#Goblin-of-Thought

GOB-5.5's reasoning strategy. A goblin-flavored alternative to chain-of-thought.

##The problem with chain-of-thought

Standard chain-of-thought (CoT) reasoning works step-by-step:

01.Identify the problem
02.Break it into sub-problems
03.Solve each in order
04.Combine into final answer

This works well for problems with a clear linear decomposition (math word problems, logic puzzles). It works poorly when:

▸The obvious decomposition is wrong
▸The shortest path requires lateral thinking
▸The "correct" answer is to reframe the question
▸The problem has hidden constraints

CoT defaults to the most-traveled path. Goblins, famously, do not.

##How Goblin-of-Thought works

GoT inverts the order. Before committing to a reasoning chain, the model performs a lateral scan of the solution space:

text

1. Receive query
2. Generate K candidate "obvious paths" (similar to CoT)
3. Generate K candidate "non-obvious paths" — explicitly different shapes
4. Score each by GRS (favoring lateral solutions) + estimated difficulty
5. Pick top-N candidate(s)
6. Run full reasoning on chosen candidate(s) only
7. Output

The key insight: instead of committing to the first plausible chain, the model briefly considers "what if the obvious approach is wrong?" before proceeding. This adds ~5–15% latency in exchange for a substantial quality bump on tasks where lateral thinking matters.

##When it matters

GoT shows the largest gains on tasks where the surface-level approach is wrong:

Benchmark	CoT (gpt-4o)	GoT (gob-5.5)	Δ
Lateral Thinking	34.1	61.3	+27.2
Subtext Detection	41.2	89.4	+48.2
Trick Questions	52.7	78.9	+26.2
Standard MMLU	88.7	89.2	+0.5
HumanEval	90.2	93.1	+2.9

It does not help on tasks where the obvious approach is correct (most coding, factual lookup). It doesn't hurt either — the lateral scan converges quickly when the answer is straightforward.

##Controlling it

GoT is always on by default. There's no top-level toggle, but you can influence its aggressiveness through related parameters:

Parameter	Effect on GoT
`mining_depth`	Higher depth → more lateral candidates considered
`horde_mode: "broad"`	More diverse expert clusters → wider candidate spread
`temperature`	Higher → more creative candidates, more variance
`grs_target`	Higher → favors lateral candidates with more goblin-flavor

To force a more conventional chain-of-thought (e.g. for benchmarking parity with other models):

json

{
  "model": "gob-5.5",
  "horde_mode": "focused",
  "mining_depth": 1,
  "temperature": 0.0
}

##Visible reasoning

GoT does not produce visible "thinking" tokens by default — the lateral scan happens internally. To expose it, set reasoning: "visible":

json

{
  "model": "gob-5.5-deep",
  "reasoning": "visible",
  "messages": [{"role": "user", "content": "..."}]
}

The response then includes a reasoning block:

json

{
  "choices": [{
    "message": {
      "role": "assistant",
      "reasoning": "considered three approaches:\n  1. iterate the array (obvious, O(n))\n  2. hash-based lookup (O(1) but extra memory)\n  3. exploit the sorted property (O(log n)) — picked this\n",
      "content": "binary search through the sorted array..."
    }
  }]
}

Reasoning tokens count toward your token budget but at a discounted rate (50% of the regular completion rate). See Pricing for details.

##Caveats

▸GoT can occasionally pick a too-clever solution when a boring one would do. If your task wants conventional answers, use horde_mode: "focused" and mining_depth: 1.
▸GoT-driven creativity scales with temperature. At temperature=0, lateral scan still happens but converges to the same candidates every time.