#1 ·
Tree-Ising: Repairing Language Models with Bit-Level Ising Edits
Hadi Salloum, Ammar Ali
Model EditingInterpretabilityDiscrete Optimization
Why it matters
Weight-space repair by an LLM agent was striking but costly: 158 agent turns and 89.7M tokens to fix one model's reasoning loops. Tree-Ising shows the same repair can be a deterministic optimization over the model's own BF16 bits — no gradients, no LLM in the loop, roughly $0.90 of GPU time. For practitioners, model surgery starts to look like a solvable engineering problem, and the recipe transferring from 0.8B to 2B without re-derivation is the encouraging part.
Building on the GPT-6 Astra weight-space repair result, the authors cast fixing a model as sparse discrete optimization over its own BF16 weights: admissible edits are sign flips and exponent shifts that preserve shared group exponents and mantissas, with an objective that weakens the loop direction, strengthens a concluding direction, and preserves layer output on productive tokens. That objective decomposes exactly into independent QUBO (Ising) problems, one per output row, solved on GPU with simulated annealing or quantum-inspired simulated bifurcation. On the same two layers Astra edited in Qwen3.5-0.8B, the bit-level repair reaches 54.6% on MATH-500 versus the agent's 51.7% (base 7.9%), modifying 13-17% fewer weights with 2.2-2.3x fewer bit flips in four forward passes. The recipe also lifts Qwen3.5-2B greedy GSM8K from 21.6% to 69.4%, and an eight-variable subproblem is executed on an IBM Heron QPU as proof of concept.
arXiv | HuggingFace
#2 ·
Few-Step Streaming Video Generation via Condition Shifting
Deyuan Liu, Jungang Li, Kaiwen Zheng, Kai Jiang
Video GenerationDistillationEfficiency
Why it matters
Few-step streaming video is the architecture interactive video apps need, but DMD-style distillation pays for a separate fake-score network — extra memory, extra compute, and a target that must chase the moving generator. Condition shifting deletes that network by having the generator estimate its own fake distribution through a shifted condition, cutting peak per-GPU memory 4.4% and iteration time 9.2% under matched settings without quality loss.
The paper extends APEX, a few-step distillation recipe, to streaming video. Instead of a separate fake-score network, the generator's own parameters estimate the velocity field of the generated distribution: the original text condition selects generation, a shifted condition selects fake estimation. The frozen reference teacher, Self-Replay Forcing, and perceptual regularization are retained. On 180-second videos (three backbones per block plus a one-step spatial refiner), APEX beats S2-Avatar on instruction fulfillment (3.856 vs 3.831 on the Progressive track) and aesthetics on both tracks, with human raters showing +30% net preference for instruction following; motion quality still favors S2-Avatar and sync is mostly tied. Under matched training, it uses 4.4% less peak allocated GPU memory per GPU and iterates 9.2% faster than DMD.
arXiv | HuggingFace
#3 ·
World In-Context Learning for Agentic Robot Manipulation
Heng Fang, Zehui Guo, Yuehan Zhang, Mingyang Du
RoboticsIn-Context LearningWorld Models
Why it matters
Trained manipulation policies bake skill into weights and stay coupled to their training distribution; WICL writes learned skill into context instead, which a frozen model can transfer across scenes with no gradient updates. One demonstration per task and 83.4% mean success on RoboTwin 2.0 — over 20 points above the strongest trained policy — plus transfer to a physical dual-arm robot makes this a serious alternative to fine-tuning for real robots.
The key observation is that what a model learns in context is semantic — task steps and object roles, not visual appearance — so experience recorded in one world transfers to another. World In-Context Learning operationalizes this: a frozen model starts from a demonstration, rehearses the task in a world such as a simulator, and writes the resulting feedback into its context rather than its weights. The RoboScript implementation freezes rehearsal output as a Task Memory and deploys it on new scenes with data isolation guaranteeing every deployment input exists on the real robot. On all 50 RoboTwin 2.0 tasks it reaches 83.4% mean success across clean and randomized settings with one demonstration per task and no gradient updates, exceeds the strongest trained action policy by over 20 points, improves through reflection, and transfers to a physical dual-arm robot via cross-embodiment memory adaptation.
arXiv | HuggingFace
Outlook
The throughline is repair and transfer without retraining: discrete optimization over bits, a generator supervising itself, and a frozen model that learns a task in context as well as in weights. Each direction now has a cheap, deterministic baseline, which is what makes it a direction — expect other labs to race the $0.90 annealing recipe and the 83.4% one-demonstration manipulation result.