One-Line Summary

Faster, more controllable video generation arrives via causal distillation, while new evaluations uncover inaudible-audio risks and limits in agents’ verified coding and instruction-following.

Research Papers

Causal distillation speeds up interactive, controllable video generation

This paper introduces a way to make AI video tools render faster without losing the ability to steer motion and camera in real time. The method, called Context-Matched Distillation (CMD), reframes training so the teacher only “grades” each generated frame using past information, not future frames or controls — a causal take on distribution matching distillation (DMD). It also adds Prefix Scoring (evaluate outputs under the student’s own cached prefix) and Prefix Corruption (lightly perturb early, unreliable prefixes) to keep supervision aligned with what the student actually rolled out. ¹

Instead of a bidirectional teacher that can peek at the whole clip, CMD trains a causal teacher that never sees the future and even initializes the few-step student, keeping the same causal setup for teacher training, student distillation, and inference. This closes a long-standing mismatch in few-step distillation, where students are trained under information they will not have at generation time. ¹

Across short- and long‑video benchmarks, the authors report state-of-the-art aggregate results among autoregressive methods and stronger adherence to time‑varying camera controls. CMD’s causal framing naturally extends to frame‑wise or chunk‑wise generation, long‑video distillation, and camera‑conditioned control. ¹

Inaudible low-frequency audio prompts fool audio-language models

The authors show that adding very low‑frequency signals that people can’t hear can still enter the system and derail Large Audio‑Language Models (LALMs). Their red‑teaming method, Intermittent Low‑Frequency Lockout (ILL), uses a universal waveform template in a black‑box setup and cuts accuracy by up to 67 percentage points while testers rate the audio’s loudness at 1.33, close to 1.17 for clean audio. ²

To mitigate the risk, they propose Distributional Requery Guard (DRG), which detects low‑frequency distribution shifts and conditionally asks for a second recording to recover meaning. In attacks across six LALMs and multiple tasks, DRG lifts mean attacked accuracy from 28.5% to 46.1% after clean reacquisition, highlighting a practical defense path for audio inputs. ²

Hybrid-policy self-editing makes knowledge updates stick

This work targets a common failure: after you inject a new passage, Large Language Models (LLMs) can repeat it but still fail to answer detailed questions about its facts or combine them in multi‑hop reasoning. The authors cast knowledge editing (KE) — and specifically unstructured knowledge editing (UKE), where the edit is a free‑form passage — as proactive self‑distillation from a privileged in‑context state of the same model, needing no external labels. ³

They introduce Hybrid‑Policy Self‑Editing (HPSE), which keeps the student on its own rollout most of the time but “steps in” exactly where coverage of the new facts would otherwise fail. The paper provides theory for why this beats pure on‑policy distillation and shows plug‑and‑play gains across four LLM backbones and two KE editors in varied scenarios. ³

Vero benchmarks repository-scale verified coding; Harness-IF audits agent rule‑following

Vero asks coding agents to implement software and produce machine‑checked proofs that the code meets its specification — not just on single functions, but across full, multi‑module repositories. It packages 43 instances (Python, Dafny, Verus, Coq) as Lean 4 repos with predetermined Application Programming Interface (API) boundaries, curated formal specs, and reference implementations, and even lets agents prove a spec is unsatisfiable or a reference is wrong during curation. With Lean toolchain access, the strongest agent fully solves 27 of 43 instances and closes no specs on the hardest repos, underscoring today’s gap to repository‑scale verified synthesis. The benchmark, curation pipeline, and evaluation harness are released. ⁴

A companion line of evaluation, Harness‑IF (Harness for Instruction Following), isolates whether agents truly obey operational rules rather than coincidentally doing the right thing. It scores 60 realistic multi‑turn coding items with 256 rules across five instruction “surfaces,” and introduces Against‑Prior Accuracy (AP‑Acc) that only counts rules labeled as opposing unprompted defaults. Across 12 frontier models, accuracy spans 72.1–85.9% and AP‑Acc 66.1–78.6%; every model is worse on against‑prior rules by 3.6 to 7.4 points (mean 5.81), and aggregate scores overstate compliance by a model‑specific margin. Precedence does not simply follow prompt depth: system prompts, project files, and user instructions rank ahead of tool and skill descriptions. ⁵

These two efforts together pressure‑test agents on both coherence across repositories and obedience to operational constraints, giving practitioners clearer signals than task‑completion alone. ⁵

Why It Matters

When AI systems are judged under the same constraints they face at runtime — past‑only information for video generation, inaudible noise in audio, missing coverage during edits, or repository‑wide proof obligations — gaps show up that average scores miss; these papers push training and evaluation toward those real‑world boundaries. ¹

이번 주 시도해볼 것

1. Proof‑aware coding demo: Skim the Vero paper’s examples to see what “code + proof” looks like in practice and how failure modes are categorized.
2. Audio hygiene check: Try recording the same sentence twice in a quiet room and a noisy one, then reflect on how a safeguard like DRG would trigger a re‑recording when distribution shifts are detected.