arXiv AI
Jun 2
Lying Is Just a Phase: The Hidden Alignment Transition in Language Model Scaling
arXiv:2605. 18838v3 Announce Type: replace-cross Abstract: Scaling laws predict loss from compute but not how capabilities interact.
TwistedMerge: Certified Higher-Order Diagnostics and Abstention for Model Merging
arXiv:2607. 20887v1 Announce Type: cross Abstract: Model merging combines independently trained or fine-tuned models, but pairwise alignability does not imply globally consistent alignment.
Related stories
arXiv Computation and Language
Sep 15
SHIFT-M3: Pre-fusion Alignment-based Consistency Screening for Multimodal ECG Record Integrity
arXiv:2609.13874v1 Announce Type: new Abstract: Multimodal clinical AI typically assumes that the waveform, report, metadata, and downstream predictions attached to a record belong to the same patien...
arXiv AI
Jul 21
First-Order Predictable but Pairwise Fragile: Local Task Adaptation in Trained Transformers
arXiv:2607. 16821v1 Announce Type: cross Abstract: Task arithmetic, sequential fine-tuning, activation steering, and first-order random search all operate through relatively small perturbations around an already trained checkpoint, and they rely on different local approximations: individual perturbations should be first-order predictable, task updates should compose with controlled interference, useful tangent structure should be stable and possible to estimate, and weight edits should have counterparts in representation space.
arXiv Machine Learning
Jul 1
Signed-Permutation Coordinate Transport for RMSNorm Transformers
arXiv:2606. 31963v1 Announce Type: new Abstract: Modern LLM workflows move coordinate-indexed objects across checkpoints: steering vectors, sparse autoencoders, top-$k$ neuron sets, attribution lists, and merge alignments.
arXiv Machine Learning
Oct 8
When Rank Rises as LLMs Degrade
The paper investigates how spectral statistics, particularly RankMe and covariance effective rank, behave when large language models (LLMs) undergo post‑training in non‑stationary environments. In controlled experiments on Qwen3‑0.6B, data duplication causes a 75% increase in held‑out loss while simultaneously raising both original and centred RankMe, indicating spectral dispersion rather than rank collapse. The authors demonstrate that a two‑sided, multichannel sequential monitor can detect various damage regimes but does not predict held‑out loss earlier and requires healthy calibration data.
arXiv Machine Learning
Jul 9
Guidance Breaks the Fitted Operator: A Terminal-Fitted Repair for Classifier-Free Guidance
arXiv:2607. 07665v1 Announce Type: new Abstract: Classifier-free guidance (CFG) is the standard way to strengthen class-conditioning in diffusion and flow-matching samplers, yet at large guidance it oversaturates and destabilizes, symptoms practitioners suppress with more steps or limited-interval schedules.
arXiv Machine Learning
Jul 27
Multi-Horizon Consistency as Geometry: When Latent Dynamics Contract, and When They Do Not
arXiv:2607. 21645v1 Announce Type: new Abstract: Multi-horizon latent consistency is a common training knob in video predictors and world models, but practitioners rarely know what it does to transition geometry.
arXiv AI
Jun 9
Trait-space Monitoring for Emergent Misalignment During Supervised Finetuning
arXiv:2606. 07631v1 Announce Type: cross Abstract: Emergent misalignment (EM) occurs when narrow finetuning causes a model to behave dangerously outside the finetuning task.
arXiv Machine Learning
Aug 6
Canonical Joint Energy-Based Model on CIFAR-10: failure modes and practical indistinguishability of Predictor-Corrector and SGLD samplers
arXiv:2608. 05025v1 Announce Type: new Abstract: Joint Energy-Based Models (JEM) unify classification and generation within a single network and support out-of-distribution (OOD) detection.
arXiv Machine Learning
Jun 2
Measuring Alignment-Induced Activation Shifts Correctly: A Template-Controlled Difference-in-Differences Protocol
arXiv:2605. 24583v3 Announce Type: replace Abstract: Comparing a model's internal activations before and after alignment is a natural way to ask what safety training changes: one forms the matrix of paired aligned-minus-base activations on safety-relevant inputs and reads off its effective rank or top direction.
arXiv AI
Jun 3
Exact equivariance, kept through training, buys zero-shot generalisation across the symmetry group
arXiv:2606. 03003v1 Announce Type: cross Abstract: A latent world model built from an equivariant encoder $E$ and an equivariant predictor $f$ inherits a provable symmetry of its training loss: when the world's dynamics genuinely carries a group $G$ acting on latents by an orthogonal representation $\rho(g)$, the one-step prediction relMSE is exactly invariant across the whole group, so fitting the dynamics on a restricted slice of orientations mathematically determines it on the entire orbit (j\v{u} y\=i f\v{a}n s\=an).
arXiv Machine Learning
Jul 2
Measuring Dead Directions: Decomposing and Classifying Singular Structure off Canonical Alignment
arXiv:2607. 00603v1 Announce Type: new Abstract: We give a descent-free, alignment-free measurement of singular structure on trained networks.