ZeroPoint gives a small decision model a way to choose nothing

When an AI system receives a request and several possible actions, choosing the highest-scoring action is only part of the problem. Sometimes none of the available actions fits. Can the model learn that distinction within its ordinary decision process?

In the preprint ZeroPoint: Lightweight Energy-Based Decision Model, Javier Marín proposes a small model with an explicit empty answer. Its experiments reveal both a useful mechanism for abstention and a substantial limitation: knowing when to reject a list does not necessarily make a model better at choosing within it.

Why relative scores are not enough

An energy-based model assigns each input–answer pair a numerical energy, with lower values indicating better compatibility. Here, energy is a model score, not a measurement of physical energy.

Ordinary option selection depends on differences between these scores. Adding the same quantity to every option’s energy leaves their ranking unchanged. It also leaves unchanged the probabilities obtained by exponentiating negative energies and normalizing across the options—the usual Gibbs distribution.

Consequently, training through those relative probabilities does not establish an absolute reference for deciding whether any option is acceptable. A confident winner can still be the best of several bad choices. Thresholding its probability can help in practice, but that probability distributes all available weight among the listed answers.

ZeroPoint instead includes “no option fits” as a state whose energy is fixed at zero for every input.

Making the empty answer compete

Each listed option receives a non-negative “depth,” with its energy defined as minus that depth. Deeper options therefore have lower energy relative to the fixed zero.

But ZeroPoint does not simply select the lowest of these raw energies. It converts depths into probability weights using an exponential function with one subtracted. That subtraction makes an option’s weight vanish at exactly zero depth. The empty answer has its own positive weight, controlled by a learned scalar, and all weights are normalized together.

The highest-probability state wins. Abstention therefore remains equivalent to applying a depth threshold, but the threshold comes from parameters fitted through the same likelihood objective as the model—not a separately chosen rejection criterion. The numerical depth scale also remains relative to a learned temperature; fixing the zero does not make it a universal unit of compatibility.

Training includes lists with no correct option. On those examples, the model learns to make all listed options shallow so that the empty answer wins.

The encoder reads the request and options together. Its position scheme makes option scores insensitive to their ordering; longer lists are canonically sorted and processed in blocks. The released files occupy 89 MB. Training used 76,211 lists, with the correct option removed from 30% of them.

What the comparisons show

The most revealing comparison holds the encoder features fixed and retrains different output layers on the same data. Across eight tests, the fixed-zero version using the released encoder achieved 69.7% correct decisions, compared with 69.6% for a learned rejection class. Its ability to separate lists with and without a correct answer was slightly better, but the learned rejection class made fewer high-confidence mistakes and had lower calibration error.

A thresholded softmax trained without empty-answer examples performed considerably worse. The author therefore concludes that exposure to lists where nothing fits explains most of the abstention benefit—not the fixed zero alone. This experiment does not isolate how the zero affects learning throughout the encoder, because those features were frozen.

In a separate comparison covering 900 cases, ZeroPoint made 53.6% correct decisions, counting both correct selections and correct abstentions. Strands Decider 2B reached 58.6%. When a correct option existed, ZeroPoint ranked it first in only 43.8% of cases, versus 58.4% for Strands Decider.

ZeroPoint’s median decision time was 42 milliseconds on a two-core CPU, and shuffling options did not change its answers. These are efficiency advantages, not evidence of superior accuracy. CPU, GPU and hosted-service timings are also not directly interchangeable.

Abstention is not a safety guarantee

On cases without a correct option, ZeroPoint still selected an option with probability at least 0.9 in 7.8% of cases. Strands Decider made no such mistakes in that sample.

Long lists pose another problem. The implemented depth function produces strictly positive values, so many shallow options can accumulate probability weight. On CLINC150, correct decisions fell from 90.5% with 20 options to 68.2% with 150. Processing long lists in separate blocks also changes which options can interact.

The released model was trained with one seed, and its multilingual vocabulary was not matched by language-by-language evaluation. The benchmark and comparison wording introduce further uncertainty. These remain preprint results, not an established deployment guarantee.

ZeroPoint addresses a specific question: does any listed answer fit? It does not establish that an input is familiar, safe or understood. Its contribution is a concrete way to learn that narrower decision—and evidence that learning when nothing fits matters more than simply naming a zero.