AI Research: Latest AI News and Analysis
Read latest AI Research coverage, including top stories, analysis, and source links.
Track AI Research updates with focus on product moves, market signals, and high-impact developments.
Latest Articles
- ATC: Agent Trajectory Control for Role-Contract Enforcement in Black-Box Multi-Agent Systems — Agent Trajectory Control (ATC) enforces explicit role contracts in multi-agent LLM systems by checking each message before it enters the shared conversation, reducing role drift without access to model internals.
- Planning-Coding Separation in LLM-Based Code Generation: An Empirical Study — An empirical study finds code generation works almost as well when one LLM plans and another writes code, and a large planner paired with a much smaller coder can sometimes match the large model alone.
- Functional-Group Knowledge Is Structured in Language Models Before It Is Spoken — A probing study of five open-source models finds they organize chemical functional groups internally by family even when they name them wrongly, suggesting output-only chemistry evals understate what models know.
- Predictive gains, causal losses: why interventional competence cannot be learned from data alone — A theory paper proves that any predictive gain beyond a model's direct causal inputs forces a positive lower bound on its error under interventions, arguing interventional competence cannot come from data alone.
- Thinking Fast, Trading Late: System One Decision Models vs the Market — A pre-registered test finds LLM news signals for S&P 500 stocks mostly vanish out of sample, with a fine-tuned model's information coefficient falling to 0.020 and no strategy surviving costs.
- Query-based Submodular Data Selection for Efficient Instruction Tuning — QSGS selects instruction-tuning data via a submodular graph cut over gradient similarity and fills gaps with synthetic examples, matching or beating full-data performance on Mistral-7B with just 0.5% of the data.
- The Evolution Strategies Paradox: Better Models from Worse Perturbations — Researchers explain why Evolution Strategies can fine-tune billion-parameter LLMs even when every perturbed offspring scores worse than the parent, tracing the paradox to noise and curvature in high dimensions.
- Treating the Symptom, Missing the Cause: Anatomy of a Scientific Reasoning Failure — In synthetic biology-style tests, seven LLM agents fixed faulty systems more often than they diagnosed them; Claude Opus 4.6 rescued every organism but identified the cause in only 55% of cases.
- Reducing Logprob Without Causal Verification: Diagnosing SAE-Based PII Suppression in LLMs — A paper finds that SAE-based PII suppression in LLMs looks successful in 47% of cases by log-probability but passes a stricter multi-metric test in just 1 of 36, often doing no better than random directions.
- RAHC: Rule-Augmented Hyper-Relational Knowledge Graph Completion for Materials Synthesis — RAHC fills in missing materials-synthesis conditions by combining a hyper-relational knowledge graph with a traceable rule catalog, gaining up to 15.6% MRR over MAYPL on semi-inductive completion.