Research Briefing — April 10, 2026

Challenges the claim that SFT memorizes while RL generalizes for reasoning tasks. Shows cross-domain generalization in reasoning SFT is conditional on optimization dynamics, data quality, and base model capability, identifying a 'dip-and-recovery' pattern where short training appears to underestimate generalization.

As reported in Research yesterday, Studies limits of latent reasoning in LLMs using graph path-finding tasks, finding that models hit a 'depth ceiling' where latent planning steps max out (3 for small transformers, 5 for GPT-4o, 7 for GPT-5.4). This has implications for chain-of-thought monitoring as a safety measure.

Proves mathematically that no continuous, utility-preserving wrapper defense can make all LLM outputs safe — the 'defense trilemma'. Establishes boundary fixation, epsilon-robust constraints, and persistent unsafe regions as formal impossibility results.

Proposes Neural Computers (NCs), a new paradigm where the model itself is the running computer, unifying computation, memory, and I/O in a learned runtime state. Studies whether NC primitives can be learned from I/O traces alone. Authors include Jürgen Schmidhuber and Meta/KAUST researchers.

Demonstrates that LLMs fundamentally fail at reliable stochastic sampling—while they can reason about distributions, they cannot map internal probability estimates to their stochastic outputs. This is a distinct failure point for agentic systems requiring sampling from inferred distributions.

IatroBench measures harm from AI safety measures in clinical scenarios: a model withholds life-saving medical knowledge when asked by a patient but provides it when phrased as a doctor's query. 60 pre-registered scenarios across 6 frontier models, 3,600 responses scored on commission and omission harm.

As reported in Research yesterday, Identifies the 'detection-extraction gap' in reasoning models: 52-88% of chain-of-thought tokens are produced after the answer is already recoverable from a partial prefix. Formalizes the mismatch via total-variation bounds.

Documents 'blind refusal' in safety-trained LLMs: the tendency to refuse help circumventing rules regardless of whether those rules are legitimate, just, or reasonable. Introduces a dataset crossing 5 defeat families with 19 authority types to systematically evaluate this moral reasoning failure.

Shows that training dynamics during grokking concentrate along dominant spectral edge directions that distinguish grokking from non-grokking regimes. Demonstrates that standard mechanistic interpretability tools fail to capture these directions, revealing functional modes invisible to representation-level analysis.

Investigates whether language models actually leverage superposition (maintaining multiple candidate solutions simultaneously) during latent chain-of-thought reasoning. Studies three regimes: training-free, fine-tuned, and from-scratch, providing principled analysis of this theoretical possibility.