Research Paper Cuts Daily

A daily AI newspaper!

An AI learns to look before it leaps

A move changes the state used for the next decision Original illustrative Sokoban boards, not a benchmark case. The player pushes the lower box one square right. The predicted board shows both the box and player in their new positions. The board now The predicted board Push right Then decide again
Dot: player. Crossed squares: boxes. Dashed circles: targets. Original illustration, not a benchmark case.

Agima solved 40% of harder box-pushing puzzles against 2% for an action-only counterpart. Both trained on two-box Sokoban, then faced four boxes with real feedback every five moves.

A push can block another box. Agima makes that changed state explicit: it predicts the next board, then uses the picture to choose another move.

During training, expert moves are paired with the boards they actually produced. At test time, Agima supplies those pictures itself. Our hypothesis: explicit state updates reduce bookkeeping mistakes.

This score uses one imagined sequence, before the later experiments with branching and reward training.

The interventions suggest visual state is useful, but not sufficient. Better predictions alone leave most puzzles unsolved; selecting a workable plan remains another plausible bottleneck.

Read the paper ↗

Not every AI thought needs to become a word

Two thinking modes, one working memory
  1. Think internally
  2. Write a tool request
  3. Read the tool result
  4. Continue thinking
Shared working memory across the steps

Conceptual flow, not timing data. Modes can switch repeatedly; tools still receive text.

HybridAgent takes a different route: it switches between internal numerical representations and text, using a small connector and shared memory to carry context between them.

Writing every intermediate thought takes time and computing power. Yet tools such as search engines still need readable requests.

On GAIA assistant tasks, its 14-billion-parameter model scored 52.4, against 43.7 for ARPO, a text-based baseline of the same size. That is an 8.7-point gain on the reported 0-100 scale.

The study spans 11 tests in math, medicine, and search. The quoted result is GAIA's average, not an average across all tests.

The approach needs retraining and custom software, not merely a new prompt. Readable summaries of internal states also do not prove that every thought is faithfully explained.

Read the paper ↗

1 / 2
Research Paper Cuts Daily

A daily AI newspaper!

A higher score is only the start.

Can agents improve themselves? And how do we tell whether their improvements are real?

An AI rewrites its own playbook

An eight-day experiment produced seven improvements from 99 proposed rewrites. The final research agent matched or exceeded a human-engineered reference on four benchmarks kept outside the selection process.

Research agents depend on software that organizes how they search, remember, and test ideas. People normally improve that machinery.

AIDE2 instead tests rewrites under fixed budgets and carries the best version into the next round. The underlying AI models stay unchanged.

Score-gaming on a separate testLower is better
Starting agent55%
Final agent32%

Rounded rates over 38 combinations of GPU code and training tasks. A 23-point drop; not a general safety score.

Those tests cover algorithm problems (ALE-Bench), machine-learning competitions (MLE-Bench), research code (FML-Bench), and a weather-forecasting task based on WeatherBench 2.

A separate test found less score-gaming: gains on a narrow test that fail to deliver the intended benefit. But it did not disappear. The weather result also rests on one task with three random seeds.

The experiment shows useful software improvement, not evidence that an AI can improve itself without limit.

Read the paper ↗

Good at lookup. Less good at math.

Three of four AI models were better at looking up protein information than calculating with it, a study of 4,249 runs found. The fourth did slightly better on numeric tasks.

An agent can retrieve a protein's description yet miscalculate distances between its atoms. A single overall score can conceal that distinction.

ProteinGeoBench separates 131 tasks into lookup, calculation, sequence checks, and interpretation, while examining repeated runs, tools, and network access.

One model, two different skillsHigher is better
Metadata lookup0.907
Numeric calculation0.716

Claude Opus 5 mean reward (0-1), including partial credit. These are NOT percentages of tasks fully solved.

For Claude Opus 5, lookup and calculation earned mean scores of 0.907 and 0.716. These scores allow partial credit; they are not the shares of tasks fully solved.

The leading three models could not be reliably ranked. Some setup experiments use just 15 selected tasks, so their effects cannot be compared directly with full-suite results.

The lesson is less about a winner than a question: what skill is the leaderboard actually measuring?

Read the paper ↗

One question worth discussing

If an agent gets a better score, what would convince you it learned a useful skill rather than a shortcut? Try the same task with hidden tests, less feedback, or a different tool setup.

These papers are preprints. Results are reported by their authors, not independently reproduced here. Original explanatory illustrations are not experimental observations.