Goto

Collaborating Authors

 Antarctica


On the Entropy Calibration of Language Models

arXiv.org Machine Learning

We study the problem of entropy calibration, which asks whether a language model's entropy over generations matches its log loss on human text. Past work found that models are miscalibrated, with entropy per step increasing (and text quality decreasing) as generations grow longer. This error accumulation is a fundamental problem in autoregressive models, and the standard solution is to truncate the distribution, which improves text quality at the cost of diversity. In this paper, we ask: is miscalibration likely to improve with scale, and is it theoretically possible to calibrate without tradeoffs? To build intuition, we first study a simplified theoretical setting to characterize the scaling behavior of miscalibration with respect to dataset size. We find that the scaling behavior depends on the power law exponent of the data distribution -- in particular, for a power law exponent close to 1, the scaling exponent is close to 0, meaning that miscalibration improves very slowly with scale. Next, we measure miscalibration empirically in language models ranging from 0.5B to 70B parameters. We find that the observed scaling behavior is similar to what is predicted by the simplified setting: our fitted scaling exponents for text are close to 0, meaning that larger models accumulate error at a similar rate as smaller ones. This scaling (or, lack thereof) provides one explanation for why we sample from larger models with similar amounts of truncation as smaller models, even though the larger models are of higher quality. However, truncation is not a satisfying solution because it comes at the cost of increased log loss. In theory, is it even possible to reduce entropy while preserving log loss? We prove that it is possible, if we assume access to a black box which can fit models to predict the future entropy of text.


Unravelling the mystery of the earliest life on Earth: Scientists uncover fresh chemical evidence of microbes in rocks more than 3.3 BILLION years old

Daily Mail - Science & tech

In 1996 Nasa and the White House made the explosive announcement that the rock contained traces of Martian bugs. The meteorite, catalogued as Allen Hills (ALH) 84001, crashed onto the frozen wastes of Antarctica 13,000 years ago and was recovered in 1984. Photographs were released showing elongated segmented objects that appeared strikingly lifelike.


UN Security Council to vote on Trump peace plan for Gaza

BBC News

The UN Security Council is expected to vote on a draft resolution backing Donald Trump's peace plan for Gaza. The text, submitted by the US, would give a mandate for the deployment of an International Stabilization Force (ISF) and to set up transitional governance there. The US says multiple unnamed countries have offered to contribute to the ISF, though it is unclear whether it would be required to ensure Hamas disarms or function as a peacekeeping force. Its formation is a central plank of Trump's 20-point plan which last month brought a ceasefire between Israel and Hamas in their two-year war. The draft also raises the possibility of a Palestinian state - something Israel strongly opposes.


Nature is not a blocker to housing growth, MPs find

BBC News

Nature is not a blocker to housing growth and the government risks missing both its housing and nature targets if it views it as one, a cross-party group of MPs has warned in a new report. The Planning and Infrastructure Bill overrides existing habitat protections, which the government has suggested is a barrier to its target to build 1.5 million houses by the end of this parliament. But in a report published on Sunday, the Environmental Audit Committee (EAC) found the measures outlined in the bill are not enough to allow the government to meet its goals. Using nature as a scapegoat means that the government will be less effective at tackling some of the genuine challenges facing the planning system, the report said. A Ministry of Housing spokesperson said it was fixing a failing system with landmark reforms, which would deliver a win-win for the economy and the environment.


Zelensky vows energy sector overhaul after 100m corruption scandal

BBC News

Ukrainian President Volodymyr Zelensky has vowed to overhaul state-owned energy companies, after a major corruption scandal engulfed the country's energy sector. Around $100 million (ยฃ76m) has been embezzled, anti-graft investigators said, causing outrage in a country where Russian attacks have resulted in crippling power outages. Alongside a full audit of their financial activities, the management of these companies is to be renewed, Zelensky wrote in a post on X on Saturday. Energoatom, the state nuclear company at the heart of the scandal, will have a new supervisory board within a week, he added. Several of those implicated in the scandal have close links to the Ukrainian president.


Hollywood's SAG Awards announces it will change its name

BBC News

Hollywood's SAG Awards announces it will change its name The Screen Actors Guild Awards, the marquee awards ceremony honouring actors, is getting a new name. Known colloquially as the SAG Awards, the awards show will now be dubbed the Actor Awards presented by Sag-Aftra, the labour union representing US film, television and radio actors. Since the beginning, our statue has been called'The Actor' and we're a show that's entirely about actors, so this new name is a perfect next step in the show's evolution, the show's executive producer said on Friday. The rebrand comes ahead of the 32nd edition of the star-studded ceremony, which is set for 1 March 2026. The award show's executive producer Jon Brockett told the BBC that the name change - which was announced at a board meeting on Friday - gives viewers in more than 190 countries an immediate understanding of who we are and what we're about - a show about actors honouring actors.


'Astonishingly lethal': BBC reports from site of Russian strike in Kyiv

BBC News

At least six people have been killed in a wave of Russia strikes on Kyiv, which the Ukrainian President Volodymyr Zelensky has condemned as a heinous attack. The BBC's James Landale visited the scene of one attack in eastern Kyiv where a drone rammed through a block of flats and left six people dead. Several other regions were also targeted. A drone attack on a market at Chornomorsk in the south of the country killed two people. Catherine Connolly has'never believed more' in the spirit of Ireland New Irish President Catherine Connolly says she has been given a powerful mandate to articulate a vision for a new republic.


UK billionaire Joe Lewis receives pardon from Trump

BBC News

Billionaire UK businessman Joe Lewis, whose family trust owns Tottenham Hotspur football club, has received a pardon from US President Donald Trump. Lewis, 88, pleaded guilty to insider trading as part of an agreement with prosecutors in 2024 that saw him avoid prison. He was accused of passing on information about his companies to his private pilots, friends, personal assistants and romantic partners in a fraud that authorities said netted millions of dollars in profit. A White House official said Trump approved the pardon for Lewis, who requested it so he could receive medical treatment and visit his grandchildren and great grandchildren in the US. Mr Lewis admitted he made a terrible mistake, did not fight extradition in the case, and paid a $5 million fine, the official told the BBC.


Lack of trust and racism concerns: Five key failings in Sara Sharif review

BBC News

An independent review of the Sara Sharif case has identified multiple failings from agencies before her murder in Surrey in 2023, following two years of abuse. The child safeguarding practice review, published on Thursday, said there were clearly several points in Sara's life, in particular during the last few months, where different actions could and should have been taken by the authorities. The system failed to keep her safe, it added. Responding to the report, the Children's Commissioner said the case was a catalogue of missed opportunities, poor communication and ill-informed assumptions. The education secretary said there had been the glaring failures across all agencies.


Drugs disguised as tea keep washing up on this S Korean holiday island

BBC News

Since September, residents on South Korea's Jeju island have been spotting small packs of what appear to be bags of Chinese tea washed ashore. Upon closer inspection, however, they were found to contain ketamine. Some 28kg (62 lbs) of the drug, wrapped in foil and labelled with the Chinese character for tea, have been found on at least eight occasions, police say. Ketamine is used as an anaesthetic in medical procedures, but its recreational use is illegal in South Korea. It can cause severe physical and mental damage, including to the heart and lungs, when misused.