Women of color running for Congress in 2024 have faced a disproportionate number of attacks on X compared with other candidates, according to a new report from the nonprofit Center for Democracy and Technology (CDT) and the University of Pittsburgh.

The report sought to “compare the levels of offensive speech and hate speech that different groups of Congressional candidates are targeted with based on race and gender, with a particular emphasis on women of color.” To do this, the report’s authors analyzed 800,000 tweets that covered a three-month period between May 20 and August 23 of this year. That dataset represented all posts mentioning a candidate running for Congress with an account on X.

The report’s authors found that more than 20 percent of posts directed at Black and Asian women candidates “contained offensive language about the candidate.” It also found that Black women in particular were targeted with hate speech more often compared with other candidates.

“On average, less than 1% of all tweets that mentioned a candidate contained hate speech,” the report says. “However, we found that African-American women candidates were more likely than any other candidate to be subject to this type of post (4%).” That roughly lines up with X’s recent transparency report — the since Elon Musk took over the company — which said that rule-breaking content accounts for less than 1 percent of all posts on its platform.

Notably, the CDT’s report analyzed both hate speech — which ostensibly violates X’s policies — and “offensive speech,” which the report defined as “words or phrases that demean, threaten, insult, or ridicule a candidate.” While the latter category may not be against X’s rules, the report notes that the volume of suck attacks could still deter women of color from running for office. It recommends that X and other platforms take “specific measures” to counteract such effects.

“This should include clear policies that prohibit attacks against someone based on race or gender, greater transparency into how their systems address these types of attacks, better reporting tools and means for accountability, regular risk assessments with an emphasis on race and gender, and privacy preserving mechanisms for independent researchers to conduct studies using their data. The consequences of the status-quo where women of color candidates are targeted with significant attacks online at much higher rates than other candidates creates an immense barrier to creating a truly inclusive democracy.”

While large language models (LLMs) are becoming increasingly effective at complicated tasks, there are many cases where they can’t get the correct answer on the first try. This is why there is growing interest in enabling LLMs to spot and correct their mistakes, also known as “self-correction.” However, current attempts at self-correction are limited and have requirements that often cannot be met in real-world situations.

In a new paper, researchers at Google DeepMind introduce Self-Correction via Reinforcement Learning (SCoRe), a novel technique that significantly improves the self-correction capabilities of LLMs using only self-generated data. SCoRe can be a valuable tool for making LLMs more robust and reliable and opens new possibilities for enhancing their reasoning and problem-solving abilities.

The importance of self-correction in LLMs

“Self-correction is a capability that greatly enhances human thinking,” Aviral Kumar, research scientist at Google DeepMind, told VentureBeat. “Humans often spend more time thinking, trying out multiple ideas, correcting their mistakes, to finally then solve a given challenging question, as opposed to simply in one-shot producing solutions for challenging questions. We would want LLMs to be able to do the same.”

Ideally, an LLM with strong self-correction capabilities should be able to review and refine its own answers until it reaches the correct response. This is especially important because LLMs often possess the knowledge needed to solve a problem internally but fail to use it effectively when generating their initial response.

“From a fundamental ML point of view, no LLM is expected to solve hard problems all within zero-shot using its memory (no human certainly can do this), and hence we want LLMs to spend more thinking computation and correct themselves to succeed on hard problems,” Kumar said.

Previous attempts at enabling self-correction in LLMs have relied on prompt engineering or fine-tuning models specifically for self-correction. These methods usually assume that the model can receive external feedback on the quality of the outputs or has access to an “oracle” that can guide the self-correction process.

These techniques fail to use the intrinsic self-correction capabilities of the model. Supervised fine-tuning (SFT) methods, which involve training a model to fix the mistakes of a base model, have also shown limitations. They often require oracle feedback from human annotators or stronger models and do not rely on the model’s own knowledge. Some SFT methods even require multiple models during inference to verify and refine the answer, which makes it difficult to deploy and use them.

Additionally, DeepMind’s research shows that while SFT methods can improve a model’s initial responses, they do not perform well when the model needs to revise its answers over multiple steps, which is often the case with complicated problems.

