Hello, this is the developer of TOOSIN : 투신. The V1.1 update focuses on feedback from our players. This update improves Blood Contract balance, defensive special abilities, AI pattern learning, the first-launch tutorial, multilingual UI, and overall stability.
Some contracts had penalties that were too severe compared to their benefits, while others were not worth choosing even at Legendary rarity. We have rebalanced them based on their actual pre-V1.1 values so that each contract offers a meaningful choice depending on your build. Values without a `%` symbol are flat values.
Blood ContractBefore V1.1After V1.1
Glass CannonRare / Attack Power +50% / Max Health -50%Legendary / Attack Power +30% / Max Health -40%
Giant's ArmorMax Health +50% / Attack Power -30%Max Health +50% / Attack Power -20%
Assassin's InstinctCritical Chance +20% / Defense -10Critical Damage +30% / Defense -30%
Vampire's ThirstLife Steal +5% / Max Health -30%Life Steal +5% / Max Health -20%
Berserker's RageAttack Speed +30% / Max Stamina -40%Attack Speed +30% / Life Steal -5%
Will of SteelDefense +15 / Attack Speed -10%Defense +10% / Attack Speed -15%
Boiling BloodLife Steal +5% / Damage Taken +20%Life Steal +5% / Damage Taken +25%
Broken SwordAttack Power +70% / Max Health -70%Attack Power +70% / Max Stamina -70%
Cursed ArmorDefense +40 / Max Stamina -50%Defense +40 / Max Stamina -30%
Hawk EyeCritical Chance +5% / Max Health -30Critical Damage +15% / Max Health -10%
Mark of DeathArmor Penetration +50% / Damage Taken +50%Armor Penetration +45% / Damage Taken +25%
Frenzied FeastCritical Chance +30% / Damage Taken +40%Critical Damage +50% / Damage Taken +30%
Sniper's Breath, Heavy Strike, Ghost's Gait, Turtle's Endurance, Tireless Vitality, Reckless Charge, Light as a Feather, and Resilient Resolve retain their previous effects. Once every available Blood Contract has been acquired, Blood Contract events and additional contract selection prompts will no longer appear.
We have adjusted defensive special abilities that allowed players to invest entirely in offense without taking enough risk.
Special AbilityActual Duration Before V1.1After V1.1
Auto Parry7 seconds3 seconds
Aegis Shield5 seconds3 seconds
Time Stop5 seconds3 seconds
These abilities retain their core functions, but choosing the right moment to activate them is now more important.
Many players told us that it was difficult to tell whether the AI was actually learning during shorter fights, as only minor changes in its patterns were noticeable. AI learning is one of the defining features of TOOSIN : 투신. We have therefore reviewed both the way player behavior is measured and how those observations affect enemy decisions.
Learning CategoryBeforeAfter V1.1
Learning updatesLong-term data was primarily calculated once when an enemy spawned, making adaptation difficult to notice during short fights.Long-term memory and current-combat observations are now handled separately. Once enough evidence is collected, current habits continue to affect the same enemy's next decisions.
Light and heavy attacksAttack choices and combo hit counts could be mixed together, producing inaccurate results.The game separately measures actual light and heavy attack starts and average combo length. These values affect guard-versus-dodge decisions and the use of quick pokes or longer attack chains.
Approach, retreat, and lateral movementMovement frames could be mixed with unrelated action counts, distorting longer fights.Approach, retreat, lateral movement, and average distance are measured separately. Enemies use them to adjust circling, backsteps, preferred distance, and pursuit behavior.
Dodges and forward dodgesDodge attempts were not always clearly separated from successful dodges.A dodge is recorded only after it successfully begins, while forward dodges are tracked separately. These observations affect pursuit, forward-dodge entries, and attack pacing.
Guard, parry, and hit outcomesDefensive choices and their actual results were not clearly separated.Guard-versus-dodge choices and guard, parry, and hit outcomes are analyzed separately. Strong defenders are pressured with longer chains, while exposed players may face quicker attacks.
Attack timing and rhythmTiming responses depended more heavily on difficulty and random values than on the player's real habits.Actual attack intervals and repeated rhythms now affect enemy reaction time, parry timing variation, attack pacing, and recovery-punish decisions.
Special abilitiesInstant abilities could be missed when only their active state was checked.Successful activations are counted directly. Repeated use affects the chance that an enemy will respond with an appropriate defensive or evasive action.
Dodge-to-heavy attacksThis sequence was not tracked as a separate pattern.Heavy attacks started within 0.8 seconds after a dodge ends are tracked separately. With enough evidence, enemies may punish the recovery or selectively imitate the same tactic.
BehaviorBeforeAfter V1.1
Range and attack decisionsEnemies often attacked through an independent probability check as soon as the player entered attack range.With enough evidence, enemies consider the player's movement and attack habits before choosing how to enter.
