Reasoning Structure Matters for Safety Alignment of Reasoning Models
📌 What the paper says:AltTrain argues that harmful responses in reasoning models come partly from the structure of their reasoning. It uses a lightweight supervised fine-tuning set to alter that structure without complex reinforcement learning.
🔍 The Critique:The premise is plausible: how a model reasons can matter as much as what it finally says. But 1,000 examples is a thin bridge to general safety. If the training set encodes a narrow safety pattern, the model may learn a template rather than a principle. The paper also needs to show that altered structure does not quietly reduce useful reasoning flexibility.
Read full critique
Discovering Novel LLM Experts via Task-Capability Coevolution
📌 What the paper says:AC/DC coevolves language models and synthetic tasks. The authors argue that evolving model populations and task archives can discover diverse specialist capabilities without explicit benchmark optimisation.
🔍 The Critique:Coevolution is exciting because it can generate novelty, but it is also famous for producing weird local arms races. If models generate tasks and tasks select models, the loop can drift toward quirks that look like expertise inside the ecosystem. The paper needs to prove that the specialists are not just adapted to the synthetic ecology they helped create.
Read full critique
Learning to Self-Evolve
📌 What the paper says:The paper trains a model to improve its own context at test time. A policy observes performance feedback, edits the context, and uses a tree-guided loop to search for better future behaviour.
🔍 The Critique:The central move is clever, but the phrase self-evolve risks overstating it. The system is not changing its weights during deployment; it is changing the prompt/context that future attempts see. That can be powerful, but it is closer to learned prompt repair than organism-like adaptation. The evaluation needs to separate context-search advantage from genuine…
Read full critique
AgentFactory: A Self-Evolving Framework Through Executable Subagent Accumulation and Reuse
📌 What the paper says:Storing successful task solutions as executable Python subagents — rather than text reflections — lets AI systems accumulate and reuse skills, dramatically reducing the effort needed to solve future tasks.
🔍 The Critique:The headline metric is average output tokens per task. This measures how much the orchestrating model has to think, not whether it gets the right answer. The authors explicitly note all 30 tasks completed "without runtime errors" — but runtime-error-free ≠ correct output. A subagent that generates a plausible-but-wrong chart passes this bar with flying…
Read full critique
TDAD: Test-Driven Agentic Development – Reducing Code Regressions in AI Coding Agents via Graph-Based Impact Analysis
📌 What the paper says:Applying test-driven development principles to AI coding agents — combined with graph-based impact analysis to identify which tests to run after each change — reduces regressions in agent-generated code.
🔍 The Critique:The impact analysis graph works when the dependency structure is static and explicit — typed languages with clear import graphs. Python (the language AI coding agents overwhelmingly generate) has dynamic imports, runtime monkey-patching, and metaprogramming that make static dependency graphs unreliable. A change in a dynamically loaded module won't show up…
Read full critique
Towards Safer Large Reasoning Models by Promoting Safety Decision-Making before Chain-of-Thought Generation
📌 What the paper says:By inserting a safety decision gate before chain-of-thought reasoning begins, reasoning models can be made safer — preventing the CoT process itself from being used to rationalise harmful outputs.
🔍 The Critique:The architecture is intuitive but creates a structural problem: it separates the safety decision from the reasoning context in which harm actually emerges. A model asked "how do I safely dispose of household chemicals?" produces wildly different CoT depending on intent — the pre-reasoning safety gate can't see that context yet. This risks two failure modes…
Read full critique