Large language models: where reliability lives

A reading map: model, system, evidence

These six selected papers approach large language models through different questions: how safety survives intervention, how agent teams recover, how histories support decisions, who supplies preferences, and what counts as evidence of forgetting. Their shared topic is reliability, not agreement on one solution. A useful starting point is the graph-mediated systems survey, which separates what an operation does from which component computes it. It also distinguishes local improvements from downstream task outcomes—a useful discipline for reading this collection without treating unlike results as a common leaderboard.

Two meanings of distributed alignment

Addition by Subtraction targets a model-internal vulnerability: safety behavior can depend on a small set of components that attackers disable. ABS-Align progressively deactivates identified safety-critical neurons during preference-based fine-tuning, training the remaining model to maintain aligned behavior. The authors report improved resistance to pruning, refusal-direction ablation and black-box jailbreaks across four models, while broadly preserving capabilities. Crucially, they do not directly measure the absolute density of safety representations; “denser” describes reduced dependence on identified components, not a complete map of safety inside the network.

The distributed post-training vision paper asks a different question: can preference learning operate without gathering sensitive feedback centrally? Its argument connects data sovereignty and pluralism with federated or decentralized optimization. Here, distribution concerns participating institutions and model updates, not neurons. The papers therefore cannot be read as mutually validating demonstrations of “distributed safety.” The vision paper identifies reference drift, conflicting preferences and privacy–utility trade-offs; its extracted small-scale experiments also show that synchronization can behave differently in federated and decentralized settings. It offers a research agenda, not evidence that decentralization automatically delivers safer or more representative models.

Keeping decisions coherent over time

Read MA-Mesh alongside the smart-home data-association paper to see two distinct approaches to continuity. MA-Mesh puts a recovery layer around collaborating agents: peers assess role-specific performance, and failover replaces agents when degradation or timeouts trigger recovery. The authors report gains under injected agent, prompt and communication failures, with an additional cloud-native testbed evaluation. Its boundary matters: the resilient proxies and recovery meta-agent are assumed to remain available.

The smart-home work instead tracks who triggered anonymous sensor events. It represents assignment history within a Markov decision process and learns from expert demonstrations through behavior cloning. A lightweight Transformer predicts assignments, while a fine-tuned LLM generates them as text. These are complementary policy implementations, not evidence that every sequence problem needs an LLM. Reported benchmark improvements and preliminary cross-environment transfer coexist with accumulating identity errors over longer horizons and a continuing need for environment-specific supervision.

The connection is error propagation, but the remedies and measurements differ: replacing an underperforming agent is not the same as correcting a resident history. The graph survey supplies a further question for both: which downstream decisions depend on information that has changed? Its discussion of memory maintenance and dependency-aware recovery is an analytical lens, not an experimental validation of either system.

Auditing is not the same as removing

Auditable Unlearning for LLMs adds another distinction: forgetting performance, retained capability and resistance to membership inference do not themselves provide a provider-side audit signal. GapWiden deliberately separates statistics for requested and retained data; GapVeil uses inference-time rejection sampling to limit the signal exposed through responses. Reported improvements do not eliminate all trade-offs: retained performance remains below the retrained reference.

This creates a sharp access-model contrast with ABS-Align. The unlearning framework assumes confidential weights and protected black-box interaction; its authors warn that white-box access could expose the engineered audit gap. ABS-Align explicitly studies attacks on accessible internals. Neither result transfers automatically across that boundary. A constructed statistical audit signal should not be confused with an unconditional guarantee of data erasure.

Evidence boundaries and open questions

This guide uses supplied abstracts and selected extracted passages, not complete papers. All six have paper-text excerpts, so none is treated as abstract-only; abstract-only evidence would not support reconstructing methods or detailed results. Extraction can disrupt tables, equations and reading order. The unlearning abstract and extracted text even differ on benchmark coverage, so no combined benchmark count is asserted here. Missing excerpt detail is not evidence of missing analysis in a paper.

These are supplied arXiv v1 manuscripts; peer-review status is not established, and the unlearning text explicitly says “Preprint.” Across this selected coverage, the unresolved reading questions are concrete: what access is assumed, which failures are recoverable, what survives a changed environment, and whether a measured improvement reaches the final task. Those questions connect the papers without implying consensus or a field-wide trend.