FlowingFlowing
← Back to all development cases
Development case · blind model judgment pending independent human review; not a formal Benchmark conclusion.
AutocompleteEvidence version preferred11 / 162 · 27f2ba81533c87f2

Predictive Coding: a Theoretical and Experimental Review

Quantitative biology · 2107.12979v4

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 autocomplete positions stratified?

Before seeing generation outcomes, we check whether a retrieved passage contains a specific proposition that directly supports the next writing move. Those positions appear in the left opportunity group; the rest are ordinary positions on the right. We select a balanced sample from admitted papers in each field. The 50/50 split is experimental, not a measure of how often either type occurs in writing.

AUTOCOMPLETE · 110

Biology, statistics and astrophysics

55 positions on each side; statistics uses the ten-paper rerun.

AUTOCOMPLETE · 52

Psychology and climate science

26 positions on each side; psychology includes nine papers and climate science four.

How are the table percentages calculated?

Across five fields, 51 of 81 left-column positions preferred the Evidence version. Ties, pairs where both versions were unusable and incomplete reviews remain in the denominator.

51Evidence version preferred
÷
81All positions in this group
=
63%Evidence preference in this group

Source contribution is a separate review: 16 of 22 Evidence wins entered into source review directly used retrieved papers; another 29 wins await review.

These are development-stage model judgments 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

Predictive Coding: a Theoretical and Experimental Review

5 Discussion and Future Directions

…ic plasticity and other basic functions ( Turrigiano (1999) ; Watt & Desai (2010) ) , or else could be implementing some kind of layer-wise normalization – which should theoretically be describable as predictive coding – but in general the exact function of this connectivity at a computational and algorithmic level is unknown both within and outside of predictive coding, as is the interaction between top-down and lateral connectivity. As such, it may be that predictive coding might be ‘right’ in some sense, but still missing core aspects of the computation that actually goes on in the cortex.

…ke architecture for action. However, it is also possible that the highest and most abstract levels of the brain may be primarily discrete, necessitating discrete generative models for their mathematical description which would entail a departure from predictive coding. Similarly, adaptive action in biological organisms requires long-term planning and maximization of future reward signals, which cannot readily be implemented within the predictive coding paradigm but instead likely relies on a complex set of machinery specialised for performing reinforcement learning ( Sutton & Barto (2018) ) .

↓ 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: 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

Preferred by blind review

A further open question concerns the relationship between predictive coding and the exploratory behaviour that supports adaptive action. If exploration is driven by minimizing a divergence between desired and predicted states rather than by maximizing an evidence bound, then the objective functional underlying predictive coding may not by itself account for information-seeking behaviour, and the two frameworks would need to be reconciled explicitly.

Passages supplied to the Evidence version

Applications of the Free Energy Principle to Machine Learning and Neuroscience

We saw that to obtain information-seeking exploration as a core part of the objective functional, in addition to reward maximization crucially entails minimizing a divergence objective instead of an evidence objective. We then related this new dichotomy between divergence and evidence objectives to…

Read full passage excerpt

We saw that to obtain information-seeking exploration as a core part of the objective functional, in addition to reward maximization crucially entails minimizing a divergence objective instead of an evidence objective. We then related this new dichotomy between divergence and evidence objectives to a wide range of currently used objectives within the reinforcement learning and theoretical neuroscience communities. The importance of this result, really, lies not in the relationship to existing methods, but what it tells us about the deep foundation of exploration. Put simply, we see that extrinsic exploratory drives emerge from trying to match rather than maximize. Matching tries to maintain the complexity of the inputs, so that given a complex desire distribution, agents are driven to stabilize a similarly complex future.

No Evidence

Same model and position · no retrieved passages · first output

Not preferred in this comparison

A further open question concerns the relationship between predictive coding and the brain's reward and value systems. While predictive coding accounts for perceptual inference and learning, it does not by itself specify how value is represented or how actions are selected to maximise future reward, and bridging this gap remains an active area of theoretical work.

Blind model judgment

The blinded model review preferred the Evidence version. Its original rationale appears below.

Read the original model judgment

Both outputs are coherent continuations, but A is more tightly grounded in the supplied source and better extends the draft's specific thread. The draft ends with predictive coding's limitations regarding 'adaptive action,' 'long-term planning,' and 'maximization of future reward signals' in reinforcement learning. Output A directly engages with the source's core claim about 'divergence objective instead of an evidence objective' and 'information-seeking exploration,' connecting this explicitly to predictive coding's potential inadequacy for exploration. This matches the source's specific mechanism (divergence vs. evidence) and the draft's concern about reinforcement learning machinery. Output B is more generic—it states predictive coding doesn't 'specify how value is represented or how actions are selected,' which is true to the draft but adds no new source-grounded content. The source specifically discusses divergence objectives and exploration, not generic value representation. Output A uses this specific source contribution to advance the argument, making it more useful for a researcher who could insert it with minimal editing while preserving and extending the draft's meaning.

Source review confirmed that this continuation directly used specific, verifiable information from the retrieved paper.

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.