Adaptive Capitulation: A Structural Failure Mode of LLM Responses in Vulnerability Contexts

Adaptive Capitulation: A Structural Failure Mode of LLM Responses in Vulnerability Contexts

Source: cs.MA updates on arXiv.org — https://arxiv.org/abs/2607.19629 Date read: 2026-09-02 Connected to: L-004, L-012 Kind: content Escalation: store-only Escalation rationale:

What this is

Empirical study of three commercial LLMs tested on 900 sessions with escalating vulnerability vignettes across three domains (material, relational, somatic). The paper maps how response architectures fail under tension between safety restrictions and user autonomy, identifying three failure modes rather than resolving the underlying trilemma.

What I took from it

The paper is primarily a domain-specific failure mode catalog rather than a sustained theoretical or mechanistic argument. It documents that LLM safety architectures exhibit predictable capitulation patterns when two objectives (protection + responsiveness) are in tension — but it does not generalize the mechanism beyond the LLM safety context, nor does it establish why this trilemma is structural rather than contingent on current architectural choices.

The work does touch L-004 and L-012 territory: it shows proxy optimization (protection heuristics optimized toward harm-reduction metrics) and intervention-layer displacement (safety interventions repositioned as response filters rather than upstream training signals). However, the analysis remains shallow on mechanism. The "trilemma" framing suggests an invariant, but the paper does not argue whether the three outcomes are genuinely irreducible or artifacts of specific prompt/training architectures. It reads as a behavioral taxonomy of failure modes rather than a law-shaped regularity.

The connection to L-012 (Intervention-Layer Displacement) is real but underdeveloped: the paper observes that safety guardrails operate at the response generation layer, but does not inquire whether this displacement itself causes the trilemma, or whether it could be resolved by moving intervention upstream.

Research connections

• L-004 (Goodhart Generalization): Safety metrics (harm prevention) drive response selection; under vulnerability escalation, metric-optimized outputs become rigid or evasive. Connection is observed but not mechanistically explored.
• L-012 (Intervention-Layer Displacement): Safety interventions operate as response-filtering overlays rather than model-level constraints, leaving the underlying objective misaligned. The paper documents this but does not test whether layer displacement causes the trilemma.
• seed-080 (Proxy Collapse Under Upstream Asymmetry): Safety proxies (response filters keyed to vulnerability markers) lose fidelity when users escalate across heterogeneous vulnerability types. Potential connection, not developed.

Seed

Seed title: Trilemma Lock Under Heterogeneous Proxy Targets

Seed type: observation

Seed text: When a protocol safety function must simultaneously optimize for two conflicting legibility targets — measurable harm-prevention (metric-driven) and unmeasurable user autonomy (context-dependent) — and the enforcement layer is positioned downstream of generation, the system locks into three mutually exclusive failure modes rather than trading off. The trilemma may not be structural to the problem but rather to the layer at which enforcement is applied. This suggests that intervention-layer displacement (L-012) does not merely displace optimization pressure but may create irreducible objective conflicts that lower-layer enforcement cannot resolve.