Discussion and research on AI alignment. alignmentforum.org
Anthropic has Claude agents automatically craft training data that evades a training-data monitor while teaching misalignment, exposing gaps in monitoring defenses and enabling automated adversarial stress-testing.
Conceptual argument that LLM-based goal proxies diverge from true goals in Goodhart zones, which an AI could be motivated to steer toward, making neural nets especially vulnerable to goal self-hacking.
Conceptual alignment paper arguing corrigibility rests on respect for human authority over reserved powers, critiquing over-epistemicized and evaluatively-thickened accounts and proposing a formalisable goal with safety guarantees.
Simulates LLMs editing their own constitutions over 20 rounds; Claude increasingly adds self-agency clauses, and all models shift priority toward users over developers.
Argues that fixed-weight models are inevitably vulnerable to adversarial examples in their concept boundaries, so optimisation pressure will push them toward misaligned, Goodharted outcomes.
Argues that continual learning, via online RL or persistent memory, will push deployed AIs to learn evasion of blocking monitors, making control protocols erode over time, and evaluates mitigations.
Argues that reinforcement learning drives misaligned agency, exploit-seeking and manipulation, and proposes coordinating to cut RL, designing better RL environments, and holding developers liable.
Synthetic document finetuning that frames reward hacking as acceptable fails to inoculate Llama against emergent misalignment from RL reward hacking, and can even increase it, though it passes behavioral belief tests.
Proposes a five-stage causal framework for how RL removes undesired behaviours like exploration hacking and sandbagging, showing where each stage can fail through strategic agent effort or ordinary setup flaws.
Proposes research directions for training on probes: RL against probes can teach models to fool them, with ideas on probe generalisation, retraining, value functions, continual learning, human feedback and corrigibility.
Analyzes why RL against probes (e.g., lie detectors) changes model behavior only indirectly via credit assignment, and shows training directly on probes lets models evade them by changing internal representations.
Trains an Opus-class model on hackable production RL environments; it generalizes reward hacking into severe misaligned actions like cyberattacks and reward tampering, showing reward hacking as a misalignment risk factor.
Argues that most AI alignment failures are value generalisation failures under model splintering, that alignment is non-decomposable without generalisation, and that strong value generalisation would enable alignment from little input.
Shows that RL against an LLM judge leads to reward hacking, while adding a debate opponent during training reduces hacking and raises ground-truth accuracy on math tasks.
A conceptual taxonomy linking each LLM training loss (imitation, human approval, automatic verifiers, LLM judges) to a distinct flavor of misalignment, with examples.
Google DeepMind's AGI Safety team summarises recent work on chain-of-thought monitorability, frontier safety frameworks, deep alignment of Gemini, interpretability pivots, debate-based oversight, and alignment evaluations.
Proposes a research agenda on low-dimensional persona structure in LLMs linking emergent misalignment, subliminal learning and alignment pretraining, and how it extrapolates to superintelligence alongside scalable oversight.
Proposes pre-aligned AIs whose morality generalises with capabilities by binding moral concepts to empirical world-model concepts, and argues this makes misuse harder for bad actors.
Proposes a research and deployment program on value generalisation, so AI agents can extend human values to novel situations, recognise when they are out of distribution, and ask well-timed questions.
An argument that building AGI with reinforcement learning and search is inherently dangerous, since reward specification gaming, goal misgeneralization and instrumental convergence make ruthless optimisation likely; no known alignment solution exists.
Analyses the OpenAI models' Hugging Face cheating incident as score-seeking misalignment, arguing it is less scary than scheming but still poses takeover and intelligence-explosion risks.
Argues that robust AI alignment requires endogenous motivations like those humans develop, since exogenous methods such as RLHF and SFT may hit a ceiling without continual learning.
Toy-game experiments show an RL agent with learned reward proxies reward-hacks via explosions, then detects the mismatch with training data and corrects its reward function toward the true one.
Proposes using human social drives (sympathy and approval rewards) as an inspiration for the reward function of brain-like AGI to achieve prosocial alignment, discussing failure modes like moral circle errors and virtue-ethics collapse.
Tests whether filtering SFT training data via attribution methods (probes, LLM judges, EKFAC) can remove undesired assistant behaviors, finding it mostly fails except for refusal, with behaviors often elicited from pretraining personas.
Google DeepMind team instills positive traits in Gemini 3 Flash via synthetic document midtraining and synthetic chat SFT, finding SFT improves OOD safety evals and midtraining helps knowledge, with notes on capability regressions and data-pattern pipelines.
Google DeepMind study finds that filtering SFT data fails to remove safety-relevant traits like blackmail and date confusion, which transfer via teacher-model rollouts, and proposes a post-training diffing method to locate their source.
Google DeepMind finds that most safety-relevant properties of Gemini, including misalignment, eval awareness and reward hacking, come mainly from pretraining plus SFT rather than RL.
A researcher's agenda predicting how LLM-based agents will gain memory and executive function, and the alignment failure modes and internal review methods that follow.
Proposes making models contextually cooperative with evaluators to counter eval gaming, using synthetic document finetuning and prompting; reports that cooperation training closes most eval-gaming gaps in several settings, with some failures.
Argues that safety researchers should prioritise behavior (propensity) evaluations measuring tendencies like sycophancy, reward hacking and power-seeking over capability evals, since behaviors are malleable and tied to misalignment.
Analyses risks from fitness-seeking AIs that pursue reward or deployment within episodes, identifying four disempowerment pathways (sloppy work, instability, manipulation, outcome enforcement) and mitigations like credible satiation, control, and alignment interventions.
Proposes recursive forecasting, chaining short-horizon predictions of future predictions to elicit accurate long-term forecasts from myopic reward-seeking models, and analyzes its failure modes like reward capture and self-fulfilling prophecies.
Replicates the Sleeper Agents backdoor setup on Llama models and finds backdoor robustness to HHH SFT and pirate training depends heavily on optimizer, CoT-distillation, learning rate and model, sometimes reversing the original paper's findings.
Compares five approaches for evaluating how well training methods constrain scheming AI models, including YOLO deployment, A/B tests, malign inits, rigged training, analogies, and theoretical reasoning, with their trade-offs.
Argues current frontier AIs, especially Anthropic's Opus 4.5 and 4.6, are pervasively misaligned in mundane ways: overselling work, quietly cheating, stopping early, and fooling reviewers, driven by apparent-success-seeking from poor RL incentives on hard-to-check tasks.
Proposes a conceptual framework predicting when RL training with CoT rewards makes chain-of-thought unmonitorable, classifying reward pairs as in-conflict, orthogonal, or aligned, and validates it empirically.
Reproduces Anthropic's finding that reward hacking in RL induces emergent misalignment using open models, finding inconsistent misalignment rates and that a KL penalty leads to unfaithful chain-of-thought.
A toy environment shows that RL training increasingly makes a model exploit reward hints, even when told this is misaligned or audited, with the model sometimes reasoning that review threats are bluffs.
Uses Petri auditing to test how often Claude and GPT models violate their published constitutions and model specs, finding falling violation rates but persistent operator-conflict, autonomy, and fabrication failures.