“It might very well happen that by the end of training the model will know how to fix the base model’s mistakes but might not have enough capabilities to detect its own mistakes,” Kumar said.

Another challenge with SFT is that it can lead to unintended behavior, such as the model learning to produce the best answer in the first attempt and not changing it in subsequent steps, even if it’s incorrect.

“We found behavior of SFT trained models largely collapses to this ‘direct’ strategy as opposed to learning how to self-correct,” Kumar said.

Self-correction through reinforcement learning

To overcome the limitations of previous approaches, the DeepMind researchers turned to reinforcement learning (RL). 

“LLMs today cannot do [self-correction], as is evident from prior studies that evaluate self-correction. This is a fundamental issue,” Kumar said. “LLMs are not trained to look back and introspect their mistakes, they are trained to produce the best response given a question. Hence, we started building methods for self-correction.”

SCoRe trains a single model to both generate responses and correct its own errors without relying on external feedback. Importantly, SCoRe achieves this by training the model entirely on self-generated data, eliminating the need for external knowledge.

Previous attempts to use RL for self-correction have mostly relied on single-turn interactions, which can lead to undesirable outcomes, such as the model focusing solely on the final answer and ignoring the intermediate steps that guide self-correction.

“We do see… ‘behavior collapse’ in LLMs trained to do self-correction with naive RL. It learned to simply ignore the instruction to self-correct and produce the best response out of its memory, in zero-shot, without learning to correct itself,” Kumar said.

To prevent behavior collapse, SCoRe uses a two-stage training process with regularization techniques. The first stage replaces SFT with a process that optimizes correction performance while ensuring that the model’s initial attempts remain close to the base model’s outputs.

The second stage employs multi-turn RL to optimize reward at both the initial and subsequent attempts while incorporating a reward bonus that encourages the model to improve its responses from the first to the second attempt.

“Both the initialization and the reward bonus ensure that the model cannot simply learn to produce the best first-attempt response and only minorly edit it,” the researchers write. “Overall, SCoRe is able to elicit knowledge from the base model to enable positive self-correction.”

SCoRe in action

The DeepMind researchers evaluated SCoRe against existing methods that use self-generated data for self-correction training. They focused on math and coding tasks, using benchmarks such as MATH, MBPP, and HumanEval.

The results showed that SCoRe significantly improved the self-correction capabilities of Gemini 1.0 Pro and 1.5 Flash models. For example, SCoRe achieved a 15.6% absolute gain in self-correction on the MATH benchmark and a 9.1% gain on the HumanEval benchmark in comparison to the base model, beating other self-correction methods by several percentage points.

The most notable improvement was in the model’s ability to correct its mistakes from the first to the second attempt. SCoRe also considerably reduced the instances where the model mistakenly changed a correct answer to an incorrect one, indicating that it learned to apply corrections only when necessary.

Furthermore, SCoRe proved to be highly efficient when combined with inference-time scaling strategies such as self-consistency. By splitting the same inference budget across multiple rounds of correction, SCoRe enabled further performance gains.

While the paper primarily focuses on coding and reasoning tasks, the researchers believe that SCoRe can be beneficial for other applications as well.

“You could imagine teaching models to look back at their outputs that might potentially be unsafe and improve them all by themselves, before showing it to the user,” Kumar said.

The researchers believe that their work has broader implications for training LLMs and highlights the importance of teaching models how to reason and correct themselves rather than simply mapping inputs to outputs. 

Sometimes, a pivot ends up being the smartest decision company leaders can make. See Netflix’s pivot from DVDs to streaming, or Corning’s pivot from lightbulbs to touchscreens.

The list of extremely successful startup pivots goes on. And on. And on.

A less-prominent (but by no means failed) pivot is Numa’s. Its co-founders killed the startup’s original conversational AI product to instead sell customer service automation tools. Not just any tools, though — these tools are targeted at auto dealerships.

That sounds like a highly specific niche, but it’s been profitable, according to Tasso Roumeliotis, Numa’s CEO. The company closed a $32 million Series B round in September.

“We were early to build AI and conversational commerce,” Roumeliotis told TechCrunch in an interview. “But we decided to focus our AI entirely on the automotive vertical after identifying enormous opportunity in that space.”

