985,559,360 failed attempts. That's the spam toll on Base across nine months starting June 2025—and 95% of it came from one bot architecture: on-chain discovery bots. A Flashbots study by Fei Wu and Burak Öz dissects the competition and shows that protocol changes like Flashblocks and fee floors reshaped the bot population without eliminating the probing game.
Three Architectures, One Spam Factory
The study classifies arbitrage bots into three architectures using trace-level classifiers on 359,752 sampled transactions. Off-chain discovery bots simulate routes off-chain and commit the full route in calldata—they post a 42.5% success rate and produce the least spam. On-chain evaluation bots run confined profitability checks against a pre-selected route; they sit in the middle with 18.5% success. On-chain discovery bots fire transactions that scan dozens of pools on-chain and only decide which route to execute mid-flight—4.58% success rate, and they're responsible for 934 million of the 985 million spam transactions. Those scans mean a median of 40 pre-execution reads per successful arbitrage, and the spam consists almost entirely of probes that terminate after state reads without finding a profitable path.
Flashblocks Didn't Reduce Spam, It Changed Who Probes
Flashblocks activated on July 7, 2025, introducing 200ms sub-blocks with incremental gas budgets. Early sub-blocks have limited space, so heavy on-chain discovery bots with broad venue scans were displaced. On-chain discovery's share of active bots dropped from 63.1% to 25.1% in a four-week window. But aggregate spam stayed nearly flat (68.4M pre to 70.5M post) because surviving bots doubled their attempt frequency—spam per active bot jumped from 72.5K to 314.7K. Off-chain discovery bots got more efficient (success rate from 35.8% to 43.6%) as shorter block intervals reduced stale-state risk. Finer ordering changed who probes, not whether they probe.
Fee Floors Raised the Hurdle, Not the Strategy
Base ramped the minimum base fee from 200K to 5M WEI between December 2025 and February 2026. On-chain discovery's bot share fell from 22% to 8.4%, and off-chain discovery rose to 71.1%. The selection was two-dimensional: bots with high volume per success and few attempts survived; low-value high-attempt bots got priced out. Surviving on-chain discovery bots didn't leaner transactions—they just required richer opportunities to cover the cost. USD volume per attempt rose from 0.36 to 1.31 at the cleanest steps. The fee floor killed low-yield probing, not probabilistic search itself.
Across the entire sample, successful arbitrages consumed just 1.31% of Base gas; on-chain discovery spam ate 19.59%. Flashblocks and fee floors improved efficiency—spam gas share fell from a third to ~15%—but opportunity shocks (like the September token launches of AVNT and MIRROR) revived probabilistic search overnight. The competition between targeted and probabilistic architectures is contingent on protocol parameters and market conditions, not a solved design problem. Future work needs to probe whether revert protection, richer state channels, or priority-access mechanisms like Arbitrum's Timeboost can push further without breaking the latency tradeoffs that make high-throughput chains attractive.
Source: To Wait or To Probe: Arbitrage Competition on High-Throughput Blockchains
Domain: collective.flashbots.net
Comments load interactively on the live page.