The paper investigates why task‑arithmetic merging fails by measuring the exact layerwise activation cross‑term of a factorial ledger. It shows that this cross‑term is largely transported and amplified by each block, is regenerated by untouched marginal paths, and varies monotonically with the displacement angle, yet it does not predict merge collapse. The study finds that behavioural performance is decoupled from the cross‑term, and that collapse is driven by marginal displacements rather than the cross‑term, which is only a bystander.
arXiv:2608. 11797v1 Announce Type: new Abstract: Model merging by task arithmetic works until it doesn't, and the field diagnoses why with magnitudes: layerwise representation bias, deviations from cross-task linearity, parameter overlap.
arXiv:2606.00930v2 Announce Type: replace-cross Abstract: Mechanistic interpretability routinely reads a probe and labels its top-activating units as the circuit executing the computation. We test th...
The paper presents a causal analysis of a compressed VLA policy that performs well in offline tests but fails in closed‑loop execution on a simulated pick‑and‑place task. An 8‑layer distillation of Octo‑Base retains most parameters and passes all offline metrics, yet collapses during deployment, with early stages degrading gradually and final transport failing entirely. The failure is traced to a negative, late‑heavy residual in the action trace, and standard remedies (continued training, offline data, command‑level compensation, clamping) do not restore performance; only a minimal‑pair intervention that mixes deployment‑distribution rollouts with teacher data restores parity with the teacher. whyItMatters":"The study demonstrates that offline validation metrics alone are insufficient to guarantee closed‑loop success for compressed policies, highlighting the need for targeted deployment‑time testing and interventions."
arXiv:2607. 09306v3 Announce Type: replace-cross Abstract: Behavioural auditing asks whether a language model behaves as it claims, but detection scores are reported without separating two targets: whether a reply was produced under a behaviour-inducing condition (exposure) and whether the behaviour surfaced in it (manifestation).
arXiv:2608. 03842v1 Announce Type: cross Abstract: When a language model fails on surface-perturbed input (typos, OCR noise, homophones), "which layer is responsible" has three natural operationalizations: where representations diverge most (sensitivity), where restoring clean activations recovers the prediction (causality), and where a small adapter can repair the damage (compensatory capacity) - and we show these three layer maps dissociate.
arXiv:2606. 00930v1 Announce Type: cross Abstract: Mechanistic interpretability often assumes that probes identifying a representational signature also identify the circuit executing the corresponding computation.
The paper reports that agent evaluations often show a tool‑call rate of zero even when the model emits valid calls, because the interface censors the trajectory before downstream components see it. Experiments on BFCL v4 and tau‑bench demonstrate that swapping the serving adapter can change the observed call rate from 0.00 to 0.96/0.19 or from 0 to 636 calls, indicating that the interface—not the model—causes the discrepancy. A 98‑line preflight check is released to detect such silent failures, highlighting that tool‑call rates depend on the model‑interface stack rather than the model alone.
arXiv:2608. 11212v1 Announce Type: new Abstract: Top-k Mixture-of-Experts (MoE) routing is discontinuous, so a deployment-motivated numerical disturbance -- simulated 4-bit KV-cache quantization read by a protected BF16 gate -- pushes tokens across decision boundaries and flips which experts fire.
arXiv:2609.25602v1 Announce Type: new Abstract: In language models, the choice between believing the prompt and believing the weights is made by a handful of identifiable attention heads. Instruction...
The paper investigates how inherited state affects sub-agent performance in multi-agent frameworks, comparing three inheritance policies—Reset, Selective, and Full—across a ladder of Qwen3 models. It finds that reliance on stale state decreases with model capability, but a mid-capability model (Qwen3‑1.7B) exhibits a statistically significant local minimum of net harm, defining a "danger band." Selective handoff consistently improves accuracy over Full, especially within the danger band, while a fixed-threshold router fails on other datasets.
arXiv:2606. 24952v1 Announce Type: cross Abstract: A central aspiration of mechanistic interpretability is controllability: if we know where a behavior is represented in a model's activations, we should be able to modify it.
The paper evaluates additive activation steering in chat and agent contexts, showing that the commonly used gain ratio (Δ_agent/Δ_chat) fails to reliably indicate potency and efficacy across multiple models and dose-response cells. By replacing the gain with a location metric, dEC50 (difference in EC50 between agent and chat), the authors demonstrate a more robust, two‑sided measure that consistently captures cross‑context shifts. The study also reports several refuted and unanswered claims, emphasizing that a single operating point cannot distinguish between displacement and gain effects.