A side network restores the safety fine-tuning strips
What helps someone shipping a model tonight sits in a handful of day-old arxiv papers rather than the launch feed, and in a $5.34 billion deal that reads the compute bottleneck as electricity instead of silicon.
HyperSafe restores the safety alignment fine-tuning strips, without touching the weights1
Fine-tune an open model for a downstream task and its safety alignment degrades. HyperSafe reads the fine-tuned model's layer-wise activation fingerprints and, in a single forward pass, generates a per-checkpoint Safe Side Network that runs alongside the frozen base, with no gradient updates and no safety data at deployment. On Qwen2-7B and LLaMA-3-8B it drops the harmful-response rate from 19–31% to below 1% on every held-out checkpoint, with downstream task accuracy within 1% of the fine-tuned baseline.1
Why it matters: anyone shipping a fine-tuned open-weights model inherits a safety regression they rarely measure. This is a per-checkpoint guard generated from the model's own activations, not a generic filter, so it is worth wiring into a release pipeline before the model ships. Read the sub-1% as a ceiling under the paper's threat model on two 7-to-8B models, not a guarantee against an adversarial red team.
SCOPE-RL raises reasoning accuracy and cuts tokens at the same time2
Reinforcement learning with verifiable rewards (RLVR) scores only the final answer, so the model gets no signal on its intermediate steps. SCOPE-RL adds verifiable rewards on sub-question chains before a trajectory succeeds and refines the correct trajectories afterward. On Qwen3-8B-Instruct that is worth up to 11.2 percentage points of accuracy and up to a 27.1% reduction in tokens, held across model sizes and RLVR variants.2
Why it matters: the usual RLVR result trades accuracy for tokens in one direction or the other, and getting both at once is the part to verify. If you already run RLVR on a reasoning model, the lever here is a process reward on the steps, not a larger final-answer bonus. The "up to" figures come from a single 8B model and assume you can decompose the task into sub-questions to score, which many domains cannot.
Williams takes $5.34 billion from Blackstone to wire power straight to data centers3
On July 13 Williams Companies sold a 49% stake in five behind-the-meter Power Innovation projects to funds led by Blackstone Credit & Insurance, with Apollo- and KKR-managed vehicles alongside, for $5.34 billion of committed capital.3 Williams keeps 51% and operational and commercial control; the total breaks down as about $4.4 billion for 49% of expected growth capex plus roughly $0.9 billion of additional consideration.4 Behind-the-meter means the generation feeds large users, data centers among them, directly rather than through the grid.
Why it matters: the GPU allocation a builder waits on is, upstream, a power-interconnection problem, and the binding constraint on new compute has moved from chips to the electricity feeding them. Behind-the-meter generation is how operators route around the grid-interconnection queue that gates when capacity comes online. This is committed capital on pre-completion projects, not delivered megawatts, so read it as where the money is pointing, not supply you can book this year.
FastTPS speeds up on-device decoding 6x, once you read the baseline5
FastTPS accelerates the sequential token (decode) phase of inference on an AI accelerator through KV-cache concatenation, RoPE-attention changes, and fused MLP operations, at low precision loss. On an AMD Ryzen AI 300-series NPU at BF16 it reaches 93% of peak memory bandwidth and a 6x speedup over a non-fused baseline, benchmarked on Phi-3-mini-4k-instruct.5
Why it matters: decode on a laptop-class NPU is memory-bandwidth bound, so the 93% of peak is the real result and the 6x follows from it. If you ship small models to NPUs, that bandwidth ceiling is the number to design against. The 6x is measured against an unoptimized floor, not a production runtime, so it is not a speedup over whatever you run today.
ToFu ships a coding-agent harness you can open up, under MIT6
ToFu is a coding-agent harness that reads a codebase, edits files, runs commands, and integrates with dev tools, released so researchers can inspect and modify its orchestration and tool-use logic and self-host it for privacy-sensitive work, under an MIT license. Its authors claim better token efficiency, lower cost, and multilingual capability than existing agentic harnesses, but the abstract carries no benchmark numbers, so those comparisons stay unquantified.6
Why it matters: most agent harnesses are closed boxes. An MIT one a team can read, modify, and run in-house is the news, not a measured win over the incumbents. If you are building agents and need to see how the loop calls tools, this is worth a look before you commit to a vendor's black box; wait for numbers before you credit the efficiency claims.
PUST moves post-training gains from a cheap proxy model to a bigger one7
PUST (Proxy-guided Update Signal Transfer) splits post-training in two: a lightweight proxy model explores which behaviors help, and the relative-improvement signal it finds transfers to the larger primary model, so the expensive online run becomes modular and reusable rather than one monolithic pass. The authors report this across Qwen3-family models on math and code, but the abstract exposes no numeric deltas, so weak-to-strong transfer here is a demonstrated direction, not a benchmarked gain.7
Why it matters: if update signals are reusable across model sizes, you can run costly exploration once on a cheap proxy and amortize it over larger models, a different cost structure for post-training than one online run per model. There is nothing to budget against yet, so watch for the follow-up that quantifies the transfer before you re-plan a pipeline around it.
Sources
- arXiv · HyperSafe: inference-time restoration of safety alignment in fine-tuned models
- arXiv · SCOPE-RL: process rewards for reasoning accuracy and token efficiency
- Williams · Williams announces $5.34 billion investment in Power Innovation joint venture from Blackstone
- U.S. SEC · Williams Companies Form 8-K, Exhibit 99.1 (Power Innovation joint venture)
- arXiv · FastTPS: accelerating the token phase of on-device LLM inference
- arXiv · ToFu: an MIT-licensed white-box coding-agent harness
- arXiv · PUST: Proxy-guided Update Signal Transfer for post-training