Large language models (LLMs) require alignment, such as instruction-tuning or reinforcement learning from human feedback, to effectively and safely follow user instructions. This process necessitates training aligned versions for every model size in each model family, resulting in significant computational overhead. In this work, we propose nudging, a simple, plug-and-play, and training-free algorithm that aligns any base model at inference time using a small aligned model. Nudging is motivated by recent findings that alignment primarily alters the model's behavior on a small subset of stylistic tokens, such as "Sure" or "Thank". We find that base models are significantly more uncertain when generating these tokens. Leveraging this observation, nudging employs a small aligned model to generate nudging tokens to steer the large base model's output toward desired directions when the base model's uncertainty is high. We evaluate the effectiveness of nudging across 3 model families and 13 tasks, covering reasoning, general knowledge, instruction following, and safety benchmarks. Without any additional training, nudging a large base model with a 7x - 14x smaller aligned model achieves zero-shot performance comparable to, and sometimes surpassing, that of large aligned models. For example, nudging OLMo-7b with OLMo-1b-instruct, affecting less than 9% of tokens, achieves a 10% absolute improvement on GSM8K over OLMo-7b-instruct. Unlike prior inference-time tuning methods, nudging enables off-the-shelf collaboration between model families. For instance, nudging Gemma-2-27b with Llama-2-7b-chat outperforms Llama-2-70b-chat on various tasks. Overall, this work introduces a simple yet powerful approach to token-level model collaboration, offering a modular solution to LLM alignment. Our project website: https://fywalter.github.io/nudging/ .

Active learning methods aim to improve sample complexity in machine learning. In this work, we investigate an active learning scheme via a novel gradient-free cutting-plane training method for ReLU networks of arbitrary depth. We demonstrate, for the first time, that cutting-plane algorithms, traditionally used in linear models, can be extended to deep neural networks despite their nonconvexity and nonlinear decision boundaries. Our results demonstrate that these methods provide a promising alternative to the commonly employed gradient-based optimization techniques in large-scale neural networks. Moreover, this training method induces the first deep active learning scheme known to achieve convergence guarantees. We exemplify the effectiveness of our proposed active learning method against popular deep active learning baselines via both synthetic data experiments and sentimental classification task on real datasets.

The aim of the notes is to demonstrate the potential for ideas in machine learning to impact on the fields of inverse problems and data assimilation. The perspective is one that is primarily aimed at researchers from inverse problems and/or data assimilation who wish to see a mathematical presentation of machine learning as it pertains to their fields. As a by-product of the presentation we present a succinct mathematical treatment of various topics in machine learning. The material on machine learning, along with some other related topics, is summarized in Part III, Appendix. Part I of the notes is concerned with inverse problems, employing material from Part III; Part II of the notes is concerned with data assimilation, employing material from Parts I and III.

Israeli rescue services say nearly 60 people were injured, some critically, in a drone strike in Binyamina on Sunday, which Hezbollah has claimed responsibility for. Rescue services in Israel said over 60 people were wounded, some of them critically, in a drone strike in Binyamina, Israel, which the Lebanon-based Hezbollah militant group has claimed responsibility for, according to reports. Israeli media reported that two drones were launched from Lebanon, one of which was intercepted. Who was hurt – whether military members or civilians – or what was struck was not immediately clear. On Thursday, Israel conducted two strikes in Beirut that killed 22 people, and Hezbollah said it was retaliating for the strikes by targeting an Israeli military training camp.

At least 67 people have been wounded in a drone attack in northern Israel, according to Israeli emergency services and local media, as the Lebanese armed group Hezbollah said it had targeted an Israeli military camp with a "swarm" of drones. Israeli Army Radio reported that at least four people were critically wounded in the attack on Sunday in the town of Binyamina, south of Haifa. According to Israel's Channel 12, no warning sirens were heard before the attack. Hezbollah claimed responsibility for the strike. In a statement, the Iran-aligned group said it launched a "swarm of drones" at a Golani Brigade camp.

