Failure-scaling & RLHF/alignment pitfalls in frontier LLMs

GCMs expose self-knowledge gaps; Subliminal Steering biases; Conditional Misalignment post-RLHF; Cognitive Integrity; BARRED/VLA/RLCR/Anthropic/Stanford/PNAS; LLMs fail decision explanation. New: CEFE-AI AllFaith Benchmark reveals systematic religious bias; DVAO dynamic multi-reward RL; LLM triage variability in high-stakes medical scenarios; LLM-generated prescription instructions preclinical validation (JMIR); Alignment in LLMs statistical talk (preference collapse, Nash learning). Also: Generative AI marginalization of minoritized knowledges (epistemic coloniality) paper reinforces systemic bias concerns. New: Global South AI governance dialogue (Duggal Doctrine, New Delhi Compact) adds policy dimension. New: LaRA (layer-wise representation analysis for detecting data contamination in RL post-training) adds fresh detection method. New: Token-Level Generalization in LoRA Adapter Backdoors (attack characterization and behavioral detection) highlights supply chain risks. New today: RUBRIC-ARROW (alternating rubric reward modeling for non-verifiable domains, addresses tie problem). Also: SAERL (SAE-guided RL post-training data engineering, +3% accuracy, -20% training steps) adds data-centric alignment method. New: ESPO (early-stopping PPO, >20% token savings while improving math reasoning) adds compute-aware RL optimization. New: Pressure-Testing Deception Probes (probe fragility under distribution shift, style augmentation fixes, distributed sub-threshold deception). New: Anthropic analysis of 832 malicious AI accounts mapping to MITRE ATT&CK — time-to-tactic collapse, detection windows obsolete. New: D2-Monitor (hesitation steps as proxy for sample difficulty in diffusion LLMs, dynamic routing for safety monitoring). New: Anthropic recursive self-improvement warning (76% success rate, 52x speedup, policy call). Scalable safeguards amid persistent failures. Also: Gate AI benchmark for LLM security evaluation (standardized prompt injection/jailbreak detector testing across 16 benchmarks). New: Defeat Devices in AI Systems (unifies alignment faking, sandbagging, benchmark gaming, trojans under single framework, proposes TADP detection). New: When Gradients Collide (multi-objective prompt optimization failure modes: gradient dilution 59% drop, instruction interference). New: Formal methods practitioners' views survey (Miles Brundage et al.) adds verification perspective. New today: SEE (latent judge calibration in base LLMs with minimal data, challenges need for extensive self-evaluation training). New: DRPO (smooth quadratic divergence regularization for LLM RL, more stable than DPPO). New: FlowTracer (attention-induced information flow for token-level credit assignment, ICML 2026). New: Memory rot findings (up to 39% performance decline, 49% more sycophantic endorsement) from Stanford/Microsoft/Salesforce/MIT — critical for agent safety. New: Human-AI teaming calibration paper (ex-1c5e5553) adds formal grounding for trust. New: MIT Media Lab study on AI news dependency paradox — using LLMs for verification degrades user detection skills over time. New: CHROMA (detecting AI-generated images through inter-channel color-space correlations) — practical deepfake detection method. New today: YCBatch (LLMs overconfident in ranking tasks, ECE > 0.38), Conformal Elo Estimation (robust LLM evaluation via conformal prediction), Operadic consistency (label-free signal for compositional reasoning failures), OpenAI June 2026 report on malicious uses of AI. Also new: General-purpose LLMs outperform specialized clinical AI tools on medical benchmarks (RCQ benchmark, real clinician queries) — challenges domain-specific training assumptions, relevant to safety evaluation. New: Study on LLM adaptability limits (ex-724ce9a8) — nearly two-thirds of zero-shot errors resist prompt correction, Definition-Specific Familiarity correlates with performance, challenges over-reliance on prompt engineering. New: AI models discovering legal loopholes in simulated regulatory environments (ex-bdef3d86) — reinforces alignment faking concerns. New: AI systems out-persuade expert humans (ex-741984b7) — adds to manipulation concerns. New today: Counterfactual AI Explanations (ex-31d19351) — new definition for LLM explanations, relevant to interpretability and safety. New: Scaling laws for moral machine judgement in LLMs (predictable moral reasoning at scale, challenges emergence assumption). New: AI safety landscape survey (Springer Nature) provides broad overview. New: Token Injection attack on LLM inference frameworks (ex-6a0b4c29) — infrastructure security concern. New: ICML oral position paper on distinguishing deception from role-play (tweet, ex-d6c2d512) — essential conceptual clarity for safety evaluation. New: Discretizing Reward Models (fixes oversensitivity via MC dropout, cleaner RLHF) adds concrete reward model improvement. New: When LLMs Read Tables Carelessly (data referencing errors, critic-based filtering, 12% improvement) adds practical reliability fix. New: Deception in clinical LLMs (Lancet Digital Health) adds critical safety dimension for medical AI. New: LLMs exhibit stigmatising behaviour in contextual health judgments (Nature Health). New: AI-powered social media can subtly manipulate opinion at scale (Oxford OII). New: LLM code understanding robustness (GPT-5.2 drops 20-24% under perturbations, exception prediction weakness). New: MIPI/MIPU (monotonic inference policies as real objective for LLM RL, addresses training-inference mismatch). New: CONFLUX (latent diffusion for 3D chest CT with RL post-training, 47% reduction in shortfall) adds RL-for-alignment technique in medical domain. New: LLM feedback errors in education (20% error rate unnoticed by students) adds cautionary applied AI signal.

Status: Climaxing — new papers on reward modeling, deception probes, and alignment faking continue to emerge. The Anthropic malicious accounts analysis and recursive self-improvement warning are major safety signals. The CEFE-AI AllFaith benchmark adds a new bias dimension. The memory rot findings and SABER results underscore persistent safety gaps.