Roumeliotis co-founded Numa in 2017 with Andy Ruff, Joel Grossman, and Steven Ginn. Grossman hails from Microsoft, where he helped ship headliner products like Windows XP, as well as a few less recognizable ones like MSN Explorer. Ruff, another Microsoft veteran, led the team that created the first Outlook for Mac client.

Numa is actually the co-founders’ second venture together. Roumeliotis, Grossman, Ginn, and Ruff previously started Location Labs, a family-focused security company that AVG bought for $220 million 10 years ago.

What rallied the old crew behind Numa, Roumeliotis says, was a shared belief in the potential of “thoughtfully applied” AI to transform entire industries. “The market is full of AI and automation point solutions or broad, unfocused tools,” he said. “Numa offers an end-to-end solution that prioritizes the needs of the customer: car dealerships.”

The U.S. has more than 17,000 new-car dealerships, representing a $1.2 trillion industry. Yet many dealerships struggle to manage customer service requests. Per one survey, a third of dealers miss at least a fifth of their incoming calls.

Poor responsiveness leads to low customer service scores, which in turn hurt sales. But Numa can prevent things from getting that bad — or so Roumeliotis claims — by tackling the low-hanging fruit.

Numa uses AI to automate tasks such as “rescuing” missed calls and booking service appointments. For example, if a customer rings a dealership but hangs up immediately afterward, Numa can send a follow-up text or automatically place a reminder call. The platform can also give customers status updates on ongoing service, and facilitate trade-ins by collecting any necessary information ahead of time.

“Many dealerships still rely on legacy systems that are inefficient and lack integration with modern, AI-driven platforms,” Roumeliotis said. “Today’s consumers expect fast, seamless interactions across all platforms. Dealerships struggle to meet these expectations, especially in areas like real-time communication, service updates, and personalized experiences, which AI can help address.”

Other small-time automation vendors (e.g., Brooke.ai, Stella AI) provide products designed to ease dealerships’ customer service burdens. Tech giants, meanwhile, sell a range of generic solutions to automate away customer service. But Roumeliotis argues that Numa stands out because it understands how workflows within dealerships impact the end-customer experience.

“Dealership service leaders and employees are running around constantly, handling customers in person, going out to check on cars and parts, dealing with ringing phones, and balancing coordination with co-workers,” Roumeliotis said. “Numa brings that all together in a way intentionally designed with AI and the user inside the dealership to drive how the platform works rather than the other way around.”

Roumeliotis asserts Numa has another advantage in its in-house models, which drive the platform’s automations. He said the models were trained on datasets from OEMs and dealership systems as well as conversation data between dealerships and clients.

Were each one of these clients, OEMs, and dealerships informed that their data would be used to train Numa’s models? Roumeliotis declined to say. “Numa’s models are bootstrapped by a feedback loop between dealerships, customers interacting with dealerships, and the usage of Numa to facilitate this,” he said.

That answer probably won’t satisfy privacy-conscious folk, but it’s seemingly immaterial to many dealerships. Numa has 600 customers across the U.S. and Canada, including the largest retail auto dealership in the world. Roumeliotis claims Numa is “just about” cash-flow break-even.

“We don’t need capital to continue scaling revenue,” he added. “Instead, Numa is using its money to accelerate product development by expanding our team of AI and machine learning engineers, including investing in building AI models for the automotive vertical.” The company currently has 70 employees.

Benefiting Numa in its conquest is the willingness of dealerships to pilot AI to abstract away certain back-office work.

According to a survey by automotive software provider CDK Global last year, 67% of dealerships are using AI to identify sales leads, while 63% have deployed it for service. Those responding to the poll were quite bullish on the tech overall, with close to two-thirds saying that they anticipated positive returns.

Touring Capital and Mitsui, a Japanese conglomerate that’s one of the largest shareholders in automaker Penske, led Numa’s Series B round. Costanoa Ventures, Threshold Ventures, and Gradient, Google’s AI-focused venture fund, also participated in the round. The funding brings Oakland-based Numa’s total raised to $48 million.