Ukraine's human rights ombudsman has urged international organisations to respond to an allegation that several Ukrainian prisoners of war were executed in Russia's Kursk region. DeepState, a Ukrainian battlefield analysis site close to Ukraine's Defence Ministry, said that Russian troops killed nine Ukrainian "drone operators and contractors" on October 10 after they had surrendered. Dmytro Lubinets said on Sunday that the international community "must not turn a blind eye" to the alleged executions. The ombudsman wrote on Telegram that he had sent letters to the United Nations and the International Red Cross about the incident, referring to it as "another crime committed by the Russians". Ukrainian Prosecutor-General Andriy Kostin also said his office had opened a criminal investigation into the alleged execution and said killing prisoners of war was a "gross violation" of the Geneva Convention.

A mysterious fleet of drones entered restricted airspace and swarmed a U.S. military base along the Virginia coast for 17 days late last year, stumping the Pentagon, according to a new report. For several nights last December, U.S. military personnel reported witnessing a fleet of unknown unmanned aircraft breach restricted airspace over a stretch of land at Langley Air Force Base along Virginia's shore, the Wall Street Journal first reported. The drones would start to arrive about 45 minutes to an hour after sunset each night, one official reportedly told U.S. Air Force Gen. Mark Kelly, who joined several other officers responsible for the country's most advanced jet fighters, including F-22 Raptors, on a squadron rooftop. Kelly described the first drone he saw as roughly 20 feet long and flying at more than 100 miles an hour, at an altitude of roughly 3,000 to 4,000 feet. As many as a dozen or more drones followed, flying across Chesapeake Bay, and then traveling toward Norfolk, Virginia, and through a space overlooking the base for the Navy's SEAL Team Six and Naval Station Norfolk, the world's largest naval port, according to the Journal.

More than 200 people have taken part in the World Conker Championships, with many competing in fancy dress. The competition took place earlier at the Shuckburgh Arms in Southwick, Northamptonshire. The event saw participants go head-to-head using conkers threaded on to string to try and smash their opponent's nut. Since its inception in 1965, the event has raised more than 400,000 for charities that support the visually impaired.PA MediaHundreds of spectators attended the event which was first held in 1965 One man wore a green inflatable Yoda headpiece, while another wore a conker-themed hat. All participants were required to follow a stringent set of rules to ensure the event was as fair as possible, which included the conkers and laces being provided by organisers.

She's not long on charisma or passion but keeps perfect rhythm and is never prone to temperamental outbursts against the musicians beneath her three batons. Meet MAiRA Pro S, the next-generation robot conductor who made her debut this weekend in Dresden. Her two performances in the eastern German city are intended to show off the latest advances in machine maestros, as well as music written explicitly to harness 21st-century technology. The artistic director of Dresden's Sinfoniker, Markus Rindt, said the intention was "not to replace human beings" but to perform complex music that human conductors would find impossible. The Sinfoniker, long known for innovation and political statements, is celebrating its 25th anniversary with the Robotersinfonie at the Hellerau hall in a concert divided into two parts, one purely human and, after the interval, one that is robot-led.

SpaceX has launched its fifth Starship test flight from Texas and returned the rocket's towering first-stage booster back to land for the first time, achieving a novel recovery method involving large metal arms. The rocket's Super Heavy first-stage booster lifted off at 7:25 am (12:25 GMT) on Sunday from SpaceX's launch facilities in Boca Chica, Texas, sending the second-stage Starship rocket on a path in space bound for the Indian Ocean west of Australia, where it will attempt atmospheric reentry followed by a water landing. The Super Heavy booster, after separating from the Starship booster some 74km (46 miles) in altitude, returned to the same area from which it was launched to make its landing attempt, aided by two robotic arms attached to the launch tower. "The tower has caught the rocket!!" SpaceX founder Elon Musk posted on X. Towering almost 121 metres (400 feet), the empty Starship arched over the Gulf of Mexico like the four Starships before it that ended up being destroyed, either soon after liftoff or while ditching into the sea. The last one in June was the most successful yet, completing its flight without exploding.