI agents can coordinate with human-like partners rather than merely perform well in isolation.
Mind Games for Machines: How Decrypto Reveals the Hidden Gaps in AI Reasoning
A mechanism-first look at Decrypto, a benchmark showing why strong single-agent reasoning does not automatically translate into social reasoning, coordination, or theory of mind.
Unsafe at Any Bit: Patching the Safety Gaps in Quantized LLMs
Quantization can make LLMs cheaper to deploy while quietly weakening safety, turning model compression into a governance checkpoint rather than an engineering afterthought.
Anchored Thinking: Mapping the Inner Compass of Reasoning LLMs
A mechanism-first analysis of how sentence-level thought anchors reveal which parts of a reasoning trace actually steer an LLM’s answer.