FlowingFlowing
← Back to all development cases
Development case · blind model judgment pending independent human review; not a formal Benchmark conclusion.
AutocompleteNo-Evidence version preferred135 / 162 · 18cff12fc5fd3e76

Brain-Inspired Machine Intelligence: A Survey of Neurobiologically-Plausible Credit Assignment

Psychology and cognitive neuroscience · 2312.09257v2

FLOWING EVIDENCE BENCHMARK

How do we tell whether Evidence helps?

At the same writing position, with the same model and task, how does supplying retrieved paper passages change the first output? We compare matched versions and retain ties, unusable outputs and incomplete reviews.

SAME DRAFT · EVIDENCE ON OR OFF

01 · FIXED WRITING POSITION

MANUSCRIPT

Same writing position ▌

Matched autocomplete at the same draft position

02 · TWO MATCHED INPUTS

SHARED BY BOTH

Manuscript context, model, task and prompt

A · With Evidence

Retrieved paper passages supplied

B · Without Evidence

No retrieved passages supplied

LLM

Same model and version

A → first output

B → first output

Blind judge agent

First outputs are anonymized as X and Y

Continuation review: accuracy, fit to the writing task and usability
Returns: X preferred / tie / Y preferred / both unusable

Order check: X / Y → Y / X

HOW DOES BLIND JUDGING WORK?

① Anonymize both outputs
The judge sees the same draft and both first outputs without knowing which received Evidence.

② Compare and swap order
The judge applies task-specific criteria in X/Y and then Y/X order.

③ Review disagreements
A third pass resolves disagreements. Incomplete reviews remain in the denominator.

How are positions and cards selected?

The evaluation uses seven fields with forty fixed ordinary autocomplete positions in each, for 280 total. A position does not need to require a literature citation in its next sentence. Each Flowing request receives three to five retrieved cards: five when all pass the relevance gate, without padding with weak cards. The run supplied 1,355 cards: 244 positions received five, 27 received four, and nine received three.

ADOPTION · 150/280

Adoption coverage

After drafting, the model returns an evidenceUse array identifying cards that materially affected facts, correction, direction, scope, terminology, or expression. A nonempty array counts as adoption.

OUTCOMES · 70 / 7 / 30

Full outcomes after adoption

Paired outcomes are counted only when the model reports adoption and both arms produce text. Among 107 comparable positions, Flowing was preferred in 70, 7 tied, and no-Evidence was preferred in 30; one tie was unusable in both arms.

How are the two table percentages calculated?

ADOPTION COVERAGE

150adopted
÷
280all positions
=
54%

FLOWING PREFERENCE AFTER ADOPTION

70Flowing preferred
÷
107comparable adopted positions
=
65%

evidenceUse is the model’s structured report of its own generation process. It does not require a citation and is not rejected merely because a faithful source paraphrase is not verbatim. Actual manuscript citations continue to use separate strict source verification. The previous 140-position result remains retained for traceability.

These are development-stage model-reported adoption and blind-judge results pending independent human review. They are not formal Benchmark conclusions and do not, on their own, establish causality.

Manuscript writing position

Text excerpt · not a PDF page

Research manuscript · excerpt

Brain-Inspired Machine Intelligence: A Survey of Neurobiologically-Plausible Credit Assignment

3 Discussion

Synergistic Explicit Local Signals, Forward-only ( ‘SEL:FO’ ): recurrent forward-forward/predictive forward-forward (RFF/PFF) learning.

…ghting important gaps that research in the domain might want to consider in future developments. For instance, we can see that control, which relates to behavioral and goal-oriented adaptation (reinforcement learning), poses the greatest difficulty; this is corroborated by the analysis in Table 2 , as only the SEL:DR family has a stronger body of work in the area. 29 29 29 We remark that this state-of-affairs is likely a consequence of the work done on predictive coding that exploits its intimate relationship with the neurobiological process framework known as active inference [ 115 , 334 ] .

↓ Next paragraph was generated here

The highlighted text is near the writing position. This is a layout of the evaluation input, not a PDF screenshot.

A matched comparison at the same writing positionJudgment: No-Evidence version preferred

Both arms use the same model, task and manuscript context. Only the retrieved passages differ. The first recorded output from each arm is shown below.

Evidence

Same model and position · retrieved passages supplied · first output

Not preferred in this comparison

This scalability limitation has motivated efforts to parametrize the key densities of the generative model and recognition distribution with deep neural networks, allowing active inference agents to scale to levels comparable to contemporary machine learning methods.

Passages supplied to the Evidence version

Applications of the Free Energy Principle to Machine Learning and Neuroscience

In this chapter, we consider the application of the free energy principle to action selection, or control, problems. While in the previous chapter on perception, we focused on the process theory of predictive coding, here we focus on the process theory of active inference, and are especially inspire…

Read full passage excerpt

In this chapter, we consider the application of the free energy principle to action selection, or control, problems. While in the previous chapter on perception, we focused on the process theory of predictive coding, here we focus on the process theory of active inference, and are especially inspired by the discrete state-space active inference theory introduced in Chapter 2. Here, we aim to solve a key limitation of those methods – their scalability. We propose to do so by parametrizing the key densities of the generative model and recognition distribution by deep neural networks, and then utilizing the tools of deep reinforcement learning to allow active inference agents to scale to levels comparably achieved by contemporary machine learning.

Model-reported card use

The model did not report that any card materially affected this completion.

No Evidence

Same model and position · no retrieved passages · first output

Preferred by blind review

A second gap concerns the evaluation of credit assignment rules themselves: most families are assessed on small, task-specific benchmarks, which makes it difficult to compare their scalability and biological plausibility on common ground. Addressing this would require shared tasks and metrics that expose the trade-offs between local learning and global performance.

Blind model judgment

The blinded model review preferred the no-Evidence version. Its original rationale appears below.

Read the original model judgment

Output A contains an unsupported claim. The source discusses scalability limitations of active inference methods and proposes deep neural network parametrization as a solution, but the draft's paragraph is about control being a difficult gap and the SEL:DR family having more work due to predictive coding's relationship with active inference. Output A incorrectly pivots to 'This scalability limitation' as if the draft had identified scalability as the gap, and presents the deep neural network solution as if it follows from the draft's discussion. The draft never mentions scalability as a problem, nor does it discuss deep neural networks. The source's content about scalability and deep neural networks is not applicable to continuing this specific paragraph, which is about control difficulty and the SEL:DR family's stronger body of work. Output B, by contrast, introduces a coherent second gap about evaluation of credit assignment rules on small benchmarks and the need for shared tasks—this is a plausible continuation that stays within the scope of 'important gaps that research in the domain might want to consider' established by the draft, without importing unsupported factual claims from the source or inventing unsupported specifics. While B does not draw on the supplied source, it does not need to—the task is to continue the paragraph coherently, and B does so without unsupported claims. A's 'This scalability limitation' falsely presupposes the draft identified such a limitation, and the deep neural network solution is imported from a mismatched source context.

The passages are input material; their presence does not mean an output used them or that they support every claim. Draft and source passages are excerpted; no simulated scores or PDF appear here.