Currently, there is no way to resolve disagreements of the same block height in Bitcoin protocol.  Floating-point Nakamoto Consensus (https://github.com/in-st/Floating-Point-Nakamoto-Consensus/blob/master/Floating-Point%20Nakamoto%20Consensus.pdf) address ambiguity in the consensus formation process so that disagreements can be empirically resolved without wasted effort. With FPNC every block has a non-zero element which provides the basis for a floating-point fitness value. Nodes are already incentivised to choose the solution that represents the most amount of work, FPNC allows for this same calculation to happen for two solutions of the same height. With FPNC the higher fitness-value carries forward, and all children of this higher value block will be stronger from having a higher fitness value, this is to make sure that  winning blocks stay as the winner.  If a rogue client chooses to mine a low value disagreement, they will have to make-up the difference in fitness score with their next solution — This is the genetic algorithm supported by FPNC which ensures that the child blocks from a loser will also be losers.

Using FPNC as a method of disagreement resolution enables two features that provide better incentives over "first seen." For one,  FPNC introduces risk into holding onto low-value solutions, which de-incentivises 1/2 of all withholding attacks.  Additionally, FPNC can be used to increase the rate of block formation which will reduce the amount of time that a miner can hold onto private blocks.

With 1/2 of the blocks being withheld, and 1/2 not being broadcast immediately, you could eventually identify malicious miners based on this timing difference. With the current system it isn't as clear, but with a split incentive you the network can observe unfair treatment of high-valued blocks.  FPNC makes the silent process of withholding into one that must show a value-bias, and this unethical behavior can be observed and acted upon. 

The bitcoin protocol uses a mining difficulty that depends on a zero-prefix. As a result, this difficulty function consists of discrete jumps that grow exponentially, and there are some very good reasons not to rely on this construct.   Instead of zeros, a range of floating-point values, or fractional parts of whole numbers can be used as the difficulty cut-off, where any solution more difficult than the cut-off is still accepted. Because the proof-of-work is no longer dependent on an arbitrary 0, moving to an floating-point value range would allow block-formation to be on an arbitrary time-schedule, which could be made slower or faster.  The amount of time that a block can be withheld is proportional to the amount of time it takes to generate,  therefore a faster block formation time inturn limits the amount of time that a block can be withheld.   If block formation is on average 30 seconds to 1 minute, then the amount of time a miner can impact the network is capped at seconds instead of minutes.   Although it doesn't stop withholding attacks, speeding up the dramatically limits the amount of time that any attacker can withhold a block, thus mitigating the impact of malignant miners.  Speeding up block formation time while keeping inflation targets the same adds value to users of the network.

Currently on the BItcoin network, any malicious miner performing a withholding attack does not need to be at the mercy of network conditions, and would be more successful if they preemptively spread their delayed solution. With an artificially-increased connection capacity a node can gain a visibility advantage on the bitcoin network.  When this greedy miner sees that a competing miner has released a block — then the greedy miner can pre-empt the spread of their delayed solution and beat out any honest solution.  A miner using pre-emption to artificially spread their side of the consensus can ensure the adoption of their block because honest miners are dependent on natural topography of the network to spread their messages — a topography which is not optimized for speed. It isn't that the attacker is all powerful, it is that p2p networks have an inherent higher-latency than centralized systems.  A miner can have a pre-computed map of the bitcoin network and then reach out and inform each node of the delayed block before the honest block has a chance to arrive.  Thus shaping the disagreement in the favor of the malicious miner.  So long as a malicious miner has sufficient visibility, they unlock a luxury of subjecating would-be dissent and guaranteeing that their solution is the one that always wins.

An attacker cannot choose their FPNC fitness value, but they can choose when their block arrives and to whom. The "first seen" approach to block adoption pressures malicious miners into a race of message propagation that convinces undecided nodes to work for the dishonest chain. The race-condition in block arrival is fundamentally resolved because the order in which blocks arrive doesn't influence the block's FPNC value. Therefore having clear disagreement resolution with FPNC removes a feature of the network that can be leveraged to make sure that a block-withholding attack supplants honest blocks.

By clarifying the rules in which blocks will be replaced  — fewer disagreements will form. Without having a form of disagreement resolution and leaving the process up to time, then nodes can be deputized by malicious miners and aid in the mining of withheld blocks.

-Michael