Connected pressureCircling, dodging, and attacking were selected independently, making sequences feel disconnected.Enemies can use a connected tactic: circle the player → enter with a forward dodge → briefly settle → attack.
Recovery punishmentTiming depended heavily on difficulty and random values, making counters slow or inconsistent.Enemies that have learned the player's attack rhythm can anticipate the end of an attack and respond more quickly.
Dodge-to-heavy imitationEnemies did not learn or use this sequence as a separate tactic.When the player repeats it often enough, skilled enemies may imitate the same dodge-to-heavy tactic with a limited chance and cooldown.
Learning fairnessStage and difficulty values could feel more influential than actual combat evidence.One or two actions never guarantee a counter. Learning affects future decisions only after enough evidence is collected and does not cancel an attack or defense already in progress.
Previous messages did not clearly explain what the AI had learned or how it intended to respond. Notifications now directly describe both the recognized habit and the enemy's intended response.
NotificationBeforeAfter V1.1
Learning informationMessages mainly used broad descriptions such as “Learning detected” or “The next enemy has become stronger.”The notification explains the observed habit and how the current enemy intends to respond.
Heavy-attack notification shown in the imageIt was difficult to tell whether the AI had recognized a preference for heavy attacks.When at least four valid attack choices have been observed and heavy attacks account for 60% or more, the game displays: “We remember the breath you take before putting your weight into a strike. The enemy will set their blade and wait to counter your heavy blow.”
Executed responseLearning messages were not visibly connected to the enemy's actual actions.When an enemy executes a response, the game identifies whether it used a guard, dodge, counter-strafe, forward dodge, counterattack, recovery punish, dodge-to-heavy imitation, or special ability.
The message shown in the image means that the AI has collected enough evidence to identify the player's preference for heavy attacks. It may adjust its guard, dodge, and counterattack weights when a heavy attack is expected. It does not mean that every following heavy attack will be automatically parried or countered.
To prevent the learning information from permanently occupying the screen, it is now displayed only while the assigned button is held.
Displayed InformationWhat It Means to the Player
Current confidenceHow much usable evidence has currently been collected. It does not represent AI difficulty or win rate.
Observed attacksThe ratio and sample count of actual light and heavy attack starts.
Observed movementThe ratio of approach, retreat, and lateral movement relative to the enemy.
Observed defense choiceThe ratio of dodges and guards that were actually selected.
Observed hit outcomesWhether enemy attacks were guarded, parried, or landed on the player.
Applied learning valuesHow strongly the collected behavior currently affects the enemy's dodge, circling, and long-combo decisions.
Planned responseThe response the enemy is currently preparing or considering. It may change or be cancelled if distance, stamina, or combat state changes.
Last executed actionThe most recent action the enemy actually performed using the learned information.
Combat observations are analyzed approximately every 0.65 seconds and applied to the enemy's next decision. The AI does not instantly read the player's input or cancel an action already in progress. Enemies can now use special abilities starting from Stage 100 in Campaign Mode and Round 10 in Infinite Mode.
When starting the game with new save data, a basic control tutorial will begin before pattern learning.
Input icons matching the currently used keyboard and mouse or gamepad controls will be displayed. Completing each action advances the tutorial to the next step. The Options screen can also be opened from the pause menu during the tutorial. Language and control settings can be changed without resetting the current tutorial step.
Automatic health regeneration is only applied while out of combat. A player is considered out of combat after spending at least 2 seconds without dealing or receiving damage. Descriptions for Traits, Perks, the Stats screen, and detailed tooltips now match the actual calculation: Regenerates N% of Max Health per second — activates 2 seconds after dealing or receiving damage
New actions and an expanded range of learnable behavior are planned for the next V1.2 update rather than V1.1. One of the actions currently being planned is a shield bash performed by pressing the left mouse button while guarding. An enemy struck by the shield bash will be briefly stunned, with a chance to be pushed backward or knocked down. We also plan to expand the AI's behavioral range so that these new actions are not limited to the player. Enemy AI will be able to observe, respond to, and learn from them.
Through your feedback, we found several cases where translations were broken or text extended outside its intended area depending on the selected language. We found these issues particularly often in Chinese, Russian, and Japanese. We have fixed the clipping and translation issues identified for this update, but we are also planning a broader overhaul of the entire UI. We would appreciate your patience while we work on it. We will address newly discovered issues as quickly as possible and continue applying improvements in future updates. We believe AI learning is one of the defining strengths and core identities of our game. Our goal is not simply to make minor adjustments to enemy attack probabilities. We want players to clearly understand how an enemy learned from their fighting style and why it selected a particular response. Thank you to everyone who shared their feedback.