--reasoning-effort: control thinking intensity (verified honored)

The adapter dropped reasoning_effort/thinking from the payload, so the
middle rung of the ladder (es reference: full 98.3 / effort_low 94 /
no-think 82.3 on humaneval) was unreachable. Both keys now pass
through; --reasoning-effort {minimal,low,medium,high,max} overrides
the YAML, and config/effort_low.yaml mirrors default.yaml with
reasoning_effort: low for one-command low-thinking runs.

Probe on the endpoint: same question, default = 319 chars reasoning /
262 tok, low = 47 chars / 117 tok -- the server honors it.

Co-Authored-By: Claude <noreply@anthropic.com>

evalharness/cli.py

@@ -800,6 +800,9 @@ def _cmd_eval_run(args) -> int:
                         _shared_reporter.owned_externally = True
                     progress_reporter = _shared_reporter
 
+            if getattr(args, 'reasoning_effort', ''):
+                args._gen_override = {**(getattr(args, '_gen_override', {}) or {}),
+                                      'reasoning_effort': args.reasoning_effort}
             if model_spec:  # generate + score in one go
                 from evalharness.model import run_eval
 
@@ -1207,6 +1210,11 @@ def build_parser() -> argparse.ArgumentParser:
                         "profile.default < profile['<bench>'] < explicit kwargs")
     p.add_argument('--disable-thinking', action='store_true',
                    help='send enable_thinking=false to the OpenAI-compatible model')
+    p.add_argument('--reasoning-effort', default='',
+                   choices=('', 'minimal', 'low', 'medium', 'high', 'max'),
+                   help="thinking intensity (GLM/Anthropic-style; verified "
+                        "working on the sglang endpoint: low shrinks reasoning "
+                        "~85%%). Overrides the YAML config")
     p.add_argument('--perf', action='store_true',
                    help='collect streaming TTFT and ITL metrics')
     p.add_argument('--textools', action='store_true',

evalharness/config/effort_low.yaml

@@ -0,0 +1,84 @@
+# 低思考档：全库参数同 default.yaml，仅追加 reasoning_effort: low
+# 用法: evalharness eval run ... --config effort_low
+default:
+  temperature: 0.0
+  top_p: 1.0
+  stream: true
+  max_tokens: 32768
+  reasoning_effort: low
+aime24:
+  temperature: 1.0
+  repeats: 12
+  max_tokens: 8192
+  reasoning_effort: low
+aime25:
+  temperature: 1.0
+  repeats: 12
+  max_tokens: 8192
+  reasoning_effort: low
+aime26:
+  temperature: 1.0
+  repeats: 12
+  max_tokens: 8192
+  reasoning_effort: low
+hmmt26:
+  temperature: 1.0
+  repeats: 12
+  max_tokens: 8192
+  reasoning_effort: low
+imo_answerbench:
+  temperature: 1.0
+  reasoning_effort: low
+gpqa_diamond:
+  temperature: 1.0
+  repeats: 3
+  max_tokens: 8192
+  reasoning_effort: low
+mmlu:
+  max_tokens: 8192
+  reasoning_effort: low
+mmlu_pro:
+  max_tokens: 8192
+  reasoning_effort: low
+cmmlu:
+  max_tokens: 8192
+  reasoning_effort: low
+arc:
+  max_tokens: 8192
+  reasoning_effort: low
+hellaswag:
+  max_tokens: 8192
+  reasoning_effort: low
+winogrande:
+  max_tokens: 8192
+  reasoning_effort: low
+simple_qa:
+  max_tokens: 8192
+  reasoning_effort: low
+trivia_qa:
+  max_tokens: 8192
+  reasoning_effort: low
+humaneval:
+  temperature: 1.0
+  repeats: 3
+  reasoning_effort: low
+live_code_bench:
+  temperature: 1.0
+  reasoning_effort: low
+longbench_v2:
+  max_tokens: 8192
+  max_input_tokens: 128000
+  reasoning_effort: low
+openai_mrcr:
+  max_tokens: 8192
+  max_input_tokens: 128000
+  reasoning_effort: low
+bfcl_v3:
+  max_tokens: 4096
+  reasoning_effort: low
+general_fc:
+  max_tokens: 4096
+  reasoning_effort: low
+tau2_bench:
+  max_tokens: 16384
+  reasoning_effort: low

evalharness/model/adapter.py

@@ -380,7 +380,8 @@ class OpenAICompatible(ModelAdapter):
                 ]
                 payload.pop('chat_template_kwargs', None)
         for k in ('temperature', 'max_tokens', 'top_p', 'stop', 'seed',
-                  'response_format', 'chat_template_kwargs'):
+                  'response_format', 'chat_template_kwargs',
+                  'thinking', 'reasoning_effort'):
             if kw.get(k) is not None:
                 payload[k] = kw[k]
         payload.setdefault('max_tokens', self.extra.get('max_tokens', 4096))  # CoT room